Campaigns built on voter behavior, measured against a control.
Brushworks reads every voter from what they do, writes the plan to the clock of the race, and measures every program against people it sealed away and never contacted. The Brushworks Desk shows all of it, every morning.
How a program is built, from terrain brief to measured result.
Terrain, audience, message, delivery and proof, the technology underneath, and how a program goes on air in fourteen days. Each movement ends in a number we can defend, measured against a control we drew ourselves on the fifth day.
Precincts of an invented county, shaded by persuadable votes in reachIllustrative
Every race has a small number of places and a small number of weeks that decide it. Before anything is built we map both: the precinct-level margin arithmetic, the persuasion window relative to the absentee and early-vote calendar, the opponent's probable lines of attack from the record, and the money that will arrive against you and when.
The output is a plan that says where the votes are, when they are reachable, and what it would cost to reach them, and, just as usefully, where they are not. A terrain brief runs to about twenty pages and is delivered inside the first seventy-two hours of an engagement.
Labor Day
Absentee ballots mailed
Early voting opens
Election Day
Persuasion window
Mobilization
Week 0
Week 7
Week 9
Exhibit 1The general-election runway. Persuasion spend that lands after the absentee drop reaches a shrinking audience; the plan is sequenced against this calendar before it is sequenced against anything else.Exhibit 2The county board. Six counties of an abstract state: last cycle's margin on the left, the modeled pool of persuadable, likely voters that neither side is reaching on the right. The plan concentrates where the second bar is long and the first is short. Illustrative.
02
Movement II of V · Day 3
Audience
Who can still be moved, read from what people do.
An invented electorate of forty-eight thousand: who is persuadable, and who is held outIllustrative
We identify deciders by what they do rather than by what a file says they are. Behavioral models score persuadability, turnout propensity and receptivity to particular arguments; individuals are then grouped by what would move them, not by demographic proximity.
Segments are small, named and described as people, so the strategist and the creative team are working from the same picture. Equally important is the group we identify in order to leave alone: the voters for whom contact measurably backfires. Move the budget slider: the point of modeling is that most of the available lift sits in a small fraction of the population.
Exhibit 3The uplift quadrant, with a budget. Each mark is a modeled individual. Drag the slider to contact people in order of expected net votes; the readout shows how much of the available uplift a given share of the population captures. Illustrative population of 170.
The segment mapExhibit 4
What the segment map records for each cohort.
Every cohort in a map carries the same dossier. How it was assembled from the electorate and how sharply it was separated from its neighbors; what moves it, in four registers; where it is reachable, with matched counts per channel; what to say to it and what never to cross-run; and the instrumentation underneath. Choose a cohort. Every figure is illustrative of a statewide general-election program; the structure is exactly what a client receives.
Exhibit 4Cohort dossiers from a segment map. Composites from the behavioral lexicon; figures illustrative. The assembly rail shows scale and discrimination in the same glance: from the state's registered electorate to the people reachable today, with not one attribute the campaign already had.
03
Movement III of V · Day 7
Message
One sentence each cohort can hold, in an order that survives the attack.
Seven steps of one narrative track; the third answers the attack before it landsIllustrative
Most campaigns write a message and then look for people to send it to. We work the other way around. The segment map tells us what each cohort is already anxious about, proud of and waiting to hear; the narrative track is built from that, one line per step, sequenced so that the attack is answered before it lands and the ask comes last.
Below, the same seven steps for three cohorts from one statewide program: three tracks, one candidate, and no line that any other cohort would have received. Each step carries its basis, the disposition or signal family that justifies the line, because a line without a basis is a slogan. The studio then produces the variants, forty to sixty across three registers for a statewide race, and every one carries its disclosure.
Exhibit 5One narrative track, seven steps, three cohorts. Choose a cohort: the header carries what she is listening for, the attack she will hear, her validators and her register mix; each step carries its line, its basis, its register and its week. Lines are illustrative masters written for this page; every real line carries a disclosure.
How the studio turns one approved master into cohort variants.
The studio does not write from a blank page. It takes in what the campaign already has (the stump speech, the approved ads, the policy pages, the endorsements, the record) and the brand guidelines that govern how the candidate sounds and looks. It decomposes that material into claims, each with its proof attached; matches each claim to the cohorts whose dispositions it speaks to; assembles a variant for each cohort in each register from one approved master; and runs every variant through four checks before a person signs it. What comes out is a tailored creative for a specific audience, built from the campaign's own material, in the campaign's own voice, carrying its disclosure. What the holdout says about it decides what runs next.
Exhibit 6The studio engine. Left, what the campaign supplies: brand guidelines, existing content, one approved master brief. Center, what the engine does: decompose into claims with proof, match claims to cohort dispositions, assemble by register, check four ways. Right, the variants a cohort receives. Choose a cohort to see the match and the output change. Lines are illustrative masters written for this page.
04
Movement IV of V · Day 12
Delivery
Three registers from one origin, with every placement accounted for.
Flights by register across a fourteen-week programIllustrative
Three registers, run as one program. Precision paid media, placed at the person rather than the zip code, across social, search, connected television, programmatic and mail. Organic advocacy (validators, relational networks, the letters page) that puts the message in a second voice. And the ambient campaign: out-of-home, a stunt calendar and earned media, so that the argument seems to come from the world rather than from the candidate.
Every register is disclosed. None of it is delivered by undisclosed personas, and none of it is bought through anyone's accounts but our own, so that every dollar can be traced from invoice to impression to the holdout that judged it.
Exhibit 7Single-origin delivery, as a loop. Every placement originates from a Brushworks account of record and flows out through six channels to the field; what the field does is measured against the holdout and fed back into the seats (reflighting, suppression, retraining) every day. The audit trail runs unbroken around the loop.
05
Movement V of V · Day 5
Proof
A control drawn on day five, and a number we can defend.
Treated against the sealed control, by week, with the intervalIllustrative
We hold out a randomized control from every program and measure against it daily. A model is not deployed unless it beats a demographic baseline and a look-alike baseline on the client's own holdout; models that fail are logged and killed, not quietly retired.
The hardest client is the one already on air. An incumbent's campaign in week four has a general consultant, a mail firm, a digital shop and a pollster, and no obvious reason to add anyone. The simulation below is that campaign. Its own program is running and working. In week four we layer in three things it cannot do for itself: the neglected persuadables its file has never contacted, the hold list that stops it paying to move likely voters the wrong way, and a sealed control that tells it what any of it did. Run it and watch the gap open.
Exhibit 8An established campaign, joined in week four. Top, the registers flighting after the join (the ochre column is the absentee drop). Below, favorability among treated voters: the campaign's own program alone, the campaign with the Brushworks layer, and a randomized control that saw neither. The gap between the two upper lines is our incremental lift; the gap to the dashed line is the campaign's. Illustrative shape; real programs are reported with confidence intervals.
TechnologyExhibits 9 and 10
The platform is built and run in-house.
Everything in the method runs on a platform we built and operate ourselves: identity, graph, registry, models, studio, seats and measurement. Nothing is rented from a vendor who also sells it to the other side. Below is the funnel a client's program passes through, stage by stage; open any stage for what happens inside it, at what scale, and under what safeguards. Enough to evaluate it, not enough to copy it.
Exhibit 9The funnel, in eight stages. Select a stage to open it. The catalog is the gate: nothing passes it without beating two baselines on the client's own holdout, and the measurement stage feeds back into it every day a program is in field.
Anatomy of a scoreExhibit 10
How a persuadability score is built, trained and allowed to ship.
A score comes from a model trained on what people like this actually did in prior programs (responded, signed up, turned out, moved) validated against two baselines on the client's own holdout, labeled with its confidence, and monitored for drift every day it is in use. The five panels below are the construction of one model in the catalog; the feature families and their weights are illustrative, the process is not.
Exhibit 10Anatomy of a score. Inputs, feature engineering, training labels, validation at the gate, and what leaves the model. Feature weights illustrative.
The graph
Sixty-two billion observed behaviors resolved to more than 270 million adults and their households, rebuilt in full every month; digital activity refreshed daily.
The registry and the dictionary
Seventeen thousand active signals against a ten-thousand-topic dictionary in twenty-five domains, curated toward what has predicted political decisions and pruned when it stops.
The catalog and the gate
Five model families trained on real outcomes. Nothing ships without beating a demographic and a look-alike baseline on the client's own holdout; failures are logged with the reason.
The studio
Cohort-specific variants generated in minutes from one approved master and the candidate's own assets, each carrying its disclosure. Forty to sixty per statewide program.
The seats
Every placement from Brushworks accounts of record across social, search, connected television, programmatic and mail. Single-origin, so a program can be audited from invoice to impression.
The desk
Seven offices across fourteen hours of the clock, watching every program in the field; the daily loop runs wherever it is morning.
ImplementationExhibits 11 and 12
Data boundaries first, then the fourteen-day calendar.
An engagement is implemented as a set of boundaries and a calendar. The boundaries say what data can touch what: a client's file enters under a written data-use agreement, is matched and scored inside Brushworks, is never combined with another client's, and leaves only as hashed audiences through our own seats. The calendar says when: fourteen days from signature to first flight, with the holdout sealed on day five.
Exhibit 11Engagement architecture. Client data enters under agreement, is matched and scored inside the firm's boundary, and leaves only as hashed audiences through single-origin seats. The randomized holdout sits outside the delivery path and reports to the client, not through us.
Day 1Terrain brief
Margin arithmetic, calendar, attack lines, money against you.
Day 3Segment map
Deciders scored and grouped; the lexicon drafted.
Day 5Holdout sealed
Randomized control drawn and locked before any spend.
First flight live; the daily loop begins; weekly field reports thereafter.
Exhibit 12The ignition sequence. The same fourteen days for every program; only the terrain changes.
What you receive
The six documents a client receives.
Terrain brief
Where the race is decided, when, and at what cost. Precinct arithmetic, calendar, opposition file.
Day 1
Segment map
Every decider segment scored, sized and described as a person, with the track assigned to each.
Day 3
Message architecture
The narrative tracks, the seven-step sequencing and the master creative briefs, with disclosures.
Day 7
Flighting plan
Three registers across the calendar, by week, by segment, by seat, with the absentee drop marked.
Day 12
Weekly field report
What ran, what it reached, what moved against the holdout, and what changes next week: four pages, every Monday.
Weekly
Proof report
The difference the program made, measured against the control, with confidence intervals and the model kill log.
Close
ProgramsWhat the method runs inside
Candidate campaigns
Federal, statewide and local. We take a small number of races each cycle and staff each with a full team, from terrain through proof.
Independent expenditures and donor programs
Donor-funded committees in which Brushworks controls budget and execution and a contributors' advisory board approves strategy, creative masters and media plans. Formed, firewalled and reported under federal coordination rules from day one, with sign-off gates and a ten-percent reallocation threshold written into the charter.
Ballot measures
A question whose text runs to dozens of pages, reduced to the one sentence each voter needs, without misrepresenting the law. For and against.
Bonds and levies
Asking taxpayers to raise their own taxes, precisely: the school, the roof, the bus route, the number. More than a hundred measures passed.
Governments and institutions abroad
Engagement, security and narrative programs with local partners, registered where the law requires it, in theaters where the ordinary playbook does not survive contact.
Performance assurance
For qualifying programs, Brushworks backs its fees with an underwritten performance assurance. It attaches only to media we place ourselves. Terms on request.
What we know about people, and how it becomes a decision.
Sixty-two billion observed behaviors, a ten-thousand-topic dictionary, a publish gate that kills more models than it ships, and three things we refuse to model at all.
S-01 · composite · The StewardIllustrative
1What a person does
Attends the school board budget session
Reads the county budget story twice
Renews the same home insurer, eleventh year
Opens the property tax notice the day it arrives
Signs a petition from the library board
2What it scores
Attention to local taxes0.91
Trust in institutions0.78
Persuasion0.58
Race, religion, healthnever
3 · What we writeA line for her, measured against the holdout.
SignalsExhibit 1
Sixty-two billion observed behaviors, each tied to an adult or a household.
A signal is an observed behavior (a purchase, a subscription, a search, a visit, a renewal, a cancellation) resolved to an adult or a household, with a topic, a date and a source. It is not an opinion, not a survey answer, and not an inference about who someone is.
We hold roughly sixty-two billion of them, resolved to more than 270 million American adults and their households, and organize them in a dictionary of about ten thousand topics across twenty-five domains. Some seventeen thousand signals are live in the registry at any time. The behavioral graph is rebuilt monthly; digital activity is refreshed daily.
Exhibit 1The active registry, to scale by domain. Ochre marks signals that entered a scored model this month. Grouping illustrative; the underlying taxonomy has twenty-five categories.
ResolutionExhibit 2
How a raw event is attached to a person.
Raw events arrive as a stream: a hashed identifier, a device, a source, a timestamp. Each passes two gates. Identity resolution attaches it to an adult or, where the source is household-level, to the household; the dictionary attaches one of ten thousand topics in twenty-five domains and one of five signal families. Only then does it reach the person on the right, whose scores, the ones this program cares about, move with every arrival and keep a history.
Nothing in the stream is an opinion. The scores are probabilities between zero and one, they are revised every day the program is in field, and none of them describes who she is, only what people like her do.
Exhibit 2Signal resolution, live. Events enter from the left, color-coded by domain (commerce, media, civic, household, health and work) pass identity resolution and the dictionary, and reach one composite decider or her household. Her six scores update on arrival and keep a history. Labels are examples of signal types, not a real person's record.
The stackExhibit 3
The seven layers of the platform.
The platform is built as a stack, and every layer is ours to operate: identity at the bottom, delivery at the top, measurement running through all of it. Nothing is rented from a vendor who also sells it to the other side, and nothing is placed through a seat we do not control.
7
Delivery, through our own seats
Social, search, connected television, programmatic and mail, bought from Brushworks accounts of record; relational and earned channels alongside. Single-origin, so every dollar is traceable.
Every flight
6
Creative variant engine
Cohort-specific variants generated from one approved master brief and the candidate's own assets, each carrying its disclosure. Forty to sixty per statewide program.
Minutes per master
5
Narrative tracks and playbooks
An authored library of message sequences, each mapped to the seven-step arc and to the dispositions that trigger it.
Authored, versioned
4
Model catalog
Five families (behavioral enrichment, dispositions, occupational, interests, affinities) trained on real outcomes and kept small by the publish gate.
Gated per client
3
Signal registry
Seventeen thousand active signals against a ten-thousand-topic dictionary; curated toward what has predicted political decisions before.
Pruned monthly
2
Behavioral graph
Sixty-two billion resolved events over 270 million adults and their households. Household-attributed where the source is; rebuilt in full each month.
Monthly rebuild, daily activity
1
Identity resolution
Individuals and households, matched on hashed identifiers; client files matched under written data-use agreements and never commingled.
Hashed activation
Exhibit 3The stack, bottom to top. Measurement, the randomized holdout, sits outside the stack and judges it.
From signals to decisionsExhibit 4
Models are trained on real outcomes and must beat two baselines.
Signals become useful when a model turns them into a score: how persuadable this person is, how likely to vote, how receptive to a particular argument. Our models are trained on real outcomes from prior programs (votes, sign-ups, responses) not on proxies.
Every candidate model faces the same gate. It must beat a demographic baseline (the file, used well) and a look-alike baseline (the platforms' own targeting) on the client's own holdout. If it cannot, it is logged with the reason and killed. The catalog stays small because the gate is honest.
Candidate modelTrained on outcomes from prior programs
Beats the file?Demographic baseline, used well
Beats the platforms?Look-alike baseline on the same task
On your holdout?Measured on the client's own control, not ours
Ships.
Exhibit 4The publish gate. A model that fails any of the three tests is logged with its kill reason. Nothing is quietly retired.
The loopExhibit 5
The daily loop: observe, score, deliver, measure and retrain.
Behavior observed today changes tomorrow's scores; tomorrow's scores change what is delivered; what is delivered is measured against the holdout; and the measurement retrains the model. The cycle runs every day the program is in field.
The practical consequence is that a program gets better as it runs. The first week is the weakest, because the models have learned the least.
Exhibit 5The daily loop, live. Each station reports what it did in the last cycle as the marker passes; the holdout sits outside the cycle so that it can judge it. Figures illustrative of a statewide program in week four.
One deciderExhibit 6
A composite voter, as the platform records her.
This is the Night Nurse from the lexicon: a composite, drawn from several thousand behavioral neighbors. The inner ring is what was observed; the middle ring is what the models made of it; the outer ring is the single narrative track she was assigned, and the register she received it in.
Notice what is absent. Nothing about her ethnicity, faith or party is modeled, and none of it was needed.
Exhibit 6One composite decider. Inner ring, observed behaviors; middle ring, scored dispositions; outer ring, the assigned narrative track. Composite, not an individual.
Governance of the human layerExhibit 7
What we refuse to model, and who checks.
Ethnicity, religion and ideology are never modeled, never bought, and never accepted as a filter from a client: partly on principle, mostly on accuracy, because identity labels are a poor proxy for what a person will do and a campaign that targets on them is speaking to a stereotype rather than a voter. That refusal is the first line of a longer ledger. Every control on the human layer has an owner inside the firm and a party outside it who can verify it.
The Steward and the Night Nurse can be any color and any faith. What they respond to is what they respond to.
Control
What it does
Who verifies
Identity exclusion
Ethnicity, religion and ideology are absent from the registry, the models and every targeting filter; a request to add them ends the engagement.
Client audit under the data-use policy; counsel of record
The publish gate
No model ships unless it beats a demographic baseline and a look-alike baseline on the client's own holdout.
Reported in the proof report with the baselines
The kill log
Every model that fails the gate or drifts in field is logged with the reason and retired; nothing is quietly kept.
Delivered to the client at close; aggregate published annually
Human review
Every creative master is read by a person before a single variant is produced; every variant carries its disclosure.
Studio lead signs the master; compliance review on day ten
Program scoping
A client's file is matched for that program alone, never commingled, and leaves only as hashed audiences.
Certified destruction at close; audit rights
The sealed holdout
Drawn on day five, before any spend, and never contacted; the only basis on which lift is reported.
Reports to the client directly
Practice walls
Candidate and independent-expenditure work run on separate teams, systems and seats.
Certification available to any committee or counsel
Retention
Program data destroyed on a written schedule after close; graph rebuilt monthly, so nothing persists by accident.
Certification on request; data protection lead
Exhibit 7The governance ledger of the human layer. Each control is written into the engagement letter and the data-use policy.
Provenance and security
Where the data comes from, and the limits on its use.
Licensed sources. Signals are licensed from established providers under written agreements and handled under our data-use policy. Nothing is scraped, and nothing is bought from a broker who cannot say where it came from.
No commingling. A client's file is matched to the graph for that client's program and for nothing else. Two clients never share a model, a seat or a segment.
Hashed activation. Audiences leave the platform as hashed identifiers, matched inside the delivery platforms; raw personal data never travels.
Retention limits. Client files and program data are destroyed on a written schedule after the engagement closes, with certification.
Independent reporting. Independent security reporting is available under agreement. Where a mandate requires it, we work inside the client's own environment.
Written walls. The candidate practice and the independent-expenditure practice are separate teams, separate systems and separate seats, documented and certifiable.
Arranged by method. Names are held under agreement; the terrain, the move and the margin are in the open.
The record, one tile per contestIllustrative layout
WonLost, and publishedWritten up here: open one
96%of programs won
500+contests over a decade
9cases written up here
4%lost, and published
The record on the map
Where the record was set.
Hover a pin or a case; click to open its file. A state is shown where the case file names it. Where a client’s states are held under agreement, the map says so instead of guessing.
Find a case.
Each case opens as its own file: the terrain, the move, the result, the exhibit, and the papers behind it.
Every contest we have designed a program for, counted here in a hundred. Results are verified against public returns and, where a holdout existed, against the control. Written references are available to qualified prospects under agreement.
Exhibit 1More than five hundred contests, shown as a hundred. Filled: won. Hollow: lost.
Office
Method
Note
Presidential
A national argument, rewritten by county
A 2016 national campaign. The same argument, rewritten for the counties that decided the electoral map, in the vocabulary each county already used.
Gubernatorial
Behavioral targeting across three distinct strategies
Three competitive states with three different theories of the race (a challenger, an incumbent, a coalition rebuild) run on one measurement standard.
U.S. Senate
The neglect algorithm
Persuadable, likely voters that both parties had stopped contacting, identified by cross-referencing contact history with receptivity scores.
U.S. House
Small-dollar lists that last
Small-dollar and volunteer growth engineered from behavioral affinity rather than purchased lists; national-profile members on both sides of the aisle.
Ballot measures
Defeating a 65-percent measure
A 2022 California measure with 65 percent support at launch and nine figures of spending behind it, defeated with 17.7 percent of the vote.
Bonds and levies
Naming the repairs
More than a hundred school bond and levy measures; more than five billion dollars approved by voters raising their own taxes.
Committees and NGOs
The permanent campaign
Trigger-based activation between cycles: the audience is kept warm by responding to events, not by a calendar.
Abroad
Institutional and diaspora programs
Twelve countries; security, engagement and narrative mandates for governments and institutions, and a five-country diaspora program.
Exhibit 2The record, by office.
Case 1Ballot measure, California, 2022
Defeating a 65-percent measure.
The terrain
An online sports-betting proposition with sixty-five percent support in the campaign's baseline polling and one of the largest advertising budgets ever assembled behind a measure. The text ran to dozens of pages. Almost no one would read it.
The move
The measure was distilled into single-sentence frames, each true, each chosen for the person reading it by behavioral profile: a family-finance frame for one household, an economic-impact frame for another, a local-revenue frame for a third. Forty-one variants across three registers.
The result
Defeated with 17.7 percent of the vote on November 8, 2022 (the official statewide result, a swing of more than forty-five points from the baseline) with a sixty-one percent increase in favorable sentiment toward the opposition's position on social platforms over the same period.
Exhibit 3Support from launch to result. The launch figure is the campaign's baseline polling; the result is the official statewide vote (17.7 percent yes). The trajectory is drawn to scale between them.
The case in full
4 min read · 3 exhibits
The situation
The measure would have legalized online and mobile sports betting across California. It was one of two sports-betting measures on the November 2022 ballot, and together they drew the most money ever raised for a ballot measure contest in the United States: about $463 million, of which $301 million was raised in support.1
The measure’s text ran to dozens of pages. The campaign’s baseline polling put support at 65 percent. The question was how to move a large, favorable electorate that would never read the text, against a budget measured in nine figures.
The method, step by step
Decompose the measureEvery consequence the text implies is listed and checked.
Write one sentence per consequenceEach consequence becomes a single true sentence a voter can hold and check. Here the frames were family finance, economic impact and local revenue.
Match the sentence to the readerBehavioral profiles, never identity fields, decide which frame each household receives. Forty-one variants were written across three registers.
Test against a sealed holdoutEach frame runs to a slice of its cohort and is kept or retired on measured movement against the control.
Retire what fadesA frame whose effect fades in the field is replaced, so the last sentence a voter hears is the one most likely to hold.
How it was measured
Support was read from the campaign’s own polling and, at the end, from the official statewide result. Social sentiment toward the opposition’s position was tracked over the same period and rose 61 percent. The record does not claim the program alone moved the measure from 65 to 17.7. It claims the frames were tested, kept on evidence and delivered to the people they moved.
What it means for a race now
On a measure, the side that says no starts with the voter’s uncertainty on its side (WP-34).
A large favorable margin at launch is a measure of unfamiliarity as much as support.
Money buys reach; the sentence decides what the reach is worth.
Money and the result. Share of the $463 million raised for the two sports-betting measures that supported them, and the yes vote on this one.Which single-sentence frame moved which cohort, measured against the holdout. Illustrative of the matrix the program kept.The method from the measure to one sentence per voter.
The Brushworks Review, WP-03, Defeating a measure that launched at 65 percent.
Case 2U.S. Senate and gubernatorial, competitive states
The neglect algorithm.
The terrain
Both parties keep contact histories, and both, over time, stop calling the same people: the voters who did not answer, did not give, did not sign. Some have moved or died. Many are persuadable people the lists have quietly dropped.
The move
Contact history was cross-referenced against behavioral receptivity and turnout propensity, surfacing tens of thousands of voters who scored high on both and had not been contacted by anyone in years. They were cheaper to reach, more open when reached, and invisible to polling that treated them as noise.
The result
In the competitive statewide races where the method has been deployed, margins in written-off counties exceeded every model that ignored them. Use the layers below to see how the population is found.
Exhibit 4An abstract territory of 448 precinct-sized cells, each about two thousand voters. Toggle the layers: who has been contacted in three cycles, how receptive each cell scores, and where the neglected persuadables are. Illustrative data; the method is real.
The case in full
3 min read · 3 exhibits
The situation
Both parties keep contact histories, and over time both stop calling the same people: voters who did not answer, did not give, did not sign. The lists drop them quietly. Some have moved or died. Many are persuadable, likely to vote, and have not heard from anyone in years.
In competitive statewide races for the Senate and for governor, those voters were concentrated in counties the campaigns had written off on last cycle’s margin.
The method, step by step
Assemble the contact historyWho has been reached, by whom and how often, across recent cycles.
Score receptivity from behaviorEach voter is scored on how open they are to persuasion, from what they do, with turnout propensity scored separately.
Find the intersectionVoters high on both scores and absent from every recent contact list are the neglected persuadables. In these states they numbered in the tens of thousands.
Reach them first, against a controlContact goes to the neglected before the contested middle, with a sealed holdout to count what it does.
Re-read the mapCounties are ranked on the size of their unreached pool as well as on last cycle’s margin (WP-18).
How it was measured
The result was read in the counties the campaigns had written off, where margins exceeded every model that ignored the neglected. The neglected cost less to reach and were more open when reached.
What it means for a race now
The cheapest persuadable voter in a state is often the one no one has called.
Rank counties by their unreached pool, not by last cycle’s margin (WP-18).
Polling that treats non-responders as noise misses the same people the lists dropped.
Registered voters by campaign contact history in three competitive states. Illustrative.From the electorate to the neglected persuadables.Cost per net vote by contact history, indexed. Illustrative of the pattern the program found.
Sources
The Brushworks Review, WP-01, The neglect algorithm; WP-18, What a county board knows.
Kalla, J. L., and Broockman, D. E. (2018). The minimal persuasive effects of campaign contact in general elections. American Political Science Review, 112(1), 148–166.
Case 3Bonds and levies, more than a hundred measures nationwide
Naming the repairs.
The terrain
Convincing taxpayers to raise their own taxes, in districts where the previous attempt had often failed, against organized opposition and general fatigue with the word "education."
The move
The conversation was moved from the abstraction to the object: the roof that leaks, the science lab from 1971, the bus route that was cut: improvements a voter can see from the road. Spend concentrated on the households most likely to be moved by them; hostile segments left untouched.
The result
More than five billion dollars in community investment approved across more than a hundred measures, with margins up to fifty-seven points.
Exhibit 5Five selected measures. District names are available under agreement.
The case in full
3 min read · 3 exhibits
The situation
School bonds ask voters to raise their own taxes. Nationally they usually pass: between 2015 and 2022 voters approved 79 percent of school bond issues, and in 2022 districts proposed 2,146 issues worth $123 billion.1 The districts in this record were the hard ones, where the previous attempt had often failed, against organized opposition and fatigue with the word “education.”
Many states set the bar above a majority. California requires 55 percent for school bonds; Idaho requires two thirds, and districts there passed 44 percent of bond elections where a simple majority would have passed 62 percent.1
The method, step by step
Walk the buildingsThe program starts with the facilities list: the roof that leaks, the science lab from 1971, the bus route that was cut.
Name the objectEach improvement becomes a sentence about something a voter can see, close to where they live.
State the cost per householdThe monthly cost to a median household is stated plainly and early, before the opposition states it for them.
Match the repair to the householdFamilies with children in the district, homeowners without them and long-time residents hear about different repairs.
Choose the dateWhere a district can choose the election date, it chooses knowing who votes in that month (WP-17).
How it was measured
Results are verified against the certified returns for each measure. Across more than a hundred measures, voters approved more than five billion dollars.
What it means for a race now
Voters approve what they can picture. Name the repair.
Say the cost first. A number the opposition announces becomes their argument.
The election date is part of the campaign.
How school bonds fare. National and state approval from the public record.Support for a bond measure by the specificity of the frame received, treated minus holdout. Illustrative.Passage rate of school bond and levy measures by election month. Illustrative of the pattern in the record.
Sources
Education Week, June 2023, citing a national database of school bond elections and ProPublica and the Idaho Statesman on Idaho.
The Brushworks Review, WP-09, School bonds won by naming the repairs; WP-17, The bond calendar.
Case 4Presidential, 2016
A national argument, rewritten by county.
The terrain
A national campaign with one argument and an electoral map decided by a few dozen counties that did not experience that argument the same way. A county that lost a plant and a county that gained one heard the same sentence and understood different things.
The move
The national narrative was rewritten at the county level, in the vocabulary each place already used (its employers, its landmarks, its grievances) and delivered as if it had been written there.
The result
Won. The states that decided the map were carried on margins that the national polling had not seen coming, in counties the national creative had not been speaking to.
A county that lost a plantThe jobs left. The promise didn't have to.
A county that gained oneKeep what's working. Fix what isn't.
A retiree countyYour check, your clinic, your county.
A county with a borderOrder at the line. Opportunity behind it.
Exhibit 6The same argument, four vocabularies. Illustrative lines; the method is the point.
The case in full
3 min read · 3 exhibits
The situation
The 2016 presidential election was decided in a small number of places. Ballotpedia identified 206 counties that voted for Barack Obama in 2008 and 2012 and for Donald Trump in 2016. They cast 5.5 percent of the national vote and accounted for about half of the Republican net gain from 2012.1
A national argument does not land the same way in a county that lost a plant and in a county that gained a warehouse. The campaign had one argument and many counties.
The method, step by step
Find the deciding countiesCounties are ranked by how much of the electoral map they can move and how open their voters are to the argument.
Read each county’s vocabularyLocal news, local employers, local grievances and the words residents already use for them.
Rewrite the example, keep the claimThe same argument is rewritten county by county in that vocabulary.
Carry it locallyEach version runs where that county reads and listens, from the campaign’s own disclosed accounts.
Measure by countyLift is read county by county against holdouts, and versions that do not move their county are retired.
How it was measured
Lift is read county by county against holdouts and set beside the national frame (the second exhibit). The record’s result: the states that decided the map were carried on margins in counties the national polling had not been speaking to.
What it means for a race now
A national message lands locally or not at all (WP-15).
Rewrite the example, keep the claim.
Small, concentrated movement still decides national elections (WP-33).
The counties that decided 2016. A small share of the vote, a large share of the change. Public record.Favorability lift by county, treated minus holdout, with the national frame for comparison.Which frame moved which county, share of available lift captured. Illustrative.
The Brushworks Review, WP-15, A national argument, rewritten by county; WP-33, Sixteen districts.
Case 5Gubernatorial, three competitive states
Three governors’ races in one cycle.
The terrain
Three governors' races in the same cycle with three different problems: a challenger nobody knew, an incumbent nobody could name anything about, and a coalition that had come apart.
The move
Three different theories of the race, three different segment maps, three different register mixes: all measured to one standard, against holdouts drawn on day five.
The result
All three won.
The challenger
The incumbent
The coalition rebuild
Theory of the race
Introduce a person before the opponent defines him
Attach a record to a name that sixty percent of voters could not attach anything to
Re-earn the segments that had drifted, one reason at a time
Audience emphasis
Late deciders and low-information likely voters
Soft supporters with high turnout propensity
Former supporters scored as recoverable, not lost
Register mix
Paid-heavy early, then validators
Earned and ambient first, paid proof second
Relational and organic, with precision paid in the last three weeks
Result
Won
Won
Won
Exhibit 7Three strategies side by side. The measurement standard was the only thing they shared.
The case in full
3 min read · 2 exhibits
The situation
Three governors’ races ran in the same cycle with three different problems: a challenger nobody knew, an incumbent nobody could name anything about, and a coalition that had come apart.
Each needed a different theory of the race. The firm’s rule was that each would be measured the same way.
The method, step by step
Write the theory of each raceIntroduce a person before the opponent defines him; attach a record to a name; re-earn the segments that had drifted, one reason at a time.
Draw three segment mapsLate deciders and low-information likely voters for the first; soft supporters with high turnout propensity for the second; former supporters scored as recoverable for the third.
Mix the registers differentlyPaid-heavy early for the challenger; earned and ambient first for the incumbent; relational and organic for the rebuild, with precision paid in the last three weeks.
Seal the holdouts on day fiveEach race drew its controls on day five of the engagement (WP-27).
Read all three on one pageOne reporting standard, weekly, with intervals, so that the three programs could be compared and corrected.
How it was measured
All three were measured against holdouts drawn on day five, with lift reported by cohort and interval. All three races were won.
What it means for a race now
The strategy should differ by race. The measurement should not.
A control drawn early is the only way to compare programs that look nothing alike (WP-06).
Write the theory of the race before the segment map.
Lift against the sealed control by cohort, with 95 percent intervals: the reporting format all three programs shared.The first fourteen days, from engagement to first flight, with the control drawn on day five.
Sources
The Brushworks Review, WP-06, A standard of proof; WP-27, Fourteen days; WP-25, The second-month test.
Case 6U.S. House, national-profile members, both parties
Small-dollar lists that last.
The terrain
House members with national followings and local districts, dependent on small-dollar donors and volunteers who were being rented from purchased lists, harvested once, and lost.
The move
Donors and volunteers were found by behavioral affinity rather than list purchase, and grown along a curve (recognition, concern, permission, ask) rather than hit with the ask on first contact.
The result
Small-dollar and volunteer bases that compounded across a cycle instead of decaying, for members on both sides of the aisle.
Exhibit 8Active small-dollar donors over twelve weeks: affinity-built base against a purchased list of the same starting size. Illustrative shape.
The case in full
3 min read · 2 exhibits
The situation
House members with national followings and local districts depended on small-dollar donors and volunteers. Their lists had been rented, purchased or harvested in a single burst, and decayed as fast as they were built.
The members came from both parties. The problem was the same: lists that lasted one cycle.
The method, step by step
Find affinity in behaviorProspective donors and volunteers are identified by what they do (causes they act on, communities they belong to) instead of by purchased lists.
RecognitionThe first contact introduces the member and the work, with no ask.
ConcernThe next contacts are about what the person already cares about, in the member’s own words.
PermissionThe person chooses to hear more: a subscription, an event, a small action.
The askOnly then the ask, sized to what the person has already shown they will do. Contact between asks follows events in the member’s work (WP-16).
How it was measured
The yardstick is retention: who is still giving and volunteering months later. In the record, the affinity-built bases compounded across the cycle for members of both parties, where purchased lists are harvested once and lost.
What it means for a race now
A list you rent is gone by the next cycle.
Ask for what a person already wants to do.
Measure a list a year after it was built.
Share of an audience still responsive by months since last meaningful contact. Illustrative.Signal families defining one composite cohort, by count of active signals: the kind of record affinity is read from.
Sources
The Brushworks Review, WP-16, The permanent campaign; WP-05, Behavior, not identity; WP-14, What a signal is.
Case 7Committees and NGOs
The permanent campaign.
The terrain
Organizations that exist between elections, whose audiences go cold between them, and whose fundraising and activation spike only when a headline does the work for them.
The move
Activation was tied to triggers rather than to the calendar (a ruling, a vote, a headline) with pre-built creative and pre-scored lists, so that response is measured in hours, not weeks.
The result
Audiences that stay warm, and activation curves that rise with each trigger instead of resetting to zero.
Exhibit 9Engagement index across a year with four triggers marked. The dashed line is the same audience under a calendar-driven program. Illustrative shape.
The case in full
3 min read · 3 exhibits
The situation
Committees and NGOs live between elections. Their audiences go cold on the calendar: a newsletter in March, an appeal in June, silence until the next crisis.
The organizations in this record needed their audiences ready when something happened, not when the calendar said to write.
The method, step by step
Define the triggersEvents that matter to the audience (a vote, a ruling, a headline) are written down in advance, with the cohort each one touches.
Pre-build the responseCreative, approvals and pre-scored lists are ready for each trigger before it fires.
Watch for the eventA standing desk watches for each trigger and confirms it.
Activate in hoursWhen a trigger fires, the right cohort hears from the organization within hours.
Measure every activationEach activation is measured against a holdout, and triggers that do not move people are retired.
How it was measured
Trigger-to-activation time is measured in hours. The record’s result: audiences that stay warm, with activation curves that rise with each trigger instead of resetting to zero (the first exhibit shows the shape, illustratively).
What it means for a race now
An audience stays warm when you speak to it about what just happened.
Prepare the response before the event.
Retire triggers that do not move people.
Engagement across a year for a trigger-based program and a calendar-driven one. Illustrative.The trigger loop.A desk’s clock, in minutes from the headline.
Sources
The Brushworks Review, WP-16, The permanent campaign; WP-23, Two hours, five markets.
Case 8Diaspora engagement, five countries
A diaspora in five countries.
The terrain
A ten-million-person global community across the United States, the United Kingdom, Canada, the UAE and Australia, during a period of acute geopolitical volatility that was fracturing sentiment toward the homeland in every one of them at once.
The move
The diaspora was treated as an electorate: segmented by behavior rather than by origin, reached through community leaders identified market by market, with sustainable channels to homeland institutions and a crisis-communications framework for rapid response.
The result
Sentiment stabilized across five markets, durable channels in place, and a playbook the institution still runs.
Market
What the community was listening for
Register emphasis
Local partner
United States
Standing at work, standing at school; the second generation's story
Organic validators, relational
Community and professional associations
United Kingdom
Belonging in a hard news cycle
Earned, then precision paid
Regional community leadership
Canada
Bilateral strain felt at the family level
Relational, crisis desk
Faith and civic institutions
UAE
Remittance, return, reputation
Precision paid, ambient
Business councils
Australia
Distance, and what it does to loyalty
Organic, relational
Student and alumni networks
Exhibit 10Five markets, one program. Partner types are given by category.
The case in full
3 min read · 3 exhibits
The situation
A ten-million-person community across the United States, the United Kingdom, Canada, the UAE and Australia carried a homeland’s story into other people’s politics, during a period of acute geopolitical volatility that was fracturing sentiment toward the homeland in all five at once.
The mandate was engagement in the open, under the firm’s rules for institutional work abroad: registered where the law requires, disclosed on every placement (WP-28).
The method, step by step
Pass the gates firstLegal review in each host country, registration where required and disclosure rules set before any contact.
Treat the diaspora as an electorateSegment by behavior, not by origin: who reads which media, in which language, through which community institutions.
Find the partners market by marketCommunity and professional associations, faith and civic institutions, business councils, student and alumni networks, by category.
Build the channels and the deskDurable channels to homeland institutions, and a crisis-communications framework for rapid response (WP-23).
Read sentiment across marketsSentiment is tracked in all five markets through volatile periods.
How it was measured
The record’s result: sentiment stabilized across five markets, durable channels were put in place, and the institution still runs the playbook.
What it means for a race now
A diaspora is a constituency with its own media and its own institutions (WP-08).
Speed matters only if the answer is disclosed and true.
Register before the work begins.
The diaspora by host country, share of the community. Rounded.Favorable sentiment across five markets through a volatile period. Illustrative.First-response time by incident through a season, minutes from headline to first placement.
Sources
The Brushworks Review, WP-08, The diaspora is a constituency; WP-23, Two hours, five markets; WP-28, Persuasion in the open.
Case 9Security and behavioral detection, the Gulf
Two million pilgrims.
The terrain
The world's largest recurring human gathering: some two million people a year moving through a small number of sites on a fixed calendar, with security that had to be rigorous and could not be seen to profile.
The move
Automated detection frameworks built on behavior (density, movement, access patterns) rather than on identity, layered from the approach roads to the sites themselves, with messaging adapted for Arabic-speaking, South Asian and Southeast Asian pilgrims.
The result
A security posture that protected the crowd without profiling the person, and a framework the ministry runs each year.
Exhibit 11Three zones, three behavioral layers. Schematic; operational detail is held under agreement.
The case in full
3 min read · 3 exhibits
The situation
The Hajj is the world’s largest recurring human gathering. Saudi Arabia’s statistics authority counted 2,489,406 pilgrims in 2019, 1,833,164 in 2024 and 1,673,230 in 2025, moving through a small number of sites on a fixed calendar.1
Security had to be rigorous and could not be seen to profile. Pilgrims arrive from dozens of countries, in many languages, and more than 420,000 people served them in 2025 alone.1
The method, step by step
Separate the zonesApproach roads and transit, perimeter and holding areas, and the sites themselves, each with its own risks (Exhibit 11).
Watch behavior, not identityDensity, movement against the expected pattern, and access at gates and checkpoints: three behavioral layers, no identity field.
Set thresholds from crowd scienceDensity thresholds follow the published crowd-safety literature.
Speak in the pilgrims’ languagesMessaging adapted for Arabic-speaking, South Asian and Southeast Asian pilgrims.
Hand it overThe framework is built for the ministry to run each year.
How it was measured
The record’s result: a security posture that protected the crowd without profiling the person, and a framework the ministry runs each year. Operational detail is held under agreement.
What it means for a race now
At the scale of a city, behavior is the only fair signal (WP-13).
Decide what you will not collect before you decide what you will.
Build to hand over.
Pilgrims by Hajj season, millions, from the General Authority for Statistics. 2020 and 2021 were restricted to residents.Crowd density thresholds and the detection layers that watch each.What the framework watches, and what it refuses to.
Programs for governments, institutions and diasporas in twelve countries.
Institutional communication, public-behavior programs and diaspora engagement, run with local partners and registered before work begins.
Where the firm works, by region.
TheatersExhibit 1
Exhibit 1The firm by longitude, live. Clocks show local time at the seven offices; the pale band is where the sun is up right now. Filled marks are institutional mandates, rings are diaspora programs, small dots are diplomatic work.
The Gulf
Saudi Arabia, United Arab Emirates
Security and behavioral detection frameworks for the world's largest recurring human gathering, some two million pilgrims a year, balancing rigorous protocols with cultural sensitivity. Narrative and partnership work in the UAE, with messaging adapted for Arabic-speaking, South Asian and Southeast Asian audiences.
South Asia
Pakistan; five diaspora markets
A diaspora engagement program for a ten-million-person global community across the United States, the United Kingdom, Canada, the UAE and Australia during a period of acute geopolitical volatility: stabilizing sentiment, building channels between the community and homeland institutions, and standing up crisis-communications frameworks for rapid response.
Japan and Southeast Asia
Japan, Vietnam, Indonesia, Malaysia, Singapore
Market entry, regulatory navigation and trade facilitation across tariff structures that were designed to be opaque. Institutional relationships in Vietnam and across ASEAN; a manufacturing presence launched from nothing.
The Levant and Iran
Israel and Palestine, Iran, UAE
Cross-border humanitarian and commercial diplomacy across lines that formal channels could not cross, using business partnership as an instrument of regional stability.
West Africa
Ghana; the coast from Senegal to Nigeria
Behavioral scoring built with local institutions for financial inclusion: the same modeling discipline applied to a population the file has never described.
The diaspora markets
United States and Canada; with the United Kingdom, UAE and Australia
Five host countries, three languages, one program: the community segmented by behavior rather than origin, reached through leaders identified market by market, with a crisis desk that answers in hours.
Institutional programsFor governments and institutions
What governments and institutions commission.
Ministries of interior and pilgrimage, diaspora and overseas-community offices, national communications directorates, sovereign institutions and the multilateral bodies that advise them have asked us for the programs below. Each is delivered with a local partner of record, registered where the law requires it, and measured, as at home, against a control rather than a survey. Every line in the evidence column is a mandate we have actually run; institutions are named by country and function under agreement.
Program
What we deliver
First result
Evidence from the record
Public engagement
Population-scale programs that change what a public understands about a policy, a service or an institution: segmented by behavior, delivered in three registers, measured against a holdout.
First read at week six; program at scale by week fourteen
Narrative and partnership programs in the Gulf with messaging adapted for Arabic-speaking, South Asian and Southeast Asian publics.
Security and behavioral detection
Detection frameworks for crowds and critical sites built on density, movement and access patterns rather than identity, layered from the approach roads to the site itself.
Framework in one season; run by the institution thereafter
The world's largest recurring gathering, some two million pilgrims a year, protected without profiling.
Diaspora engagement
A nation's overseas community treated as an electorate: segmented, reached through its own leaders, connected to homeland institutions through channels that outlast a crisis.
Pilot market in ninety days; five markets in a year
A ten-million-person diaspora across the United States, the United Kingdom, Canada, the UAE and Australia through a period of acute volatility.
Narrative and crisis desk
Standing capacity to answer an incident within hours across markets and languages: pre-scored audiences, pre-built creative, validators identified in advance.
Desk live in thirty days; first response within two hours thereafter
Crisis-communications frameworks built for the diaspora program and run through subsequent incidents.
Sentiment measurement
Continuous, behavior-based measurement of how a public or a diaspora is moving, independent of the surveys that tell an institution what it hopes to hear.
Baseline in thirty days; weekly thereafter
Behavioral scoring built with local institutions in West Africa for a population the file had never described.
Terrain and counterparty advisory
Market, regulatory and counterparty analysis for institutions entering environments where the rules are unwritten; screened, registered, described in public.
Brief in fourteen days
Market entry and trade facilitation across Vietnam, ASEAN and Japan; cross-border diplomacy in the Levant.
Exhibit 2Institutional programs. Timelines are typical for a signed mandate with a partner of record in place; evidence is drawn from the firm's mandates, anonymized under agreement.
Diaspora as an electorate
Modeling a diaspora as an electorate.
A diaspora is a nation's most valuable and least managed asset: millions of people who carry the homeland's story into boardrooms, neighborhoods and voting booths in five other countries at once. It is also, increasingly, a decisive bloc in the host countries' own elections, and almost no one models it as such.
We model it as one. The method that finds the Night Nurse in a Midwestern state also finds the second-generation engineer in a suburb of Toronto, the remittance-sending nurse in Dubai, the small-business owner in Bradford, and tells us what each of them is listening for.
The seven officesExhibit 3, live
Seven offices across fourteen hours of the clock.
Seven offices spanning fourteen hours of the clock run as a single desk. A program in the field is watched around the clock; a crisis in one market is answered from the office that is open; and a client in Riyadh, Ottawa or Tokyo has a Brushworks team working its own business day. The desk is also where the platform lives: a client's data is hosted in the region its mandate requires, replicated, and reachable from the office that is on watch.
Specification
The desk
Why it is built this way
Coverage
Twenty-four hours, every day a program is in field. Handoff Tokyo to Dubai to Munich and London to Ottawa, New York and Washington, with a written watch log.
A ruling, a headline or an attack lands on someone's business day, not on ours.
Languages
English, Arabic, Urdu, French, German, Japanese and Vietnamese in-house or through standing partners of record; others by mandate.
A crisis desk that answers in translation is late.
Data residency
Client data hosted in the United States, the European Union or the Gulf as the mandate requires; never moved between regions without written instruction.
Some mandates cannot leave the jurisdiction. The platform is built to stay.
Redundancy
Every program environment replicated across two regions with daily point-in-time recovery; restoration tested each quarter and recorded.
A program in week eight cannot lose a day.
Access
Named users, multi-factor authentication, logged access to every client file, separate seats and systems per practice.
The wall is enforced by the system, not by memo.
Platform reach
Activation across social, search, connected television, programmatic and mail in every market where the platforms operate; local media through partners of record where they do not.
Single-origin delivery has to work in Bradford and Sharjah, not only in Ohio.
Response
Crisis desk first response within two hours in any time zone; weekly field report every Monday, wherever the client sits.
The cadence is the same at home and abroad, so the standard is.
Exhibit 4The desk, specified. Residency, replication and access are written into the engagement letter for institutional mandates; independent security reporting available under agreement.
How we work abroad
The terms every program abroad runs under.
Through local partners. Every program abroad is run with a local partner who owns the relationships and the language. We bring the method; they bring the ground.
Registered. Where a mandate touches the United States, it is registered under the laws that govern foreign representation before any work begins. We do not test the line.
Screened. Counterparties are screened against sanctions and end-use restrictions. We decline mandates where the client, the funds or the purpose cannot be described in public.
Disclosed. The same rule as at home: paid voices are labeled, and there are no undisclosed personas, in any language, on any platform.
Ours to answer for. Work abroad is signed by the firm, not by a subsidiary built to be deniable. If we would not put our name on it in Washington, we do not do it in Riyadh.
Four products that plan, run and measure a program.
The Index is the data engine underneath. Terrain is the map the race is planned on. Program writes a sourced line for each cohort and shifts weight daily toward what works. Response answers the news as it happens. Each is built so that a client, a counsel or an independent reviewer can check it.
Who can still be moved. Where the race is decided. What to say to each cohort, and what it did against the people left alone. And what to say when the news changes.
Indexthe data engine
The data engine: identity, the behavioral graph, a registry of 17,000+ signals and a model catalog held at a gate on your holdout.
The data engine that Terrain, Program and Response draw on.
The Index resolves identity, builds the behavioral graph, curates the signal registry and trains the model catalog, then holds every model at a gate on the client’s own holdout before it reaches a single voter. Terrain, Program and Response all draw from it.
Four layers and a publish gate, with measurement kept separate.
LAYER 1Identity resolution
Records from licensed sources are matched to people and households. Identifiers are keyed hashes before anything is stored; no name travels past this layer.
Keyedhashes, at the edge
Per programmatching, under a written data-use agreement
LAYER 2Behavioral graph
62 billion observed behaviors resolved over 270 million adults and their households. The graph is rebuilt every month; activity arrives every day.
62Bobserved behaviors
270M+adults and households
LAYER 3Signal registry
A signal is an observed behavior, never an opinion and never an inference about who someone is. Signals that stop predicting are pruned every month.
17,000+active signals
10,000topics, twenty-five domains
LAYER 4Model catalog
Five model families trained on recorded outcomes. Nothing ships without beating two baselines on the client’s own holdout; failures are logged.
5model families
34 of 84candidates survive the gate, a representative year
BEFORE ANY VOTERThe gate
A model ships only if it beats the stronger of two baselines on your holdout by a margin set in advance: 0.03 in AUC and ten percent relative in top-decile lift.
+0.03AUC margin, pre-registered
+10%top-decile lift, pre-registered
BESIDE THE STACKMeasurement
Measurement sits beside the stack, not inside it: it can see everything and change nothing, which is what makes its judgment worth having.
62Bobserved behaviors, resolved over 270 million adults and their households
10,000topics in the dictionary, in twenty-five domains
17,000+active signals in the registry, pruned monthly
5model families, trained on recorded outcomes
ILLUSTRATIVE · THE REGISTRY DRAWN BY FAMILY; HOVER A TILE FOR ITS FAMILY. BRIGHT TILES ENTERED A SCORED MODEL THIS MONTH.
The gateBefore any voter
A model ships only if it beats two baselines on your holdout.
MODEL CARD · PERSUASION · PROGRAM 0417
Stronger baseline0.64
Candidate0.69
Margin+0.05 (pre-reg. +0.03)
Top-decile lift+14% (pre-reg. +10%)
Brier score0.19
Ledger entryDR-0417-0021
SHIPPED
The exclusion check, before the gate.
Every candidate model and signal is tested for whether it can predict an attribute the data use policy excludes: health conditions, religion, sexual orientation, union membership, immigration status and the others listed there. Above an AUC of 0.60, it does not go forward.
ILLUSTRATIVE · HOVER A POINT OR A BIN FOR ITS VALUE
DriftAfter the gate
Checked daily in the field, retired after three days below threshold.
Drift is checked every day after launch.
Every card states an in-field discrimination level. A model that falls below it for three consecutive days is retired, the cohorts it scored go back to the stronger baseline, and the retirement is written into the kill log with its reason.
Drift is how a model fails quietly: the electorate moves and the signals that predicted it stop predicting. Watching for it is the discipline that keeps a score honest after launch.
ILLUSTRATIVE · THREE MODELS IN FIELD; THE THRESHOLD IS SET PER CARD
What you receiveEvery program
Four documents ship with every score.
DELIVERABLE 01
Model card
Baselines, the pre-registered margin, discrimination and calibration on your holdout, and the ledger entry that recorded the decision.
DELIVERABLE 02
Registry extract
The signals that entered your program’s models, by family and month, with the ones pruned since.
DELIVERABLE 03
Kill log
Every model tested for your program: the baselines it faced, the result and the reason for each retirement.
DELIVERABLE 04
Exclusion report
The test for excluded attributes on every candidate model and signal, with anything struck before the gate.
The Index, second editionIn the build now
The same behavior, read in depth.
Being built now
62Bobserved behaviors
270M+adults and households
17,000+active signals
ILLUSTRATIVE · SEVEN KINDS OF EVIDENCE MEET ON ONE ID; EACH BEHAVIOR MEASURED EIGHTEEN WAYS AND READ EIGHT WAYS
Today’s Index stays. The second edition adds depth beneath every reading.
The Index already reads 62B observed behaviors across 270M+ adults and holds every model at a gate on the client’s own holdout. The second edition, being built now, keeps all of it.
Beneath each reading it adds seven kinds of evidence on one permanent id, eighteen measures of every behavior and eight reads of every construct, from who people usually are to where they are heading and when.
Every score will carry its range, its basis and a status it has to earn against the program’s own control.
23domains in one tree, the civic domain our own
18measures of every behavior
8reads of every construct
7kinds of evidence on one id
6grains, person to community
5bases on every score
9checks before a model is offered
4statuses, grounded to proven
DESIGN COUNTS FOR THE SECOND EDITION, NOW BEING BUILT · NO RESULT IS CLAIMED FOR THEM
Each line keeps today’s read and adds to it.
Inputs
TODAYBehavior and the electoral record
SECOND EDITIONSeven kinds of evidence on one permanent id
Each behavior
TODAYRead by topic
SECOND EDITIONEighteen measures, regularity and context stability among them
Reads
TODAYTrait or state, one per model
SECOND EDITIONEight, adding trajectory, next, timing, value and interaction
Certainty
TODAYA score and its model card
SECOND EDITIONA score with a range, a basis and an evidence depth
Grading
TODAYThe gate on the client’s holdout
SECOND EDITIONThe gate, forward tests on held data, and a scorecard per program
History
TODAYThis week’s reading
SECOND EDITIONAny reading as it stood on any date
RECORD · READINESS TO ACT · DISTRICT 0036Illustrative
Position76 of 100
Range70 to 81
BasisBlended
Evidence depthB
Two ranks4 of 65 · 3 of 22
Last evidenceTwo days ago
TrendRising, 30 days
TenureSix weeks
ExpiryThree weeks
Top reasonsNews ↑ sign-ups ↑
Model and gradev3 · Checked
Rules of useOpen tier
Every score will travel with its own record.
A score on its own invites more trust than it has earned. In the second edition no score moves without its range, the evidence it rests on, the date of its last evidence and the status its model has earned.
A recipient can be sent a score without its basis; never the other way round. Every score is kept as it stood each week, so a program can read back exactly what was known on the day it acted.
Inside a program’s own instance the same record reads a person in full: open models only, reached through an audience and never by name.
A model earns its status in order.
01GroundedPublished research supports the construct, with the passage on file.
02CheckedNine checks pass on our own data, each one dated.
03Forward-testedIt ranks later behavior on held data it never saw, before any program.
04ProvenA program read it against its sealed holdout.
ConvergentDiscriminantKnown groupsStabilityCoverage and spreadProxy: under 0.60 for any excluded traitSensitivityIncrementalDecay
A health condition, read on a person or a place
Sexual orientation, immigration status or criminal history
Anyone under 18
A score built on emotional state, distress or money trouble
A lookup by name, or an eligibility decision of any kind
More depth calls for a firmer line.
Identity reaches a program by a separate and narrower door: as the composition of districts, in cells of at least twenty-five, or, for a single person-level use, under a written approval that names the trait, the purpose, who sees it and when it is deleted.
Emotional tone is read for groups of twenty-five or more. It is never a score for a person and never a condition of an audience.
The race mapped precinct by precinct, with the cohorts that decide it.
Terrain is the map a program is planned and run on: the race precinct by precinct, the cohorts that decide it placed where they live, the plan that reaches them by week and seat, and the brief that changes the plan every Monday.
A campaign on an invented county, from first brief to result.
The cohortsChoose a cohort
The cohorts that decide the race, placed where they live.
S-01 · COMPOSITE · 7.4%The Steward
Defined by tenure: the same address, employer and insurers for more than a decade, and civic activity that runs through institutions rather than platforms.
3.8M
registered electorate, the state
1.42M
observed on the defining behaviors
612K
scored above threshold, five families
318K
separated from three near-neighbor cohorts
281K
reachable today, hashed, six channels
S-02 · COMPOSITE · 4.9%The Splitter
Where this cohort lives on the map, by precinct share. Its dossier carries what moves it, where it is reachable and what never to cross-run.
S-03 · COMPOSITE · 3.1%The Night Nurse
Where this cohort lives on the map, by precinct share. Its dossier carries what moves it, where it is reachable and what never to cross-run.
S-04 · COMPOSITE · 9.2%The Renter
Where this cohort lives on the map, by precinct share. Its dossier carries what moves it, where it is reachable and what never to cross-run.
S-05 · COMPOSITE · 2.6%The Pharmacist
Where this cohort lives on the map, by precinct share. Its dossier carries what moves it, where it is reachable and what never to cross-run.
S-06 · COMPOSITE · 5.5%The Contractor
Where this cohort lives on the map, by precinct share. Its dossier carries what moves it, where it is reachable and what never to cross-run.
S-07 · COMPOSITE · 41% · HELD OUTThe Certain
Held out of every send. The Certain are the measure of turnout, not its target.
The planHover a week
Flights timed to the absentee and early-vote calendar.
Paid, organic and earned media, planned by week, cohort and seat.
The plan is written to the clock of the race. In most states the persuasion window closes weeks before November, when absentee ballots drop.
Precision paidOrganic advocacyAmbientHeld out
Fourteen daysChoose a day
From the engagement letter to the first flight.
THE IGNITION SEQUENCEDay 1 · Terrain brief
Margin arithmetic, the calendar and the return curve, the opposition file, the money, and the county board with its second column, delivered within seventy-two hours.
1234567891011121314
THE IGNITION SEQUENCEDay 3 · Segment map
The file is matched, cohorts are scored, sized and separated from their near neighbors; each receives its lexicon entry. The client signs the map.
1234567891011121314
THE IGNITION SEQUENCEDay 5 · Holdout
Ten percent of every cohort drawn at random and sealed, reported to the client directly. It has never moved in the firm’s record.
1234567891011121314
THE IGNITION SEQUENCEDay 7 · Masters
One human-written master per narrative track, from the campaign’s own content, approved in writing.
1234567891011121314
THE IGNITION SEQUENCEDay 10 · Compliance review
Disclaimers, coordination protocols, registration where it applies, sanctions screening, platform verification, AI-content disclosure. Signed.
1234567891011121314
THE IGNITION SEQUENCEDay 12 · Seats
Single-origin accounts provisioned, hashed audiences matched, the hold list and the holdout applied at every seat.
1234567891011121314
THE IGNITION SEQUENCEDay 14 · On air
The first flight, the daily loop, and the weekly field report from that day forward.
1234567891011121314
Two fixed points: the holdout on day five and the compliance review on day ten. Nothing goes on air before both, and the sequence is the same for a school-board race and a statewide program.
The clockThe general-election calendar
Most ballots are cast before Election Day.
60%
of 2024 ballots were cast before Election Day (Census Bureau).
In most states the persuasion window closes weeks before November, when absentee ballots drop. Terrain writes the plan to that clock: persuasion weight moves to mobilization on the day the ballots drop, not the day the polls close.
SEE WP-02 AND WP-21
The briefFrom week one
A field report every Monday, and a revised plan with it.
Field report · week 3ILLUSTRATIVE
1 · WHAT RANPrecision paid in four cohorts, organic advocacy in two, ambient in the hospital district. Seventeen variants live.
2 · WHAT IT REACHEDReach and frequency by cohort and register, against the plan for the week.
3 · WHAT MOVED AGAINST THE HOLDOUT, POINTS
S-04 The Renter+3.1
S-01 The Steward+1.8
S-03 The Night Nurse+1.3
S-06 The Contractor-1.6
4 · WHAT CHANGES NEXT WEEKS-06 contact paused and its variant retired; weight moved to S-04; the inoculation step brought forward a week.
The Monday report runs to four pages.
What ran, what it reached, what moved against the holdout, and what changes next week. The brief is the plan’s conscience: it is written from the measurement, not from the delivery reports.
Every lift carries its interval. A cohort moving the wrong way is reported on the first page, not the last.
The limitsWritten into every engagement
What Terrain never shows.
I
No names on the map.
Each mark is one household. People are never shown, searched or exported from Terrain.
II
The Certain are held out.
The cohort that will vote either way receives nothing, and is the measure of everything else.
III
Invented on this page.
The county and every value shown here are illustrative: no real place, person or result.
BRUSHWORKS DESK · 03 · MESSAGING AND TARGETINGProgram
Messages written for each cohort, reweighted daily by what works.
Program turns one approved master into sourced, disclosed variants for each cohort, places them at the seat across three registers, and moves weight every day toward what moved the cohort against its sealed control.
Each cohort hears its own line, drawn from one approved master.
WHAT THIS COHORT HEARSListening for competence under pressure.
Eleven years on a trauma floorHEARS
Wrote the transit billNEVER RECEIVES
Cut the prescription copayHEARS
Answered the tax attackHEARS
Precision paidMOBILE VIDEO, SHIFT NIGHTS · 30S“You watch the news between rounds. Here is the thirty seconds that matters.”Paid for by the campaign
Organic advocacyA NURSE · 45S, ON HER OWN FLOOR“She ran a trauma floor for eleven years. You’ll hear she never ran anything.”Validator, disclosed
AmbientOUT-OF-HOME, HOSPITAL DISTRICT · BOARD“Eleven years on a trauma floor.”Sponsor named
WHAT THIS COHORT HEARSListening for someone who keeps things running.
Eleven years on a trauma floorNEVER RECEIVES
Wrote the transit billHEARS
Cut the prescription copayNEVER RECEIVES
Answered the tax attackHEARS
Precision paidCONNECTED TV, EVENINGS · 15S“Same street, same hardware store, same answer: fix it first.”Paid for by the campaign
Organic advocacyA NEIGHBOR · LETTER TO THE EDITOR“She showed up to the library board for six years before anyone asked her to run.”Author named
AmbientMAIL, FIRST-RING SUBURB · CARD“Fix it first.”Sponsor named
WHAT THIS COHORT HEARSListening for what the month will cost.
Eleven years on a trauma floorNEVER RECEIVES
Wrote the transit billHEARS
Cut the prescription copayHEARS
Answered the tax attackNEVER RECEIVES
Precision paidSOCIAL VIDEO, WEEKDAY MORNINGS · 15S“Rent is due before the paycheck clears. Here is her plan for that.”Paid for by the campaign
Organic advocacyA TENANTS’ ORGANIZER · 30S“I asked her about the lease law. She had read it.”Validator, disclosed
AmbientTRANSIT SHELTER · BOARD“Read the lease law.”Sponsor named
✓Brand: palette, type, voice✓Disclosure on the master✓Compliance review✓A named person reads and signs
LINES ARE ILLUSTRATIVE MASTERS WRITTEN FOR THIS PAGE. EVERY VARIANT TRACES TO A CLAIM; EVERY CLAIM TRACES TO A SOURCE THE CAMPAIGN SUPPLIED.
TargetingThree registers
Placed through seats the firm holds, with the control suppressed at each.
Audiences are matched as hashes inside each platform.
Audiences travel to each platform as keyed hashes and are matched inside it. The sealed control and the hold list are suppressed at the seat, so nothing reaches them. All paid placement runs through the firm’s own seats, so every dollar traces from invoice to impression to the holdout that judged it.
Three registers, all disclosed: precision paid, organic advocacy with a real second voice, and the ambient campaign with its sponsor named.
Learning in real timeHover a day
Weight shifts every day toward the variants that work.
■ variant A■ variant B■ variant C■ variant D
Each variant is read against the control, daily.
Every variant is read against the cohort’s sealed control, with its interval. What moved is reflighted; what backfired is suppressed the day it shows; what was wrong retrains the match.
↑Reflight: weight to the variants that moved the cohort
×Suppress: contact that hurts stops the day it shows
↺Retrain: a wrong match goes back to the Index
ILLUSTRATIVE · HOVER A DAY FOR EACH VARIANT’S SHARE OF WEIGHT
The narrative trackA single cohort’s track
A seven-step narrative track, with the ask at the end.
S-01 · THE STEWARD · LISTENING FORCost-of-living arithmetic told with respect; the fair deal; who wrote the bill
THE ATTACK COMING“The tax plan.” Weight arrives in week three from the opposing committee
VALIDATORSA parish trustee, a shop foreman, a retired county clerk
“You’ve kept the same job, the same house and the same promises for twenty years.”BASIS · TENURE SIGNALS: SAME ADDRESS AND EMPLOYER, RENEWAL CADENCE
MailWk 1
2Concern
“The people who wrote the last budget never had to keep one.”BASIS · FEAR 0.77: THE NEXT BILL, AND WHO WROTE IT
Connected TVWk 2
3Inoculation
“They’ll call it a tax plan. Here is what it actually does to a household like yours.”BASIS · OPPOSITION SPEND FORECAST: ATTACK LANDS WEEK THREE
Mail, connected TVWk 2–3
4Proof
“One vote, on the record, that kept the county’s rate where it was.”BASIS · INTENT TOPICS: PROPERTY TAX, LOCAL GOVERNMENT
RelationalWk 3–4
5Contrast
“The other choice voted the other way, twice. We’ll show you the dates.”BASIS · DECISION STYLE: READS THE MAIL, DECIDES AT THE TABLE
MailWk 5
6Permission
“A parish trustee and a shop foreman on why they’re voting this way.”BASIS · TRUST RUNS THROUGH INSTITUTIONS, NOT PLATFORMS
Organic validatorsWk 6
7The ask
“Your ballot arrives Tuesday. Here is the line, and here is when it has to be back.”BASIS · ABSENTEE DROP: WINDOW CLOSES; MOBILIZATION BEGINS
Mail, SMSWk 7
Inoculation: the attack answered in advance.
A weakened dose of the attack, delivered early, with the rebuttal attached. Placed before the opposition’s weight arrives in week three.
Why the ask comes last.
A request made before recognition, proof and permission is a solicitation. Made after them, it is a conclusion the voter has already reached.
LINES ARE ILLUSTRATIVE MASTERS WRITTEN FOR THIS PAGE; EVERY REAL LINE CARRIES A DISCLOSURE
The readoutAgainst the sealed control
Lift reported daily, with its confidence interval.
CHOOSE A COHORT
+5.0 pts±1.0, 95% interval, week 8
+3.8 pts±1.1, 95% interval, week 8
+2.9 pts±1.3, 95% interval, week 8
+2.3 pts±1.3, 95% interval, week 8
THE CONTROL
90 in every hundred can be reached. 10 are sealed on day five and are the measure of everything else.
ILLUSTRATIVE · HOVER A WEEK FOR ITS VALUE. A RESULT IS REPORTED WITH ITS INTERVAL, OR NOT AT ALL.
The recordAppend-only
Every send recorded in an append-only ledger.
DR-0417-0001Program opened
prev genesis… h 3a91…
DR-0417-0007Segment map signed
prev 3a91… h c04e…
DR-0417-0011Holdout sealed
prev c04e… h 9e1c…
KL-0417-0002Model retired at the gate
prev 9e1c… h 51bd…
DR-0417-0038Week 3 reallocation
prev 51bd… h e7f2…
Every decision, written down and chained.
The record is the program’s memory of what was decided, by whom, on what evidence. Each entry carries the hash of the one before it, so an entry cannot be altered or removed without breaking every entry after it. The chain head is delivered with each weekly report, and a client may read their program’s record at any time.
Decision records (DR) are written when a gate is passed, a plan is signed or an envelope is exceeded; kill-log entries (KL) when a model is retired.
Single origin: from invoice to the holdout.
1Invoice
2Seat
3Impression
4Cohort
5Readout
Every placement runs from the firm’s own accounts of record across social, search, connected television, programmatic and mail, so each one is traceable to an invoice and each result to a cohort and its control.
FROM THE DOCUMENTATION, FIGURE D4
The limitsWritten into every engagement
What Program will not send.
I
No synthetic depictions of real people.
Candidates and validators appear as themselves, with their consent, or not at all.
II
No result without its interval.
Every lift is reported with its interval, against the sealed control, or not at all.
III
Measurement grades, never edits.
It reads every layer and writes to none; the delivery team cannot read it.
IV
No claim without a source.
Every line traces to a claim, and every claim to a source the campaign supplied.
V
No variant without a disclosure.
Every placement names its sponsor; AI-generated content carries a clear label.
VI
No send without a signature.
A named person reads and approves every master before a variant is made from it.
Rapid response, sent only to the cohorts an event affects.
Response reads the wire, the record and each cohort’s attention around the clock. When an event lands, it finds the cohorts it moves, drafts from the campaign’s approved message house, and sends tailored, disclosed answers through every channel the plan uses once a named person signs.
An opponent’s ad on the tax plan starts running on morning television.
06:02DetectedAd tracking and the morning wire
06:14Cohorts readAttention to the topic rising in S-01 and S-04
06:21DraftedFrom the approved answer to the tax attack
06:38CheckedClaims register, disclosure, counsel’s list
06:44SignedThe communications director
06:50LiveSocial, search and CTV, two cohorts
Day 2ReadAgainst each cohort’s sealed control
COHORTS IT MOVES
S-01S-02S-03S-04S-05S-06S-07
S-07, THE CERTAIN: HELD OUT
THE APPROVED ANSWER · ILLUSTRATIVE“You’ll hear she raised your taxes. Here is the bill she voted for, and what it cost a household like yours.”
Paid for by the campaignFrom the message houseSigned by a named person
The transit question comes up in the second half of the debate.
21:31DetectedLive debate desk
21:36Cohorts readSearch and video attention in S-02 and S-06
21:40DraftedFrom the transit answer in the message house
21:52CheckedClaims register, disclosure
21:58SignedThe campaign manager
22:05LiveSearch and social, two cohorts
Day 2ReadAgainst each cohort’s sealed control
COHORTS IT MOVES
S-01S-02S-03S-04S-05S-06S-07
S-07, THE CERTAIN: HELD OUT
THE APPROVED ANSWER · ILLUSTRATIVE“She wrote the transit bill. The route you take to work is in it.”
Paid for by the campaignFrom the message houseSigned by a named person
The county board puts the school bond on the ballot.
14:05DetectedPublic record, board agenda
14:20Cohorts readCivic attention in three cohorts
14:31DraftedFrom the measure’s fact sheet
14:55CheckedClaims register, committee disclaimer
15:02SignedThe committee treasurer
15:10LiveMail queue, social and email
COHORTS IT MOVES
S-01S-02S-03S-04S-05S-06S-07
S-07, THE CERTAIN: HELD OUT
THE APPROVED ANSWER · ILLUSTRATIVE“The bond fixes the roofs on four schools. Here is the list, and what it costs per month.”
Paid for by the campaignFrom the message houseSigned by a named person
ILLUSTRATIVE · TIMES, LINES AND COHORTS ARE EXAMPLES OF THE SEQUENCE, NOT A RECORD OR A PROMISE OF SPEED.
One night on the deskSeven offices
Events logged overnight, and what was done about each.
THE OVERNIGHT LOG · ILLUSTRATIVE
21:58WASDebate answer signed by the campaign manager; live on search and social, two cohorts
23:40TYOTokyo takes the desk; overnight attention read by cohort
01:15DXBCounty board agenda posted; the bond fact sheet queued for review at 08:00 local
04:12LONS-06 contact paused: the transit variant reads negative against its control
04:40MUCModel M2-c retired at the gate; S-02 falls back to the stronger baseline
06:00NYCMorning brief filed: what moved overnight, what changed, what waits for a signature
The desk is staffed around the clock.
The daily loop runs wherever it is morning. Variants are retired and contact is paused inside the envelope the client signed; anything outside it waits for a named person on the campaign.
By six in the morning in Washington, the campaign has a brief: what moved overnight, what the desk changed, and what is waiting for a signature.
How a response is decidedThe decision path
Eight steps from event to send, two of them signed by a person.
01EventThe wire, ad tracking, the public record, a debate desk
02Who it movesAttention to the topic, read cohort by cohort
03An approved answer?The message house and the claims register
04AdaptOne line per cohort from the approved master
05CheckCounsel’s list for the jurisdiction, disclosure, AI label
06SignA named person on the campaign
07SendEvery channel the plan uses
08ReadAgainst each cohort’s sealed control
!No approved answer, no send. When the message house has nothing that fits, the desk drafts for the campaign to decide and holds. Speed never replaces the signature.
The kitPrepared in advance
Answers written in advance, then tailored to each cohort.
01
The message house
The campaign’s approved positions, proof points and the answers to the attacks it expects, written before anything happens.
02
Answered before it lands
Inoculation lines on the record for the attacks the opposition’s spend and record make likely, so the answer is ready when the attack runs.
03
A line per cohort
Each answer is adapted from the approved master to what each cohort is listening for, in the studio Program uses.
04
Every channel the plan uses
Social, search, connected TV, email, the mail queue and field scripts, through the firm’s own seats.
The clockFourteen hours of the clock
Which offices hold the desk right now.
NOW, 09:12 UTC · LIT: AT THE DESK
Washington04:12 local
New York04:12 local
Ottawa04:12 local
London09:12 local
Munich10:12 local
Dubai13:12 local
Tokyo18:12 local
Prepared forFrom the Review
Three uses: between cycles, in a crisis and before an attack.
BETWEEN CYCLES
The permanent campaign
Committees lose their audiences in the years between elections. Pre-built creative and pre-scored lists are released on a ruling, a vote or a headline, not on the calendar.
A crisis desk is a set of decisions made in advance: who is reachable, with what, in which language, from which office. The clock starts when the headline does.
A campaign on an invented county, from the first brief to the measured result.
Nine steps on an invented county: how the Index reads it, how Terrain plans it, how Program and Response run it, and how the result is measured against a sealed control. Every value is invented and every figure agrees with the others.
Reporting on the founder’s work with governments and partners abroad. Each story is written by the firm and dated; the original coverage is linked beneath it.
The Minister for Communication, Digital Technology and Innovations receives Adam Syed in Accra. Photograph: Miles Creative Studio.
Accra, Ghana · Reported August 1, 2025 · Government
Ghana opens talks on sovereign AI infrastructure with a delegation led by Brushworks’ founder
Samuel Nartey George, Minister for Communication, Digital Technology and Innovations, received the delegation and asked that the work proceed with the National Information Technology Agency.
Ghana has set out to lead the continent in artificial intelligence, and its government is choosing the partners it will build with. In meetings reported on August 1, 2025, the Minister for Communication, Digital Technology and Innovations, Samuel Nartey George, received a U.S. delegation led by Abdullah “Adam” Syed, who went on to found Brushworks. The agenda was secure, scalable and energy-efficient data infrastructure, capacity in artificial intelligence, and digital sovereignty.
The delegation included its Ghanaian partner, OMNI Strategies, led by its chief executive, Dr Francis Blay. Ing. Dr Collins Yeboah-Afari, Director-General of the Kofi Annan Centre of Excellence in ICT, joined the meeting. The minister described AI as an area in which Ghana intends to be a continental leader, welcomed the firm as a strategic partner, and asked that the work be developed with the National Information Technology Agency so that it follows national policy and international practice.
No agreement was announced at the meeting.
What it shows
Engagement at ministerial level on national AI and data policy, alongside the agencies that set it.
On Nigerian television, Brushworks’ founder sets out how data can follow stolen public money home
A fifteen-minute interview on Anti-Corruption Crusade with Hon. EJ Agbonayinma, a former member of the House of Representatives and a Federal Commissioner of the Code of Conduct Bureau.
The host put the country’s hardest question directly: how money taken from Nigerians, including the proceeds of undeclared crude oil, can be traced and returned. Abdullah “Adam” Syed answered with the record that every transaction leaves behind.
Company formations, asset purchases and cross-border transfers each leave a trail; for crude oil, so does the satellite record of where each vessel loaded, sailed and docked. Read together, he argued, those records can identify an illicit transfer and support the cooperation with international law enforcement that freezing and repatriating assets requires.
He closed with an appeal to Nigerians to invest in their own capabilities in data and artificial intelligence. The interview is hosted here in full so that it stays available.
What it shows
Public, on-the-record engagement with Nigerian officials on data in accountability work.
Philadelphia and Vietnam · Announced May 1, 2025 · Partnership
A U.S.–Vietnam manufacturing and logistics partnership, built for resilient supply chains
An equity partnership with DVR International, a Vietnam- and U.S.-based manufacturing and logistics platform, announced by the firm Abdullah “Adam” Syed founded before Brushworks.
To meet shifting commercial needs in global trade, the firm Abdullah “Adam” Syed founded before Brushworks announced an international manufacturing, shipping and risk-mitigation division on May 1, 2025, built on an equity partnership with DVR International.
According to the announcement, DVR brings in-house and partner production in Vietnam across textiles, consumer goods, plastics, furniture, packaging, electronics and mechanical assembly, with bonded warehousing and international shipping. The division offers U.S. companies production from prototype to scale, end-to-end shipping with real-time visibility, and support with customs, zoning and permitting in Vietnam and the United States.
What it shows
Operating relationships in Southeast Asia, and programs built to satisfy regulators in two countries at once.
In 1984, about 190 congressional districts voted for one party for president and the other for the House. In 2024, sixteen did. The swing voter did not disappear. She moved to the decisions where the party label does not reach.
By Brushworks research deskNovember 2026Working paper
Abstract. Split-ticket outcomes have nearly vanished from American federal elections: sixteen House districts in 2020 and in 2024 split their presidential and House votes, against about 190 in the landslides of 1972 and 1984, and only three states will send a split Senate delegation to the 119th Congress, the fewest since popular election of senators began. We read the decline from the record, separate what it says about partisanship from what it says about persuasion, and set out where the persuadable voter still decides elections: in the choice to vote, in ballot measures and local races, and in primaries.
In brief
The finding
Party-switching at the top of the ticket has collapsed. Movement has not. It now happens between voting and staying home, and in contests without a party label.
For a campaign
Plan persuasion where the party cue is weak and mobilization where it is strong. In a partisan general election, the swing voter is usually a sometimes-voter.
For an institution or a ballot committee
Measures and local races are where cross-party majorities still form. They are decided by people who agree with the other party on almost everything else.
For most of the twentieth century American voters split their tickets often enough to keep both parties in power at once. In the Nixon and Reagan landslides, about 190 House districts voted for one party for president and the other for Congress.1 In 2020 the number was sixteen, and in 2024 it was sixteen again.2, 3 This paper reads that decline, asks what it means for persuasion, and sets out where the voter who can still be moved now makes up her mind.
The record
The count of split districts fell in steps. At least 145 districts split in 1988.4 By 2000 the number was 86, and after a brief rise to 83 in 2008 it fell to 26 in 2012, 35 in 2016 and 16 in 2020.2, 3 The Senate shows the same pattern: in the incoming 119th Congress, only Maine, Pennsylvania and Wisconsin will have senators from different parties, the fewest split delegations since direct election began in 1914.5
House districts voting for one party for president and the other for the House. Counts from Pew Research Center, Brookings and the Crystal Ball; earlier years rounded as published.
Two things drove most of the fall. The old split was regional: between 1972 and 1988, 41 percent of split districts were in the South, where conservative voters backed Republican presidents and Democratic members of Congress.4 As that realignment completed, the splits went with it. The second is nationalization: voters increasingly treat a House race as a vote on the national parties, and dislike of the other party has become a stronger motive than attachment to one’s own.6, 7
The voters who did move
None of this means that no one changes their mind. In 2016, 206 counties that had voted twice for Barack Obama voted for Donald Trump. They cast 5.5 percent of the national vote and accounted for 51 percent of the Republican net gain from 2012.8, 9 Small, concentrated movement still decides presidential elections.
The pivot counties of 2016: a small share of the vote, a large share of the change.
The same counties show how durable a shift can be once it happens. By 2024, 179 of the 206 had voted for the Republican nominee three times running. The other 27 voted Democratic at least once afterward: 16 in 2020 only, 9 in both 2020 and 2024, and 2 in 2024 only.10
What became of the 206 pivot counties after 2016.
Where persuasion now happens
If voters rarely switch parties within a ballot, they still decide many things. Four decisions remain open to persuasion, and each needs a different program.
Whether to vote. In a nationalized general election, the voter whose choice can change the result is often the one deciding whether to show up. That is a mobilization job, and it is measured in turnout against a control.
Ballot measures. A measure carries no party label. Voters who agree with one party on almost everything else regularly cross on a specific question, and the published experiments find persuasion effects there that they do not find in partisan general elections.11
Local and nonpartisan races. School boards, city councils and county offices are often decided without a party cue, by small electorates that skew older (WP-35).
Primaries. Within a party, the label is the same on every candidate, and the voter has to decide on something else.
Where the party cue is weak, persuasion has room to work. A qualitative reading of the record and the experimental literature.
How the split ticket ended
The split ticket of the 1970s and 1980s was, more than anything, a Southern institution. In 1972, two thirds of all Southern House districts voted for one party’s presidential nominee and the other’s candidate for Congress. As recently as 1988 more than half still did. Between 1972 and 1988, 41 percent of all split districts in the country were in the South, and more than nine in ten split districts nationally voted Republican for president and Democratic for the House.4
The pattern had a plain cause. Conservative Southern voters had moved to Republican presidential candidates a generation before they moved to Republican members of Congress, many of whom did not yet exist. Long-serving Democratic incumbents held their seats on personal standing, seniority and local service. As those members retired and Republican candidates contested the seats, the split closed. By 2012, voters in only eight Southern districts split their presidential and House votes, five percent of the country’s split districts that year.4
The South and the split ticket. Bar lengths approximate the shares as Pew Research Center describes them.
Outside the South, the decline followed a slower and broader change. Voters came to see House elections as choices between national parties, local news that once covered members of Congress closely thinned out, and the share of voters who dislike the other party more than they like their own rose.6, 7 A member’s local record, which used to win the votes of people who disagreed with her party, now buys less.
The lesson for a campaign is not that local standing is worthless. It is that local standing has to be carried to the voter by the campaign itself, because the institutions that once carried it, from the local paper to the party organization, carry much less of it now.
How to find a persuadable voter in a nationalized electorate
If the persuadable voter rarely switches parties, she is found by what else she does. Four kinds of behavior separate her from the settled partisan, and all of them can be read without asking about identity.
Her vote history across contest types. A voter who votes in presidential years and skips midterms, primaries and local elections decides each time whether to vote at all.
Her ballot behavior below the top line. Roll-off (leaving lower races blank) and crossing on measures show a voter who decides question by question.
Her attention. A voter who follows the race closely has usually decided. One who reads local news and little national politics often has not.
Her recent changes. A move, a new job, a new child or a new bill changes what a voter is listening for, and often whether she votes.
Behaviors that separate the persuadable from the settled, and what each suggests. A qualitative reading.
A worked example: one state’s persuadable pool
The figures in this example are illustrative, for an invented competitive state. Of every hundred registered voters, a campaign might find that about four are plausible switchers at the top of the ticket, about fourteen are sometime voters who lean one way and decide whether to vote, and about ten are cross-pressured on a measure on the same ballot. The remaining seventy-two have settled both their party and their habit of voting.
One invented state’s registered voters by the decision still open to them. Illustrative.
A plan written for the four would spend its persuasion budget fighting for a small and contested group. A plan written for all three groups spends most of its money on the fourteen and the ten, and measures each program against its own control.
Objections
Turnout is not persuasion. For a campaign, it is the same decision in a different form: whether this voter’s choice counts this time. The voter who leans your way and stays home is the most valuable persuasion target in a nationalized election.
Many voters call themselves independent. They do, Most of them lean toward one party and vote like its partisans.12 A self-description is a weak guide to a vote; a record of behavior is a better one.
Ticket-splitting will come back. It may. If it does, the voters who return to it will show it first in their ballot behavior below the top line, and a program that reads behavior will see them before a program that reads party registration.
Questions to ask of a targeting plan
How many voters does the plan count as persuadable, and by what behavior?
How many of them are switchers, and how many are sometime voters deciding whether to vote?
Is there a measure on the same ballot, and who crosses on it?
What control will say whether the persuasion and turnout programs each worked?
Further reading
Hopkins, D. J. (2018), cited above: how American political behavior nationalized.
Abramowitz, A. I., and Webster, S. (2016), cited above: negative partisanship and the decline of split tickets.
Burden, B. C., and Kimball, D. C. (2002). Why Americans Split Their Tickets: Campaigns, Competition, and Divided Government. University of Michigan Press: the account of split-ticket voting at its height.
Implications
For a campaign in a partisan general election. Count the persuadable voters with care, and expect them to be fewer than the plan assumes. Spend the persuasion budget early and on the people still undecided (WP-32), and treat most of the close as a turnout program.
For a ballot committee, a company or a public body. Measures and local questions are where cross-party majorities still form. They are won by finding the voters who agree with you on this question and not on much else, and by saying one sentence each of them can check (WP-03).
For anyone reading a map. A county that moved once can move again, and most of the ones that moved in 2016 did not move back. Read behavior across several cycles before assuming that last time’s margin is this time’s ceiling.
Notes and references
Brownstein, R., in The Atlantic, as quoted by Ballotpedia; and Axios, July 2022.
Crystal Ball (University of Virginia Center for Politics) counts for 2000 and 2024, as reported in Jacobson, L., “Presidential downballot losses: an updated history.”
Hopkins, D. J. (2018). The Increasingly United States: How and Why American Political Behavior Nationalized. University of Chicago Press.
Abramowitz, A. I., and Webster, S. (2016). The rise of negative partisanship and the nationalization of U.S. elections in the 21st century. Electoral Studies, 41, 12–22.
Kalla, J. L., and Broockman, D. E. (2018). The minimal persuasive effects of campaign contact in general elections: evidence from 49 field experiments. American Political Science Review, 112(1), 148–166.
Petrocik, J. R. (2009). Measuring party support: leaners are not independents. Electoral Studies, 28(4), 562–572.
Basis of figures. Split-district counts as published by Pew Research Center, Brookings and the Crystal Ball; 1972 and 1984 rounded to about 190 as published. Pivot-county figures from Ballotpedia. The exhibits on where persuasion happens and on behavioral signals are qualitative readings; the state pool is illustrative. No Brushworks program figures appear in this paper.
A political practice that models voters as individuals.
Strategy, research, modeling, creative and delivery work as a single team, and every program is measured against a sealed control.
Seven offices, one deskLocal time
0006121824 UTC
Brushworks is a political intelligence firm. Strategy, behavioral research, modeling, creative and delivery sit in one team under one measurement discipline, so that what the research finds is what the field executes and what the report shows.
The practice grew out of a decade of campaign work, more than five hundred contests across every level of American politics, and a parallel body of institutional work abroad. The two inform each other. The hardest theaters teach discipline about evidence and restraint; American campaigns teach speed.
We are a small firm by design. We take a limited number of programs each cycle, staff each with a full team, and decline the rest.
The clients we keep are the ones who want to be told what happened rather than what they hoped.
Three practices and two studios, walled where the law requires.
StudioCreative studio
Master briefs, narrative tracks and variant production across three registers, every piece disclosed.
Register IRegister IIRegister III
ICandidate practice
Federal, statewide and local campaigns, from terrain through proof.
The wall
IIIndependent-expenditure practice
Donor-funded committees and issue programs, separated from the candidate practice by structure, systems and seats, not by memo.
IIIInternational practice
Institutional and diaspora mandates for governments and institutions, run with local partners across seven offices.
StudioResearch and modeling
The human layer: signals, models, the publish gate, the holdout. Serves all three practices without crossing the wall.
WatchThe desk
Seven offices across fourteen hours of the clock, watching every program in the field.
OfficesLocal time, live
Exhibit 1The seven offices. Drag to turn.
Where the seven offices are, and what each covers.
Washington and New York carry the candidate and independent-expenditure practices; Ottawa runs diaspora and North American institutional work; London and Munich serve Europe and the desk; Dubai serves the Gulf and South Asia; Tokyo serves Japan and Southeast Asia. Every office is on the same daily loop, so a program in the field is watched by whoever is awake.
The rules we work byFive things we will not do
Five things we will not do.
Written into every engagement letter, and the first page a donor's counsel or a ministry's compliance office is shown.
I
Target on identity
We model behavior only. Ethnicity, religion and ideology are never modeled, never bought, and never accepted as a filter.
So the Steward can be any color and any faith.
II
Hide the messenger
Paid advocacy is labeled. We do not operate undisclosed personas or accounts, on any platform, in any language, for any client.
So the second voice is a real one.
III
Report without a control
Every program carries a randomized holdout sealed on day five, and every model beats a demographic and a look-alike baseline before it ships.
So the number in the report is the number that happened.
IV
Cross a wall
Independent-expenditure work is separated from candidate work by structure: teams, systems and seats. Foreign work is registered before it begins.
So a donor's counsel can certify the posture, not take our word for it.
V
Run what we could not describe
If a program could not be explained on the record, in public, in the client's name and ours, we do not run it. Mandates that fail the test are declined in writing.
The governance of Brushworks is built so that a client, a donor's counsel, a committee treasurer or a ministry's compliance office can verify what we say rather than take our word for it: separate practices behind a documented wall, a compliance review that no program goes on air without, written sign-off gates with a reallocation threshold, a conflict register, and a channel for concerns that answers in writing. Each control has an owner inside the firm and a party outside it who can check it.
Structure
A Wyoming corporation whose whole business is political and institutional work, organized as a candidate practice, an independent-expenditure practice and an international practice.
Verified by: engagement letters; counsel of record
Practice walls
Candidate and independent-expenditure work run on separate teams, systems and media seats. No person crosses during a cycle; no data, creative or plan does either.
Verified by: certification available to any committee or counsel
Compliance review
Every program passes a signed review on day ten and before any material change: disclaimers, coordination, registration, sanctions, platform policy.
Verified by: counsel of record; the signed review
Client sign-off gates
Strategy, creative masters and the media plan are approved in writing before execution. Reallocations above ten percent of any line need the same approval.
Verified by: the client, or the contributors' advisory board
Conflicts
Never both sides of the same contest or question. A conflict register is reviewed before every engagement; mandates that cannot be described in public are declined in writing.
Verified by: the register, on request
Data and models
The data use policy binds every employee and vendor; a data protection lead owns retention and rights requests; the publish gate and the kill log are reported to clients.
Founder and chief executive of Brushworks. He leads the firm's practices and its teams across seven offices, and has run behavioral programs in more than five hundred contests, from presidential campaigns and statewide ballot measures to bond elections and institutional mandates on three continents.
500+contests run, from presidential to school board
7offices led, across fourteen hours of the clock
12countries of institutional and diaspora mandates
Adam Syed founded Brushworks to plan campaigns from the behavior of individual voters and to prove each program's effect against a sealed control.
Leading Brushworks
As chief executive he directs the firm's three practices (candidate, independent-expenditure and international) and the research, studio and desk functions that serve them, with teams in Washington, New York, Ottawa, London, Munich, Dubai and Tokyo. He sets the firm's method and its rules: behavior rather than identity, a sealed control in every program, single-origin delivery, disclosed messengers, and the kill log. The governance on this site (the practice walls, the compliance review, the concerns channel) is his design, and he is accountable to the board for it.
He is also the founder and chief executive of Moonbrush, the behavioral intelligence company whose platform and graph underlie the firm's work, and he oversees both organizations' teams globally. Brushworks is Moonbrush's political and institutional practice, operated as a separate firm under its own name, rules and disclosures.
Campaigns
His campaign record spans more than five hundred contests across both parties and every level of American politics, and it has run at national scale. He has designed and run behavioral programs for a presidential general-election campaign, for a congressional insurgency that unseated a long-serving incumbent, for gubernatorial campaigns in large and competitive states, for United States Senate races, for statewide ballot measures and for more than a hundred bond and levy elections. The programs have won ninety-six percent of the time. The losses are published in the casebook, because a record that hides them is not a record.
PresidentialA national general-election campaign in 2016, in which the campaign's narrative was rewritten county by county for the places that decided the race: four counties in one state received four different campaigns, each built on the county's own fact and measured against its own holdout.
CongressionalA primary insurgency in New York in which a first-time candidate unseated a long-serving incumbent. Turnout was modeled from behavior among voters the file had never seen, and a relational register made a small budget behave like a large one.
GubernatorialA gubernatorial campaign in a large Midwestern state run as three distinct regional strategies under one standard of proof, part of a portfolio of gubernatorial programs across three competitive states with three different theories of the race.
United States SenateSenate campaigns in competitive states built on the neglect algorithm (the persuadable, likely voters both parties had stopped contacting) the county board's second column, and register-level holdouts that told the campaign each week what was working.
Ballot measuresThe 2022 California online sports-betting proposition: sixty-five percent support in the campaign's baseline polling, defeated with 17.7 percent of the vote, with forty-one variants and every frame verifiable against the text.
Bonds and leviesMore than a hundred measures asking taxpayers to raise their own taxes, more than five billion dollars approved, passed on concreteness: the named roof, the named bus route, the monthly cost stated beside what it buys, and a calendar chosen deliberately.
Committees and organizationsIndependent-expenditure and issue programs behind a documented wall, and permanent campaigns for committees and organizations whose audiences move on events rather than dates.
What unites the record is the method rather than the party: behavior rather than identity, a sealed control in every program, single-origin delivery, disclosed messengers, and a report at close that says what the program did rather than what the race did.
Governments and institutions
Abroad he has led mandates for governments and institutions in twelve countries: a behavioral security framework for the world's largest recurring human gathering; a five-market engagement program for a ten-million-person diaspora through a period of acute volatility; cross-border humanitarian and commercial diplomacy in the Levant; market-entry and trade work in Japan and Southeast Asia; and behavioral scoring for financial inclusion built with institutions in West Africa. He established the firm's seven-office desk so that a program anywhere is always on someone's business day, and the crisis architecture the desk runs.
Before Brushworks
Before founding Brushworks and Moonbrush, he founded A&S Academic Services, which was acquired by the Texas Education Agency in 2021. He is a dual citizen of the United States and Canada and works from the firm's Washington and Ottawa offices.
Writing
He writes for the Brushworks Review, where his papers set out the firm's method and its doctrine.
Everything needed to describe the firm accurately: approved descriptions, facts that can be checked, the founder's biography and portrait, the wordmark and the rules for using it. Press inquiries are answered within one business day.
Press contact
Presspress@brwrks.com: answered within one business day, in any of the firm's offices' hours.
Research and data requestsresearch@brwrks.com: questions about a paper in the Brushworks Review, its sources or its exhibits.
VerificationRegistrations, disclaimers and policies are confirmed on request; client identities held under agreement are not disclosed. The verification ledger says where each claim can be checked.
Speaking and briefingsThe founder speaks on behavioral persuasion, measurement and the governance of political data. Requests through the press address.
Boilerplate
Short.
Brushworks is a political intelligence firm. It models how individual voters behave, determines what would move each of them, and puts persuasion into the field with a standard of proof: a sealed control group in every program. The firm works for campaigns, committees, governments and institutions from seven offices in six countries.
Long.
Brushworks is a political intelligence firm. Its programs model how individual voters behave, from sixty-two billion observed behaviors resolved to more than 270 million American adults and their households, determine what would move each of them, and deliver persuasion through three disclosed registers from the firm's own media seats. Every program is measured against a randomized control drawn before any money is spent, and every model must beat the voter file and the platforms' own targeting before it is used. The firm does not model ethnicity, religion or ideology, does not operate undisclosed accounts, and publishes the models it rejected. Its record spans more than five hundred contests across both parties at every level of American politics, more than a hundred bond and levy measures, and institutional and diaspora mandates in twelve countries. Brushworks is a Wyoming corporation with offices in Washington, New York, Ottawa, London, Munich, Dubai and Tokyo.
Facts that can be checked
Fact
As stated
Basis
Record
More than 500 contests; programs won 96 percent of the time; losses published in the casebook
The firm's record; casebook
Bonds and levies
More than 100 measures; more than $5 billion approved by voters
The firm's record; public results
California, 2022
Online sports-betting proposition defeated with 17.7 percent yes, from 65 percent support in the campaign's baseline polling
Secretary of State statement of vote; campaign polling
The graph
62 billion observed behaviors; 270 million+ adults and households; 10,000 topics; 17,000+ active signals; rebuilt monthly
The firm's platform documentation
Abroad
Institutional and diaspora mandates in 12 countries; a five-market program for a ten-million-person diaspora; a security framework for a gathering of some two million
The firm's record; institutions held under agreement
Offices
Washington, New York, Ottawa, London, Munich, Dubai, Tokyo
Public
Founded
Wyoming corporation; founder and chief executive Adam Syed
Public record
Founder
Adam Syed is the founder and chief executive of Brushworks. He has designed and run behavioral programs in more than five hundred contests (a presidential general-election campaign, a congressional insurgency, gubernatorial and Senate campaigns, statewide ballot measures and more than a hundred bond elections) and institutional and diaspora mandates for governments and institutions in twelve countries. He leads the firm's practices and its teams across seven offices, writes for the Brushworks Review, and works from Washington and Ottawa. Before Brushworks he founded A&S Academic Services, acquired by the Texas Education Agency in 2021. Full biography
Wordmark and portrait
The Brushworks wordmark is the word set in Archivo Bold with the ochre brush sweep beneath it. It is used on white or on the firm's dark plate; it is never recolored, rotated, outlined or set in another face, and the sweep is never separated from the word. Minimum clear space is the height of the capital B on all sides. The founder's portrait may be used for editorial coverage of the firm with the credit "Courtesy of Brushworks."
We confirm registrations, disclaimers and policies; we describe methods; we do not disclose client identities held under agreement, and we do not comment on contests in which we are engaged while they are live. Corrections to anything published on this site are made with a dated note. Interviews with the founder are on the record unless agreed otherwise in writing before they begin.
Candidate and independent-expenditure work are taken by separate teams; say which you are. Governments and institutions: tell us the mandate and the jurisdiction. Within a week we will tell you whether we can move it and what it would take.
The first fourteen days of an engagement · illustrative
Campaigns, governments, candidates and the people who fund them.
Brushworks is a political intelligence firm: behavioral research, strategy, creative and delivery for campaigns, committees, governments and institutions, measured against a sealed control. This page says what each kind of client receives, what it costs them to start, and what to ask us before they do.
A statewide, congressional, legislative, municipal or ballot-measure program with a standard of proof: run inside your campaign's plan, not around it.
You receive
When
What it is
Terrain brief
Day 1
The margin arithmetic, the calendar and your state's own return curve, the opposition file, and the county board with its second column: the modeled pool of persuadable, likely voters neither side is reaching
Segment map
Day 3
Deciders scored, sized and named as composites (the Steward, the Night Nurse, the Renter) with the hold list of people the program should never contact
The holdout
Day 5
Ten percent of every cohort sealed and reported to you directly, so every claim of lift is checked against people we left alone
Masters and variants
Day 7 onward
One approved master per narrative track, built from your existing content and brand; forty to sixty disclosed variants for a statewide race
Weekly field report
From day 14
Lift by cohort and register with intervals; what was reflighted, suppressed and retrained; the media trail from invoice to impression
Proof report and kill log
Close
What the program did, not what the race did, including the cohorts that moved the wrong way and the models we retired
Where it fits: alongside your pollster (we measure acts, they measure opinion), your media buyer (we place through our own seats and hand you the trail) and your field program (we tell you whom not to knock). Start with the method in five movements, the first fourteen days, and the neglect algorithm. Then brief the firm.
For governments and institutions
Institutional communication, public-behavior programs and diaspora engagement outside an election: registered before work begins, labeled on every placement, measured against a control.
Mandate
What the firm does
Example in the record
Public communication at scale
Cohorts by behavior rather than by census category; messages in the registers each public trusts; a sealed control so the ministry knows what moved
Narrative and regulatory programs across the Gulf and Southeast Asia
Diaspora engagement
A ten-million-person community across five host markets segmented by behavior, reached through the leaders it already trusts, with a crisis desk that answers in hours
The five-market program in the casebook
Security and safety behavior
Density, flow and access frameworks against a site's own baseline, identity absent by design
A gathering of some two million people a year
Crisis architecture
A rehearsed clock, pre-approved masters, register-level response within two hours, measured
The seven-office desk, fourteen hours of the clock
How it is governed: an acceptance screen for client, funds, purpose, jurisdiction and sanctions; registration under the laws that govern foreign representation before any activity touches a market; local partners of record; in-country data holding where the mandate requires it; certified destruction at close. Read Persuasion in the open, the work abroad, and the compliance packs for the jurisdictions you touch. Mandates that cannot be described in public are declined in writing.
For candidates
A standard you can hold your team to: a program that says it moved voters should be able to prove it against people it left alone.
What you will be asked to decide
Why it is yours
The masters
Every variant descends from a master you signed. Nothing goes out in your voice that you did not read
The rules
Behavior only, so no cohort is a demographic label; disclosed messengers, so no one speaks for you who cannot say who asked; a sealed control, so the number in the report is the number that happened
The hold list
The people whose behavior says they decided long ago and who move the wrong way when contacted. The most valuable decision a program makes is whom to leave alone
The record
Our losses are published in the casebook with their post-mortems. Yours would be too
What it takes from you: the file on day one, an approver for masters, your existing content and brand guidelines, counsel's contact. What it gives back: a weekly report that says what moved, a program that can be described in public, and a standard your opponent's vendor cannot meet. Read A standard of proof and Backfire, and the five things we will not do.
For donors, committees and their counsel
Independent-expenditure and issue programs behind a documented wall, with the diligence pack a treasurer or counsel needs to sign off.
What counsel receives
What it shows
Wall certification
The candidate and independent-expenditure practices are separate teams, systems and seats; no person crosses during a cycle; certified by counsel
Compliance review, day ten
Disclaimers, coordination protocols, registrations, sanctions screening and platform verification, signed before anything goes on air
Media trail
Every placement from one account of record: invoice to placement to impression to holdout comparison
Proof report
Lift against a sealed control with intervals: the return on the contribution, stated so that it can be checked
Performance assurance
Underwritten by a third-party carrier for qualifying political programs placed through the firm's seats; described only in the engagement letter, never as a guarantee of an outcome
What a donor should ask any vendor: show me the control; show me the models you killed; show me the wall. The answers are on this site: the reports, the disclosures and the diligence pack. Then write to the firm and say who your counsel is.
Questions we are asked
What is a political intelligence firm?
A firm that models how individual voters or publics behave, determines what would move each of them, and puts persuasion into the field with a measurement design that can fail: a sealed control group in every program. Brushworks combines behavioral research, strategy, a creative studio and single-origin media delivery under one standard of proof.
How does Brushworks measure whether a campaign program worked?
Ten percent of every cohort is drawn at random before any spend, sealed, and never contacted. Lift is the difference between the treated and the control on recorded acts or panel measures, reported with intervals, by cohort and register, weekly and at close. Raw outcomes among the treated are never presented as effects.
Does Brushworks target voters by race, religion or ideology?
Programs are targeted on behavior, recorded acts with a time. Ethnicity, religion and ideology are not in the behavioral registry. Identity enters a program only as the composition of districts or, for a single person-level use, under a written approval for that one program, with counsel’s review; ethnicity and religion never leave in an export, and no trait is ever used to discourage anyone’s vote. Every behavioral model is tested for reconstructing an excluded attribute by proxy and retired if it does.
What does a campaign need to provide to start?
The voter file or supporter file under a data-use agreement on day one, a named approver for creative masters, the campaign's existing content and brand guidelines, and counsel's contact. The terrain brief is delivered within seventy-two hours; the program is on air on day fourteen.
Can Brushworks work for a government or institution outside the United States?
Yes, through its international practice: registered before work begins under the laws that govern foreign representation and lobbying where they apply, with a local partner of record in each market, labeling in the market's language, reporting on the regime's cadence, and certified destruction at close. Mandates aimed at influencing a United States election are declined.
Does the firm work for both parties?
Yes. Clients of both parties are accepted, never on opposing sides of the same contest, and the same rules apply to every program: behavior only, disclosed messengers, a sealed control, a documented wall, and nothing the firm could not describe in public.
What does an engagement cost?
Programs are priced by scope in the engagement letter after a terrain brief. Within a week of a brief, the firm says whether it can move the race or the mandate and what it would take. Media is placed through the firm's own seats at cost with a disclosed fee, and the invoice-to-impression trail is delivered with every report.
How can a donor or a committee's counsel verify the firm's claims?
Through the verification ledger on the policies page, which says where each claim can be checked; the reports page, which publishes the methodology, the kill-log summary and the transparency report; and the diligence pack, available under non-disclosure agreement, which includes the wall certification and a redacted proof report.
Technical documentation for the Brushworks platform.
Technical documentation for the Brushworks platform: how data enters, how a person is resolved, how a score earns the right to ship, how a variant is made, how a dollar is traced, and how a result is proved. Enough to evaluate it; not enough to copy it.
The Index as a flow, from behavior to scored cohorts · illustrative
7layers in the stack, all operated by the firm
62 bnobserved behaviors in the graph
270 m+adults and their households, rebuilt monthly
10,000topics in the dictionary, 25 domains
17,000+signals live in the registry, five families
2baselines every model must beat on your holdout
10%of every cohort sealed on day five, never contacted
14 hof the clock covered by seven offices
0
Read me first
Stable
This is the public technical reference for the platform every Brushworks program runs on. It is written for the people who have to evaluate the firm before they can trust it: a campaign's data director, a ministry's chief information officer, a committee's counsel, an independent assessor, a journalist who reads schemas. It describes architecture, contracts, guarantees and the checks a client can run for themselves. It does not describe model architectures, feature definitions, the registry's contents or the playbook library, which are the firm's property and are not published.
0.1 How to read it
Sections are numbered and each carries a status. Stable means the behavior is a commitment: it is written into engagement letters, the methodology notes and the data use policy, and a change is a versioned change with notice. Evolving means the design is settled but the parameters or interfaces may move between editions of this document. Under agreement marks material that exists in full but is delivered only to clients and prospective clients under non-disclosure.
Three conventions run throughout. Identifiers are shown in monospace and follow the formats in Section 15. Schemas are illustrative: field names and shapes are real, the values are invented, and optional fields are marked. Where a number is a default, it is called a default, and the place it is fixed for a given program (the terrain brief, the model card, the engagement letter) is named.
0.2 The standard these documents are held to
Everything here is consistent with four documents that take precedence over it where they differ: the methodology notes (how a result is measured), the data use policy (what may be done with data), the security summary (how data is protected) and the governance policy (who may decide what). This reference explains the machinery that carries those commitments out.
What is deliberately absent. Model architectures and hyperparameters; the feature definitions behind any family; the topic dictionary and the signal registry as artifacts; the narrative playbook library; the identity of any licensor; any client's program data. A reader who finds one of these here should tell us at privacy@brwrks.com; it is a defect.
1
Architecture
Stable
The platform is a seven-layer stack with a single loop running through it, and a measurement system that sits outside the stack and judges it. Every layer is designed, built and operated by the firm. No layer is rented from a vendor who also sells it to the other side, and no score, audience or result ever passes through a third party's model.
1.1 The stack
DeliveryAccounts of record across social, search, connected television, programmatic and mail. Single-origin: every placement traceable to an invoice.every flight
StudioCohort-specific variants from one approved master and the sponsor's own assets, each stamped with its disclosure and a provenance manifest.minutes per master
Narrative tracksAn authored, versioned library of message sequences mapped to the seven-step arc and the dispositions that trigger each.authored, versioned
Model catalogFive families trained on recorded outcomes. Nothing ships without beating two baselines on the client's own holdout; failures are logged.gated per program
Signal registrySeventeen thousand active signals against a ten-thousand-topic dictionary in twenty-five domains, curated toward what has predicted political decisions.pruned monthly
Behavioral graphSixty-two billion resolved events over 270 million adults and their households; household-attributed where the source is.monthly rebuild, daily activity
Identity resolutionIndividuals and households matched on keyed hashes. Client files matched under written data-use agreements, per program, never commingled.hashed at the edge
MeasurementThe sealed holdout, the register arms, the estimand and the daily readout. It has read access to every layer and write access to none. The delivery team cannot read it.
Figure D1. The stack, bottom to top. Measurement sits beside it, not inside it: it can see everything and change nothing, which is what makes its judgment worth having.
1.2 Planes
The platform is divided into a data plane and a control plane, and the division is enforced by identity, not by convention. The data plane holds events, the graph, the registry, models, cohorts, variants and placement logs. The control plane holds the things that decide what the data plane may do: program definitions, access lists, the decision ledger, model cards, kill logs, sign-off gates and the schedule. A person can hold a role in the control plane for a program without holding any read access to its data, and this is the normal condition for a client's counsel, a program lead and the desk's incident lead.
1.3 Tenancy
Isolation has three levels. Each practice (candidate; independent expenditure; international and institutional) runs in its own environment with its own identities, storage, key hierarchy and delivery seats, and no environment can read another's data. Within a practice, each program is a project with its own access list, its own keyed hash for the client file, its own holdout and its own model catalog. Within a program, the measurement project is the only place where holdout assignment and outcomes meet, and it is closed to the delivery team and to the studio. Two clients never share a model, a segment or a seat; two practices never share an environment.
1.4 Design principles
Outcomes, not proxies. Models are trained on recorded acts (a validated vote, a returned ballot, a contribution, a sign-up), never on survey answers or on other models' scores.
Seal before spend. The holdout is drawn and hashed on day five, before any placement, and the hash leaves the firm's hands the same day.
Kill by default. A model that cannot beat the voter file used well and the platforms' own look-alike targeting on the client's holdout does not ship, and the reason is recorded where the client can read it.
Hash at the edge. Raw identifiers are hashed with a program-scoped key on receipt; audiences leave as hashes; raw personal data never travels.
One origin. Every placement is bought from an account the firm holds, so the trail from invoice to impression is unbroken.
Everything versioned, everything reviewed. Dictionaries, registries, catalogs, playbooks, variants and code carry versions; every production change has a second reviewer and a rollback.
Separation of judgment. The team that delivers cannot see the holdout; the team that measures cannot place; the studio sees cohorts, never records.
2
Ingestion and data contracts
Stable
Data enters the platform from five classes of source, and nothing enters without a contract: a licence or agreement that says where it came from, what it may be used for, how long it may be kept and how often it is refreshed. The contract is a record in the platform, not a document in a drawer, and every derived table carries a pointer back to it.
2.1 Source classes
Class
What it is
Cadence
Retention class
Notes
Licensed behavioral
Observed acts (purchases, subscriptions, renewals, visits, media consumption) from established providers under written agreements
Monthly full; daily digital
L (provider terms)
Nothing scraped; no source that cannot state its own provenance
Public and registry
Voter files, validated-turnout histories, precinct returns, census geography, administrative boundaries
As published
P (public)
Used as the demographic baseline every model must beat
Client first-party
The client's own file: supporters, donors, contact history, panel data
On receipt; on client refresh
C (program life)
Keyed hash on receipt; matched under data-use agreement; destroyed within ninety days of close
Program telemetry
Placement logs, delivery and frequency records, spend, creative exposures from the firm's own seats
Hourly to daily
C (program life)
The media trail; reconciled nightly to the accounts of record
Field and panel
Relational-organizing contact records, returned instruments, institutional administrative outcomes
As collected
C or M (measurement)
Enters the measurement project only; never the delivery path
2.2 The event envelope
Every observed behavior is normalized into one envelope before it can touch the graph. The envelope carries the subject reference (already hashed), the topic and family assigned by the dictionary, the observation and ingestion times, the source and licence, and a retention class that the lifecycle system enforces without human intervention.
{
"event_id": "EV-01J9Q7W4T8N2K3M5R6",
"subject_ref": "h:3f9a…e21c", // keyed hash, environment-scoped
"household_ref": "hh:7c04…91be", // optional; present when source is household-level
"observed_at": "2026-09-14T03:12:41Z",
"ingested_at": "2026-09-15T00:41:07Z",
"source_id": "SRC-0117",
"license_id": "LIC-0117-2026",
"topic_id": "T-COM.4412", // dictionary topic; 25 domains × ~400 topics
"family": "behavioral", // one of the five families
"kind": "renewal",
"value": null, // optional magnitude, source-defined units
"confidence": 0.97, // resolution confidence, band A
"retention": "L",
"lineage": ["SRC-0117/2026-09/part-0412"]
}
2.3 Contract tests
Each delivery from a source is a drop, and each drop passes seven tests before it is admitted: schema conformance against the registered schema version; volume within the source's historical band (a drop more than two standard deviations from its trailing twelve is quarantined); null rate by field; duplicate rate; referential integrity to the dictionary; timestamp sanity (no observation after ingestion, no observation older than the licence permits); and licence validity on the date of ingestion. A drop that fails any test is quarantined, the source is flagged on the desk's board, and nothing from it reaches the graph until a named engineer has released or rejected it. Releases and rejections are ledger entries.
2.4 Lineage
Lineage is a first-class column. Every row in every derived table (graph, registry snapshot, model training set, cohort, audience, readout) carries the identifiers of the drops it was computed from, so that a licence expiring, a source being withdrawn or a data-subject request arriving can be traced forward to everything that depends on it and acted on. The lifecycle system uses lineage to expire, and the privacy system uses it to delete; neither needs a person to remember where a fact went.
2.5 Freshness
Object
Refresh
Freshness objective
What happens if missed
Behavioral graph
Full rebuild, monthly
Complete by the fifth business day
Prior graph remains authoritative; programs are told which edition they are scored on
Digital activity
Daily
Landed before the daily loop opens in the first office
Loop runs on the previous day's activity and the field report says so
Client file
On receipt
Matched within one business day
Terrain brief notes the match date and rate
Placement logs
Hourly to daily by platform
Reconciled nightly
Reconciliation gap is an alert; unreconciled spend is reported as such
Registry snapshot
Monthly, with the graph
Versioned and frozen at snapshot
Programs continue on the snapshot named in their model cards
3
Identity resolution
Stable
Identity resolution answers one question for every event: which adult, or which household, does this belong to? It is designed to prefer a missed match over a wrong one, to work on hashed keys rather than names, and to make it impossible to join two clients' files even for someone with access to both.
3.1 Keys and hashing
Identifiers are normalized (case, whitespace, diacritics, address standardization to the postal authority's format, telephone to E.164) and then hashed with a keyed function. The key is scoped: licensed data is hashed under the practice environment's key; a client file is hashed under a key that exists only for that program and is destroyed at close. Because the keys differ, a hash from one program cannot be equated with a hash from another; the match to the graph happens once, inside the program's project, through a bridging step that holds both hashes in memory and persists neither. Raw identifiers from a client file are not retained beyond the match.
3.2 Tiers
Tier
Basis
Confidence band
Permitted uses
Deterministic
Exact agreement on one strong key (hashed email, hashed telephone) or two corroborating keys (name + standardized address, name + date of birth where lawful)
Standardized address plus surname agreement; the household is the unit when the source cannot see the individual
A (household), B (individual)
Household-level modeling and activation; individual attribution only with a band-A key
Probabilistic
Weighted agreement across sparse keys with a scored threshold; never a single weak key
B or C
B: modeling and cohort assignment. C: modeling only, weighted down; never activation, never measurement
Precision is preferred to recall throughout. The thresholds are tuned each month against a labeled reference set and the results are recorded on the graph edition's card; the match rate on a client file (the share of the file resolved at band A or B) is reported in the terrain brief so that the client knows how much of its own list the program can see.
3.3 Never-merge rules
Some pairs are never merged whatever the score: two records with different band-A keys of the same type (two distinct hashed emails claimed by one person are kept as two candidate identities until a stronger key reconciles them); a record flagged deceased by a registry source with any live activity; and any record that a data-subject request has placed on the suppression ledger, which is resolved only for the purpose of continuing to suppress it. Minors are not in the graph: sources are contracted to supply adults only, and a registry age below eighteen removes the record and its events at the next rebuild.
3.4 The household graph
Households are first-class nodes. An event from a household-level source is attributed to the household, not divided among its members; scores that use it are household scores, and the methodology notes require that assignment to the holdout is made at the household where any household score is in use, so that a treated member and a held-out member never share an address. The household graph is rebuilt with the individual graph each month.
4
The dictionary and the registry
Evolving
The dictionary is the vocabulary: about ten thousand topics in twenty-five domains, each a thing a person can be observed doing, buying, reading, joining, renewing or leaving. The registry is the working set: the seventeen thousand or so signals, built on those topics, that are live in models at any time. The dictionary changes by edition; the registry is pruned monthly.
4.1 Topics
A topic is a node with a stable identifier (T-<domain>.<n>), a definition written for a human curator, the source kinds that can evidence it, and its position in a shallow hierarchy (domain → group → topic; there is no fourth level, by design, because a taxonomy deep enough to be clever is too deep to be checked). Topics are assigned at ingestion by a deterministic mapping from each source's own categories, maintained per source and versioned with the dictionary edition. Where a source's category cannot be mapped without inference, the event is admitted with a domain and no topic, and it is not available to any model until a curator resolves the mapping.
4.2 Signals
A signal is a computation over events on one or more topics for one subject over a window: a recency, a frequency, a monetary or duration magnitude, a change, a co-occurrence. Every signal belongs to one of five families, and the family determines what a model may do with it.
Family
What it captures
Typical window
Decay
Behavioral enrichment
Recency, frequency and magnitude of observed acts on a topic
30 to 730 days
Exponential, family-default half-life; overridden per signal on evidence
Dispositions
Stable patterns across topics that have predicted how a person responds to an argument (risk posture, novelty, deference to institutions), derived from acts, not surveys
365 to 730 days
Slow
Occupational
Work and profession as evidenced by licensing, subscription and purchase behavior
730 days
Slow; refreshed on new evidence
Interests
Sustained attention to a subject across media and purchase
90 to 365 days
Moderate
Affinities
Attachment to institutions, places, brands and communities as evidenced by membership, renewal and giving
365 to 1,095 days
Slow
4.3 Lifecycle
A signal moves through four states. Proposed: written by a curator or generated by the registry's search over topic pairs, with a hypothesis about what it should predict. Candidate: computed for the full graph and evaluated in shadow (available to training, unavailable to any shipped model) across at least one program cycle. Active: admitted to the registry after it has entered at least one model that passed the publish gate; eligible for any program. Pruned: retired at a monthly review when it has not entered a passing model in twelve months, when its source is withdrawn, or when the exclusion check in Section 5.4 finds it proxying an excluded attribute. Pruning is not deletion: a pruned signal's definition is retained, versioned, so that a historical model card can be read.
The registry is curated toward one thing: behaviors that have predicted political decisions in prior programs, on the client's own holdout, against the two baselines. Volume is not a virtue. A signal that is merely interesting is not admitted; a signal that is predictive but proxies an attribute the data use policy excludes is pruned regardless of its performance. The curation review is monthly, minuted, and attended by the head of research, a program lead from each practice and counsel. Its decisions are ledger entries.
4.5 Domains
The twenty-five domains are grouped, for illustration, as commerce, media, civic, household, health-adjacent consumer behavior, and work. The full domain list, with definitions, is delivered under agreement with the segment map. Two rules govern the domains: no domain admits a topic that is itself an excluded attribute (a diagnosis, a religion, an orientation, a union card), and no signal may be constructed whose best available interpretation is one of those attributes, whatever the topics it is built from. The check that enforces the second rule is described in Section 5.4.
5
Models and the publish gate
Stable
A score is a probability between zero and one that a person like this one will do a particular thing: respond to an argument of a given kind, return a ballot, give, turn out, move. It is a weight, not a verdict. It is trained on recorded acts from prior programs, validated against two baselines on the client's own holdout, labeled with its confidence, and watched every day it is in the field. This section describes how a score earns the right to ship and how it loses it.
5.1 Families and catalog
The catalog is organized in five model families, named for the signal family that dominates their inputs: behavioral enrichment, dispositions, occupational, interests and affinities. Within a program, models from these families are composed into the program's task scores: response to an argument (persuadability by frame and register), turnout and participation, contribution and volunteering, institutional uptake (enrollment, attendance, compliance behavior for mandates), and movement (the change in a leading indicator over a window). The gate tests every model that contributes to a task score the program uses. A program's catalog is the set of models that have passed the gate for that program, and it is usually small: a statewide program runs on a catalog of six to twelve models, and the number is a fact of the model cards, not a target.
5.2 Training basis
Labels are recorded acts: a validated vote from a state's published history, a returned ballot, a contribution in a public filing or the client's own ledger, a sign-up, an enrollment. A label never comes from a survey answer or from another model's output. Training uses cycles before a cutoff date stated on the card; the current program's own outcomes are never used to train the models that score it. Features are drawn only from signals in the active registry as of the snapshot named on the card. Feature definitions are proprietary and not published; the families and the windows are.
5.3 The gate
CandidateTrained on prior outcomescutoff and snapshot stated on the card
Test 1Beats the file?the demographic baseline: the voter file, used well
Test 2Beats the platforms?the look-alike baseline on the same task
Test 3On your holdout?measured on the client's own control, by matched exposure
ShipsModel card issueddelivered with the segment map
Any failure → kill log. The entry names the baseline that beat the model, the margin, the edition and the date. Nothing is quietly retired.
Figure D2. The publish gate. Both baselines are built for every program; the margin over the stronger one is pre-registered on the card, by default 0.03 in AUC and ten percent relative in top-decile lift.
Discrimination is reported as area under the curve and as lift in the top decile; calibration is reported as a reliability table in ten bins and as the Brier score, and a model whose scores are discriminating but badly calibrated is re-fitted before it can ship, because a score that says 0.7 must mean seven in ten. Every figure on the card is computed on the client's holdout, never on the training data and never on a holdout drawn by the firm.
5.4 The exclusion check
Before the gate, every candidate model and every candidate signal is tested for whether it can predict an excluded attribute (the categories the data use policy names: health conditions, religion, sexual orientation, union membership, immigration status and the others listed there) from its inputs, using the labeled reference sets the firm maintains for the purpose and nothing else. An area under the curve above 0.60 against any excluded attribute retires the model or prunes the signal regardless of its performance on the task, and the kill log records which attribute. This is a structural check, run by code on every edition, and its thresholds are stated in the methodology notes.
5.5 Model cards
Every shipped model carries a card stating its family, training basis and cutoff, registry snapshot, validation split, both baselines and their figures, the gate margin, calibration, the exclusion-check result, the field kill threshold, and a named owner. The card is delivered to the client with the segment map on day three and is retained for seven years.
A shipped model is scored against its holdout every day the program is in the field. Its seven-day rolling discrimination is compared with the kill threshold on its card (by default the stronger baseline's figure); three consecutive days below it retire the model, rescore the cohort on the next model on the card or on the file, tell the client in that week's report, and write the kill-log entry with the drift figure. Drift in the inputs is watched separately, feature by feature, with a population-stability statistic, and an input that has drifted beyond its band triggers a retrain overnight in the office whose morning it is.
5.7 What a client receives, and does not
Clients receive scores for their program's cohorts, the model cards, the kill log and the readouts. They do not receive the models, the feature definitions or the training sets, and they may not re-identify a score to a person beyond the activation the program is licensed for. Scores are never delivered with an individual-level explanation; explanations are given at the cohort level, as the composite persons on the segment map, because a score describes what people like this do and not who this person is.
6
The studio
Evolving
The studio turns one approved master and the sponsor's own assets into the forty to sixty cohort-specific variants a statewide program runs, in minutes rather than weeks, with every variant carrying its disclosure, a claims check and a provenance manifest. It receives cohort descriptions and masters. It never receives records.
6.1 Inputs
Four things enter the studio for a program. The brand kit: the sponsor's guidelines, typefaces, palette, voice rules and existing creative, ingested once and compiled into a tokenized style specification that every variant is checked against. The claims library: every factual assertion the program is permitted to make, each with its source and its approval, and nothing else may be asserted. The master: the approved creative brief for a narrative track, with the seven-step arc marked. The cohort descriptions: the composite persons from the segment map with the frames and register each responds to, and no individual data of any kind.
6.2 Pipeline
Frame selection. For each cohort and step of the arc, the track names the frame; the studio selects the register-appropriate treatment of it.
Drafting. Copy and layout are generated from the master, the style specification and the frame, in the sponsor's voice, for each cohort, register and format. Drafting is assisted by the firm's own generation systems, which are trained and operated inside the platform.
Constraint checks. Every draft is checked by code before a person sees it: every assertion resolves to an entry in the claims library; the disclosure required for the register, jurisdiction and format is present and legible; the style specification is satisfied; reading level is within the cohort's band; contrast and caption rules for accessibility are met; no likeness, voice or name appears that is not in the approved asset list.
Human review. Two reviewers, one from the program and one from compliance, approve each variant. The second reviewer cannot be the drafter.
Release. The variant is versioned, stamped with its disclosure, given a provenance manifest and pushed to the seats for its flight. Nothing enters a seat that has not passed step four.
6.3 Provenance manifest
Each variant carries a signed manifest recording the master, the track and step, the cohort and register, the claims it makes and their library entries, the assets it uses and their licences, the reviewers, the disclosure text and the jurisdictions it was approved for. The manifest travels with the variant into the placement logs, so that a placement can be traced from the impression back to the claim and the person who approved it.
variant_id: V-0417-S03-PP-04-02 # program · cohort · register · arc step · n
master: MB-0417-T2@v3
track: "Cost of the everyday" (T2), step 4 of 7
cohort: S-03 "The Night Nurse"
register: precision_paid
formats: [social_1x1, ctv_15, mail_6x9]
claims: [CL-0417-018, CL-0417-021] # each resolves to a sourced, approved assertion
assets: [AS-0417-007 (sponsor-owned), AS-0417-012 (licensed, LIC-A-3391)]
disclosure: "Paid for by …" · jurisdiction CA · placement rule set 2026-07
reviewed_by: [program lead, compliance reviewer] # distinct from drafter
synthetic: none # no synthetic voice, likeness or name
released: 2026-09-12T16:03:55Z
6.4 Synthetic media
The studio does not produce a synthetic likeness, voice or name of any real person, including the sponsor. Generation is used for copy, layout, composition and variation of assets the sponsor owns or has licensed. Where a jurisdiction requires a label for generated content in political communication, the label is applied by the release step and recorded in the manifest, and the compliance packs list the jurisdictions that require it.
7
Seats and delivery
Stable
Every placement a program makes is bought from an account of record that the firm holds: social, search, connected television, programmatic and mail, with relational and earned channels run alongside. Single origin is what makes the media trail continuous from invoice to impression, and what makes it possible to keep the holdout out of every register at once.
7.1 Activation
Audiences leave the platform as keyed hashes of the identifiers the delivery platform accepts, matched inside the platform, and nothing else: no scores, no cohort names, no signals. An activation is scoped to a program and a flight and carries a time to live after which the platform is instructed to delete it. The holdout and the hold list travel to every seat as suppression lists before the first placement and are refreshed daily. A placement log entry that matches the holdout hash is an incident, and the check runs every day.
7.2 The three registers
Delivery runs in three disclosed registers, and the flighting plan assigns each cohort a mix. Precision paid is the sponsor's own voice placed at the person: social, search, connected television, programmatic and mail through the firm's seats. Organic advocacy is a real, disclosed second voice, sequenced with the paid register. Ambient and earned is the argument from the world. Where a program's population allows register arms, the contribution of each is measured separately; otherwise all three are measured together against the single holdout and the report says so.
7.3 Controls
Control
Rule
Enforced where
Suppression
Holdout; hold list; opt-outs and Global Privacy Control signals where applicable; registry-deceased; any record on the suppression ledger
Platform-side lists, refreshed daily; matched against logs nightly
Frequency
Caps per cohort, register and format, set in the flighting plan; a cohort whose register effect turns negative is suppressed in that register
Platform caps plus the daily loop's reallocation
Pacing
Weekly budget by cohort and register with the absentee and early-vote calendar marked; no flight may exceed its week without a decision record
Seat configuration; ledger
Disclosure
The variant manifest's disclosure must match the placement's jurisdiction and format before the seat accepts it
Release step; seat pre-flight check
Ad libraries
Every placement in a register the platform makes public is visible in that platform's library under the sponsor's disclosed name
Platform; verified weekly by the desk
7.4 The media trail
Each night, placement logs from every seat are reconciled to the accounts of record: spend by platform, register, cohort and variant, matched to invoices and to the flighting plan's week. The reconciled trail is a table with one row per placement carrying the variant manifest identifier, and a client may read it for their program at any time through the interfaces in Section 13. Unreconciled spend is reported as unreconciled, never absorbed.
8
Measurement
Stable
Measurement is the part of the platform that can see everything and change nothing. Its design is fixed in the methodology notes; this section describes how the platform carries that design out so that the result of a program is a difference between two groups that were the same on day five.
8.1 The fourteen-day timeline
Day 1Terrain briefoutcomes and covariates pre-registered
Day 3Segment mapcohorts sized; model cards issued
Day 5Holdout sealedframe frozen; keyed hash to the client, outside the delivery path
Day 6Balance tablestandardized differences; > 0.1 investigated
Day 7Message architecturetracks, arc, masters
Day 12Flighting planregisters by week, cohort and seat
WeeklyField reportlift with intervals, by cohort and register
CloseProof reportestimand, intervals, kill log, media trail
Figure D3. The first fourteen days and the two reports that follow. Nothing is placed before day five, and the seal is delivered before the first dollar is spent.
8.2 Drawing and sealing the holdout
Within each cohort, a simple random ten percent is assigned to the holdout by a seeded generator whose seed is recorded in the program file. Assignment is at the household wherever any household score is in use. The frame is frozen on day five and hashed; the assignment list is hashed with a keyed function and delivered to the client directly, outside the firm's delivery path. The firm's seats receive the holdout only as a suppression list. The delivery team cannot see which individuals are held out; the measurement team can. A cohort created mid-program receives a new holdout drawn the same way; an existing holdout is never re-drawn.
program: BW-C-2026-0417
sealed_at: 2026-09-05T21:14:03Z # before any placement
frame_hash: sha256:4b7e…c0a1 # frozen cohort frame, all members
assignment: hmac-sha256(program_key, member_ref ‖ arm) # key held by the client and measurement only
seed: recorded in program file, disclosed at close
arms: [holdout, precision_paid, paid_and_organic, all_registers]
unit: household # household scores in use
delivered_to: client counsel · outside the delivery path
suppression: seats receive hashes only, as a do-not-contact list
8.3 Estimation
The estimand is the intention-to-treat effect of assignment on the pre-registered outcome in the cohort's population: everyone assigned to treatment counts as treated whether or not the program reached them. Lift is the difference in means or proportions between treated and holdout; a regression-adjusted estimate using the pre-registered covariates and no others is reported beside the unadjusted one. Standard errors are clustered at the household where assignment was at the household. Intervals are ninety-five percent and two-sided. Register arms and cohort-by-frame cells are reported with a false-discovery-rate adjustment across the family of comparisons, and the report states the family.
8.4 Outcomes
The primary outcome is chosen from the highest available rank (a validated vote; a returned ballot; a recorded contribution, sign-up or enrollment; a registered institutional outcome) and pre-registered in the terrain brief. Where the primary outcome is an attitude, a panel is drawn from treated and holdout and instrumented identically at the same times. Leading indicators used in the daily loop are named in the brief and are never substituted for the primary outcome in the proof report.
8.5 Power
The terrain brief sizes every program from the methodology notes' power tables: eighty percent power, a two-sided five percent test, treated at least as large as the holdout. A cohort holdout of a thousand can see roughly a three- to four-point lift at ordinary base rates; five thousand sees under two; twenty thousand sees under one. A program underpowered for its client's question is redesigned before engagement, by widening the universe or pooling cohorts, and the firm will not run a holdout that cannot see a three-point effect.
8.6 Integrity
Two checks run for the life of the program. Every placement log is matched against the holdout hash every day, and a match is an incident under the security summary's timeline. Where the client can supply other vendors' contact records, holdout contact by those vendors is measured and reported as a limit. The balance table is re-run whenever the segment map changes. What invalidates a holdout is written in the methodology notes, and a program whose holdout is invalidated reports that fact with the same prominence as a result.
9
The daily loop
Stable
Every day a program is in the field, five things happen in order, and they happen in whichever of the seven offices it is morning. The loop is a scheduled pipeline with a human decision at the end of it, and every decision it produces is a ledger entry the client can read.
Ingest. New behavior since midnight enters the graph and is resolved to people and households. Digital activity lands before the loop opens; if it has not, the loop runs on the previous day's activity and the field report says so.
Rescore. Every score the program uses is recomputed for every member of every cohort. The hold list is refreshed: members whose scores have crossed the cohort's floor move out of contact, and the seats receive the updated suppression list.
Reweight. Spend moves toward the variants and registers that moved people yesterday, within the flighting plan's approved lines. Optimization within a line is the firm's responsibility and is reported weekly; a reallocation of more than ten percent of any budget line requires the client's written approval, recorded as a decision record before it takes effect.
Read. The measurement team reads, by cohort and register, the treated-minus-holdout difference on the leading indicator, its sign, its trend and its interval, and applies the three rules: reflight where a register's effect exceeds its threshold, suppress where it is negative, retrain where a model breaches its kill threshold.
Retrain and retire. Models whose inputs have drifted are retrained overnight; models that have breached their kill threshold for three consecutive days are retired and their cohorts rescored on the next model on the card, or on the file.
9.1 Timing
The loop is scheduled per program in the program's local time, opening after the day's digital activity has landed and closing before the first placement window of the day. The office whose business day it is holds the watch; handoff between offices is a written note in the program's ledger, and the desk's board shows every program's loop state (queued, running, read, decided, handed off) to every office at once. Seven offices cover fourteen hours of the clock; for the remaining hours, alerts page the on-call lead in the office that opens next.
9.2 Rules and thresholds
Condition
Action
Recorded where
Told to the client
Reflight
A cohort's register effect exceeds its threshold: weight in that register is raised for that cohort within the week's envelope
Reallocation log
Weekly report
Suppress
A cohort's register effect is negative with its interval below zero: that register is suppressed for that cohort; if all registers, the cohort moves to the hold list
Suppression log with the date
Weekly report, with the same prominence as a gain
Retrain
A model breaches its kill threshold: model retired; cohort rescored; next model or the file takes over
Kill log
Weekly report
Contamination
A placement log matches the holdout hash
Incident record
Within the security summary's timeline
10
Ledger and provenance
Stable
The ledger is the platform's memory of what was decided, by whom, on what evidence. It is append-only, hash-chained and per program, and a client may read their program's ledger at any time. It is the mechanism behind the firm's sign-off gates and the record an independent reviewer works from.
10.1 Structure
Figure D4. A program's ledger is a chain: each entry carries the hash of the entry before it, so that an entry cannot be altered or removed without breaking every entry after it. The chain head is delivered to the client with each weekly report.
10.2 Entry types
Type
Prefix
Written when
Must carry
Decision record
DR
A gate is passed, a plan is signed, an envelope is exceeded, a source is released from quarantine, a program is opened or closed
The decision, the evidence pointed to, the signers, the alternative not taken
Kill-log entry
KL
A model fails the gate or the field threshold; a signal is pruned; a variant is withdrawn
What was killed, by what test, by what margin, the edition, the date, who owned it
Seal
SL
A holdout or arm assignment is frozen
Frame hash, assignment scheme, unit, seed reference, recipient
Access event
AC
Any read of record-level data; any export; any privilege change
Identity, scope, purpose code, volume
Handoff
HO
The watch passes between offices
State of every open item; the receiving lead
Destruction
DX
Program data is destroyed at close; a data-subject request is executed
Lineage of everything destroyed; the certificate reference
{
"id": "DR-0417-0038",
"program": "BW-C-2026-0417",
"type": "reallocation",
"at": "2026-09-23T14:02:11Z",
"office": "the office whose morning it was",
"decision": "raise precision-paid weight for S-03 from 22% to 31% of week-3 envelope",
"evidence": ["RO-0417-W03-S03-PP"], // readout: +2.9 pts on returns, 95% CI [0.8, 5.0]
"rule": "reflight",
"alternatives": ["hold weight; wait one more day of readout"],
"signers": ["program lead", "measurement lead"],
"prev": "sha256:51bd…8e03",
"hash": "sha256:e7f2…4c19"
}
10.3 Sign-off gates
The governance policy requires the client's written approval of strategy, creative masters and the media plan before execution, and of any reallocation of more than ten percent of a budget line. The platform enforces this as decision records the client signs at fixed points: the terrain brief (day one), the segment map and model cards (day three), the message architecture and masters (day seven), the flighting plan (day twelve), any reallocation above the threshold, and close. The seal (day five) is recorded the same way with the client as recipient. The platform will not release a placement, a variant or an export that depends on an unsigned record.
10.4 Audit rights
A client may, under their engagement letter, read their program's ledger in full, verify its chain, obtain the seeds and hashes at close, and appoint an independent reviewer who is given read access to the measurement project for the program. The firm's own annual independent review, described in the methodology notes, samples program files, checks holdout seeds, balance tables and estimation code, and publishes its findings in the transparency report.
11
Security
Stable
The security summary is the firm's public statement of its controls and the timelines it commits to; the full independent assessment, penetration-test summary and incident runbook are available under non-disclosure. This section is the engineering view of the same controls: where in the platform each one lives.
11.1 Identity and access
Named accounts only, through a single identity provider with phishing-resistant multi-factor authentication for every person and every administrative action. Access is granted per program, from the roster, and expires with the program; it is recertified quarterly by the program lead and revoked the same day on any role change. Administrative access is a separate identity from analytical access, and administrative sessions are recorded. Service identities are short-lived and bound to a workload; there are no long-lived shared credentials anywhere in the platform.
11.2 Encryption and keys
TLS 1.2 or higher for every connection, internal and external; AES-256 for storage, backups and exports. Keys live in a dedicated key-management service with one hierarchy per practice, rotated on a schedule and on any suspected compromise; key administrators and data administrators are different people. Client files are hashed on receipt with a keyed function whose key is per program and destroyed at close, which is what makes a client's hashed records unmatchable to any other program's.
11.3 Segregation, network, endpoints
Separate environments per practice, separate projects per program, separate seats per practice. No inbound path to a data environment exists except through the identity provider; components speak over private connectivity; egress is restricted to approved destinations (delivery platforms, licensors, the hosting provider), and any new destination is a change with a reviewer. Endpoints are managed, encrypted, patched and monitored; no program data is stored locally; a personal device cannot reach a data environment.
11.4 Logging and detection
Access and change logs for every environment are immutable, retained for one year and reviewed monthly by someone other than the administrator whose actions they record. Alerts fire on anomalous access, bulk reads, exports and privilege changes; the export of any record-level data is an alert by default. Every placement log is matched daily against the holdout hash, and a match is an incident. Logs feed the ledger's access events, so a client can see who read their program's record-level data and why.
11.5 Change and vulnerability management
Every production change is reviewed by a second engineer, is logged, and is reversible; secrets are kept out of repositories and scanned for. Dependencies and configuration are scanned continuously; critical vulnerabilities are patched within seven days and high within thirty. An independent penetration test of internet-facing and internal systems runs annually and after any material change, with tracked remediation.
11.6 Continuity
Every environment is backed up, encrypted, to storage separate from production; restoration is tested quarterly and the results recorded. Recovery objectives: measurement and delivery within four hours; the studio within one business day. Operational continuity comes from the desk: a program in the field is watched by the office whose business day it is, with a written handoff.
11.7 Incident response
An incident is declared by the named lead within one hour of confirmation; counsel is engaged from the first hour; affected identities are revoked, keys rotated and environments isolated as required, with evidence preserved; clients whose data is affected are notified within seventy-two hours with what is known, what is being done and what they should do; a written post-incident review follows, and the transparency report counts incidents by class each year.
12
Privacy engineering
Stable
The privacy and data use policies say what the firm will and will not do with personal data. This section describes the machinery that makes those statements true without relying on anyone's memory: retention classes enforced by the lifecycle system, a suppression ledger that every path consults, a request pipeline that uses lineage to find every copy, and jurisdiction rules that decide where a program's data may live.
12.1 Minimization
The platform holds what a program needs and no more. Raw identifiers are hashed on receipt and not retained beyond the match; the studio receives cohorts, never records; delivery receives hashes, never scores; measurement receives assignment and outcomes, never creative. Excluded attributes are not collected as topics and are not permitted as signals, and Section 5.4 describes the check that prevents them being reconstructed.
12.2 Retention classes
Class
Covers
Retained
Destroyed
C
Client file and derived program data (scores, cohorts, audiences, readouts)
Life of the program
Within ninety days of close; certificate issued
M
Holdout assignment and measurement outcomes
Per the client's instruction; for validated-turnout measurement, until the state publishes its history
On the client's instruction; certificate issued
A
Activation audiences inside delivery seats
Life of the flight (time-to-live set at activation)
Deleted from seats at close; deletion confirmed in the media trail
L
Licensed behavioral data
Per provider terms; superseded on refresh
Superseded copies destroyed on refresh
R
Model cards, kill logs, proof reports, ledgers
Seven years
Secure deletion
G
Access and change logs
One year
Secure deletion
12.3 Data-subject requests
A request to access, correct, delete or opt out is verified, recorded and executed by a pipeline that uses lineage to find every derived row that depends on the subject's records. Deletion removes the subject's events, resolved identity and household edge from the graph at the next rebuild and from every program project immediately, and places a keyed hash of the identifier on the suppression ledger so that the subject is not re-admitted by a later drop. Opt-outs, including a Global Privacy Control signal where the law gives it effect, are written to the same ledger and propagate to every seat's suppression list at the next daily refresh. Requests are acknowledged and completed within the periods the applicable law sets, and the transparency report counts them each year.
12.4 Jurisdiction and residency
Every program carries a jurisdiction profile from its terrain brief, and the profile decides three things by rule: which disclosure and labeling requirements the studio must satisfy; which suppression signals must be honored; and where the program's data may be stored and processed. Institutional mandates that require residency are held in-country in the hosting provider's regional facilities, with replication disabled or confined as the mandate specifies, and a residency and replication attestation is delivered at program start and at any change. The compliance packs summarize the rules the profile encodes, jurisdiction by jurisdiction.
12.5 Certified destruction
At close, the lifecycle system enumerates every object in class C and A for the program by lineage, destroys it, verifies the destruction, and writes a destruction entry to the ledger with the lineage list. A certificate signed by the head of the desk and counsel is issued to the client within ninety days, and the program's keyed hash key is destroyed with the data.
13
Interfaces and exports
Evolving
A client sees their program through three interfaces: the six documents, delivered on the fourteen-day schedule; signed export bundles, machine-readable and verifiable offline; and a read-only program interface for clients whose engagement letter provides for it. All three read from the same tables the firm reads from, so there is no client version of the truth.
13.1 Export bundles
Every weekly field report and the proof report are accompanied by a bundle: the report as a document, its tables as comma-separated and columnar files, the data dictionary for each table, the model cards in force that week, the kill log to date, the reconciled media trail and the ledger chain head. The bundle carries a manifest listing every file and its digest, and the manifest is signed with the firm's publishing key, whose public half is delivered with the engagement letter and can be confirmed out of band with the program lead. A client can verify a bundle without any access to the platform.
$ tar -xf BW-C-2026-0417-W03.tar
$ sha256sum -c manifest.sha256 # every file's digest matches the manifest
readouts/lift-by-cohort-register.csv: OK
readouts/lift-by-cohort-register.parquet: OK
cards/M-DISP-0142.yaml: OK
kill-log.jsonl: OK
media-trail/week-03.csv: OK
ledger-head.json: OK
$ ssh-keygen -Y verify -f brushworks-publishing.pub -I publishing@brwrks.com \
-n bw-export -s manifest.sha256.sig < manifest.sha256
Good "bw-export" signature for publishing@brwrks.com with ED25519 key
13.2 The program interface
Where the engagement letter provides for it, a client's named users may read their program through a versioned, read-only interface. Access is by short-lived tokens scoped to one program and issued to named identities through the client's own single sign-on; institutions may require mutual TLS. Every read is an access event in the ledger. The interface exposes what the documents expose and nothing more: no scores at the individual level, no signals, no models. Resources are versioned in the path and paginated by cursor.
Resource
Returns
Notes
GET /v1/programs/{program}
Program definition, jurisdiction profile, gate status, edition identifiers
Gate status shows which of the six sign-offs are recorded
Models, feature definitions, training sets, the registry and dictionary as artifacts, individual-level scores or explanations, any other client's anything. Exports of record-level data are alerts by default and require a purpose code that the ledger records; the only routine record-level export is the hashed activation to the firm's own seats.
14
Reliability and operations
Evolving
The platform is operated by the desk: seven offices, one board, one set of identities, a written handoff. This section states the objectives the desk runs to and what a client is told when one is missed.
14.1 Service objectives
Objective
Target
Measured how
If missed
Daily loop completion
Closed before the first placement window, every day in field
Loop state on the board; ledger timestamps
Field report states the days the loop ran late and on what data
Suppression propagation
Holdout, hold list and opt-outs at every seat within the daily refresh
Seat list versions matched to the suppression ledger
Any placement to a suppressed record is an incident
Media reconciliation
Nightly, with unreconciled spend under one percent of the week
Reconciliation report
Unreconciled spend reported as such, never absorbed
Graph edition
Monthly rebuild complete by the fifth business day
Edition record
Programs continue on the prior edition and are told
Program interface availability
99.9 percent monthly, excluding announced maintenance
Synthetic probes from three regions
Reported in the next field report
Recovery
Measurement and delivery within four hours; studio within one business day
Quarterly restoration tests
Test records available under agreement
14.2 Change management
Platform releases are versioned and scheduled; a release that changes any behavior a document in Section 0.2 depends on is a versioned change to that document with notice to clients in the field. Changes roll out by environment, the international practice last, with a rollback rehearsed before the release and a decision record for every production change. No change ships during the final ten days before a program's election day unless it is a security fix, and a security fix is released with a note in the ledger of every program in the field.
14.3 Environments
Three tiers: development (synthetic data only; no licensed or client data ever), staging (licensed data under the practice's terms; no client data) and production (per practice, per program). Promotion between tiers is a reviewed change; production data is never copied downward. The studio's generation systems are trained and operated in production environments under the same identities and logs as everything else.
14.4 Status and support
Clients in the field have a named program lead and the desk's incident line. Questions about this document go to research@brwrks.com; security disclosures to privacy@brwrks.com, which is read by the head of the desk and counsel, and acknowledged within one business day.
15
Conventions
Stable
15.1 Identifiers
Object
Format
Example
Program
BW-<practice>-<year>-<seq>, practice C (candidate), I (independent expenditure), N (international and institutional)
Timestamps are ISO 8601 in UTC with a trailing Z; calendars (flighting, absentee drops, election day) are in the program's local time and say so. Probabilities are decimals in [0, 1]; lift is in percentage points; intervals are ninety-five percent, two-sided, written as a pair. Hashes are SHA-256 unless stated; keyed hashes are HMAC-SHA-256 with the key named. Text is UTF-8; names of people never appear in any table a client receives, and cohorts are always their composite.
15.3 Versioning
This document is versioned by edition and dated. Model cards, registry snapshots, dictionary editions, playbooks and variants are versioned independently and referenced by identifier; a program's readouts name the editions they were computed on. A change to a Stable section is announced to clients in the field and listed in the changelog with the date it took effect.
16
Glossary
Stable
Arm
A randomly assigned subset of a cohort's treated population that receives a defined register mix, used to measure each register's marginal contribution.
Baseline
A model a candidate model must beat: the voter file used well, and the platforms' look-alike targeting, both built for every program.
Catalog
The set of models that have passed the publish gate for a given program.
Cohort
A group on the segment map who share what would move them, drawn from behavior, named as a composite person and sized before design.
Composite
The named, described person who stands for a cohort in every document; never a real individual.
Data plane / control plane
The data plane holds records, models and logs; the control plane holds the definitions, access lists, ledger and gates that decide what the data plane may do.
Decision record
A ledger entry stating a decision, its evidence, its signers and the alternative not taken.
Dictionary
The ten-thousand-topic, twenty-five-domain vocabulary that every event is assigned into at ingestion.
Drop
One delivery from one source, tested as a unit before admission.
Edition
A versioned, frozen snapshot of the graph, registry, dictionary or catalog.
Envelope
The normalized record every observed behavior is written into; also, in flighting, a week's budget bound for a cohort and register.
Estimand
The intention-to-treat effect of assignment to the program on the pre-registered outcome, in the cohort's population.
Excluded attribute
A category the data use policy forbids as a topic, a signal or a prediction target, and that the exclusion check tests for.
Field kill threshold
The in-field discrimination level, stated on the card, below which a shipped model is retired after three consecutive days.
Hold list
Members moved out of contact for the program because their scores crossed a cohort floor or a register effect turned negative.
Holdout
The random ten percent of each cohort sealed on day five to receive nothing from the program in any register.
Keyed hash
An HMAC over a normalized identifier with a key scoped to a practice or a program; the form in which identifiers exist inside the platform and leave it.
Kill log
The program's record of every model, signal or variant retired, with the test it failed and the margin.
Ledger
The append-only, hash-chained record of decisions, seals, access, handoffs and destruction for a program.
Lineage
The identifiers of the drops and editions a derived row was computed from, carried on every row.
Master
The approved creative brief for a narrative track from which the studio produces variants.
Media trail
The reconciled record from invoice to placement to impression, by register, from the accounts of record.
Model card
The document every shipped model carries: family, basis, baselines, gate result, calibration, exclusion check, kill threshold, owner.
Practice
One of three separately walled parts of the firm: candidate, independent expenditure, international and institutional.
Program
A bounded engagement with one client, one purpose, one file and one measurement design, from terrain brief to proof report.
Publish gate
The pre-registered test a model passes before it is used: beat both baselines, by the stated margin, on the client's own holdout.
Register
One of three disclosed channels of persuasion: precision paid, organic advocacy, ambient and earned.
Registry
The working set of signals that are live and eligible for models, pruned monthly.
Seat
An account of record the firm holds on a delivery platform, through which every placement is bought.
Seal
The act, and the ledger entry, of freezing a holdout or arm assignment and delivering its keyed hash to the client.
Signal
A computation over events on one or more topics for one subject over a window, in one of five families.
Task score
A program-level probability composed from catalog models: response to an argument, turnout, contribution, institutional uptake, movement.
The desk
The seven-office operations function that watches every program in the field and runs the daily loop wherever it is morning.
Variant
One cohort-, register- and format-specific creative produced from a master, versioned, disclosed and carrying a provenance manifest.
17
Changelog
Stable
Edition
Date
Change
Docs 1.0
September 2026
First public edition. Consistent with methodology notes v1.0, the security summary (September 2026), the data use policy and the governance policy as published on this site.
Cite as: Brushworks (2026), Technical documentation, Docs 1.0, September 2026, brwrks.com/docs. Corrections and questions: research@brwrks.com.
Reports
What the firm publishes, and how to read it.
The methodology every program is measured by; the models we retired; what we register, disclose and decline; how the firm is secured; and what a prospective client receives under agreement. Written to be checked.
A year at the gate: shipped and killed · illustrative
Open as PDFA4 · opens in the browser · cite as Brushworks (2026), Methodology notes
The standard of proof, written down. These notes govern every field report and proof report the firm issues; the Review's papers cite them.
In short. Ten percent of every cohort is drawn at random before any spend, sealed, and reported to the client directly. Lift is the intention-to-treat difference between treated and control, on recorded acts where they exist, reported with a ninety-five percent interval, by register and cohort, weekly. A program that cannot produce this is not a Brushworks program.
1. The holdout
Within each cohort on the segment map, a simple random sample of ten percent of individuals (household-attributed records are sampled at the household) is assigned to the control on day five, before the first media plan is approved. The assignment list is hashed and delivered to the client; the firm's seats receive it as a suppression list. Cohorts smaller than two thousand are pooled for measurement with their nearest neighbor on the map, and the pooling is stated in the report. The control is never contacted by the program in any register; contact by the campaign's other vendors is recorded where the client can supply it and reported as contamination.
2. Register-level holdouts
Where the program's size allows, the treated population is further divided at random into arms that receive precision paid only; paid and organic; and all three registers in sequence. The differences between arms are the registers' marginal contributions; the interaction is reported separately. Where size does not allow, registers are measured against the single control and the limit is stated.
3. Estimand and outcomes
The estimand is the intention-to-treat effect: a voter assigned to treatment who was never reached counts as treated. Outcomes are recorded acts where they exist (validated turnout from the state's record, returned ballots, sign-ups, contributions, event attendance) and panel measurement of favorability and intent where an act does not exist, using the same instrument in both arms at the same times. Self-reported turnout is never used as an outcome.
4. Intervals and power
Holdout size
Minimum detectable lift, base rate near half
Design implication
2,000
4.4 points
Pool cohorts; measure registers together
5,000
2.8 points
Cohort-level lift; registers pooled
10,000
2.0 points
Cohort and register lift separately
20,000
1.4 points
Frame-by-cohort tests in the second month
40,000+
1.0 point or better
Decay and backfire watched daily with confidence
Intervals are ninety-five percent, two-sided, computed on the difference in means (or proportions) between arms; where assignment is at the household, standard errors are clustered at the household. Power is eighty percent against a treated group of similar size measured on the same instrument; the terrain brief states, before the program starts, which effects the design can see.
5. Baselines and the publish gate
A model is used only if, on the client's own sealed holdout, it discriminates the outcome better than two baselines: the voter file used as a competent consultant would use it (participation history, party, precinct, age band) and the delivery platforms' look-alike targeting seeded with the campaign's supporters. Discrimination is reported as area under the curve on a time-split validation; the gate is a pre-registered margin over the stronger baseline. Models are also tested for reconstructing an excluded attribute by proxy and retired if they do. Every shipped model carries a card and a field kill threshold.
6. What is reported, and what is not
Weekly: lift by cohort and register with intervals; reallocations; suppressions; models retired. At close: the proof report: design, lift, decay, backfire, the kill log, and the invoice-to-impression media trail. Never: raw outcomes among the treated presented as effects; the race's result presented as the program's; a cohort's negative lift omitted from the report.
Methodology notes, version 1.0, September 2026. Reviewed annually by an independent reviewer; revisions posted here with a new version number.
Kill-log summary
Open as PDFA4 · opens in the browser · cite as Brushworks (2026), Kill-log summary
Every model retired at the publish gate or in field, in aggregate, with the reasons. Program-level logs are delivered to each client and held under agreement.
Launch edition. The firm publishes this summary each year. The table below is the format, populated with representative proportions from the firm's programs; the first audited annual count, for the year ending September 30, 2026, is published with the annual report in spring 2027. Counts here should be read as the shape of a year, not as audited totals.
Family
Tested
Failed the file
Failed look-alike
Proxy reconstruction
Shipped
Killed in field
Behavioral enrichment
22
8
3
1
10
2
Dispositions
19
5
2
1
11
3
Occupational
14
7
1
0
6
1
Interests
16
9
2
1
4
0
Affinities
13
6
1
0
3
1
All families
84
35
9
3
34
7
Reasons, defined
Failed the file. Did not beat a competently used voter file on the client's holdout for the outcome. The file is used instead, and the report says so.
Failed look-alike. Beat the file but not the platforms' own targeting seeded with supporters. Not used for that channel.
Proxy reconstruction. Discriminated an excluded attribute (ethnicity, religion, ideology) above the tolerance. Retired regardless of performance.
Killed in field. Shipped, then fell below its pre-set discrimination threshold on the daily watch; the cohort was rescored by the next model or the file.
Kill-log summary, launch edition, September 2026. Next edition: with the annual report, spring 2027.
Transparency report
Open as PDFA4 · opens in the browser · cite as Brushworks (2026), Transparency report
What the firm registered, disclosed, declined, received and resolved, in aggregate, twice a year.
Launch edition. The first reporting period runs from the firm's public launch on September 18, 2026 to December 31, 2026, and is published in January 2027. This edition sets out what will be reported and how each figure is defined, so that the first numbers can be read against a fixed standard.
Reported
Definition
Basis
Registrations
FARA and LDA registrations filed and active; foreign-influence and third-party registrations abroad, by jurisdiction
Public registries
Disclaimers
Programs on air; paid placements carrying disclaimers; placements visible in platform ad libraries; deficiencies found and corrected
Compliance reviews; ad libraries
Practice wall
Certification issued for the period; any exception, with the counsel's finding
Wall certification
Mandates declined
Count by category of the declination policy (opposing sides; not describable in public; sanctions; violence or civil-rights denial; voter suppression; foreign influence on a U.S. election; undisclosed messengers or synthetic people)
Acceptance screen records
Concerns
Received, acknowledged within two business days, investigated, resolved, outstanding; anonymous share
concerns@ log under counsel
Incidents
Security, compliance or operational incidents; clients notified within seventy-two hours; regulators notified
Incident log
Rights requests
Access, deletion, correction and opt-out requests received and answered within statutory time; Global Privacy Control signals honored
privacy@ log
Models retired
Pointer to the kill-log summary
Kill logs
Transparency report, launch edition, September 2026. First period report: January 2027; semiannual thereafter.
Security summary
Open as PDFA4 · opens in the browser · cite as Brushworks (2026), Security summary
The public overview of how program data is protected. The independent assessment, residency and replication attestations and restoration-test records are available under non-disclosure agreement.
Area
Standing control
Evidence under NDA
Access
Named-user accounts, multi-factor authentication, least privilege scoped to a program, quarterly recertification, same-day revocation
Access matrix and recertification records
Encryption and keys
TLS 1.2+ in transit; AES-256 at rest; keys in a dedicated service with rotation and separation of duties
Configuration attestation
Segregation
Separate environments per practice; client data isolated; delivery seats separated by practice
Environment diagram; wall certification
Monitoring
Immutable access and change logs retained one year, reviewed monthly; alerts on anomalous, bulk or export access
Log samples; review records
Vulnerability management
Continuous dependency scanning; severity-based patch windows; annual independent penetration test with tracked remediation
In-country holding for mandates that require it; replication attestations to the client
Attestations
Incident response
Written runbook; on-call across the desk; client notice within seventy-two hours; post-incident review within thirty days
Runbook summary; aggregate incident log
Security summary, September 2026. Independent assessment: annual; summary available under NDA to clients and prospective clients.
Annual report, Vol. I
Open as PDFA4 · opens in the browser · cite as Brushworks (2026), Annual report, Vol. I
Spring 2027, in preparation. The record for the cycle, published in one document.
The annual report will carry: the contests of the cycle by office and state, with the programs' aggregate lift against holdouts and the losses with their post-mortems; the kill-log summary with audited counts; the transparency figures for both periods; the governance ledger with the independent-review summaries; the Review's papers of the year in a single volume; and the policies as revised. It is published on this page and delivered to every client of the cycle.
Section
Contents
Basis
The record
Contests by office and state; aggregate lift vs holdouts; the losses
Proof reports
Models
Audited kill-log summary; the catalog by family
Kill logs
Transparency
Both semiannual periods, combined
Transparency reports
Governance
Wall certification; independent-review summaries; concerns and outcomes
Independent reviewers; counsel
The Review
The year's papers, bound
Research desk
Diligence pack
Open as PDFA4 · opens in the browser · cite as Brushworks (2026), Diligence pack
For prospective clients, committee treasurers, donors' counsel and ministries, under non-disclosure agreement.
Document
What it shows
Wall certification
The candidate and independent-expenditure practices' separation, certified by counsel for the current cycle
Sample proof report (redacted)
A full program report: design, lift by cohort and register with intervals, decay, backfire, kill log, media trail
Sample kill log and model card
The gate in practice: baselines, results, reasons; what a shipped model documents
Processor list
Every processor by category, with the terms that bind it
Sample provenance record
What is recorded about a data source before it is used
Independent-review summaries
Security controls; the practice wall; the measurement methodology
Carrier's certificate
The underwriting behind the performance assurance, for qualifying political programs
Compliance review template
The day-ten checklist, unfilled
Request the pack through the contact page (choose "a government or institution" or "a campaign"), and say which documents you need first. The NDA is the firm's mutual form; counsel's redlines are welcome.
What data and political firms must obey, country by country.
An expert-level map of the privacy, political-advertising, communications, foreign-influence and synthetic-media rules that bind firms like ours (the United States, Canada, Europe and the rest of the world) with the thresholds, the definitions that decide a case, the enforcement that has actually happened, and what each regime means in practice for a program. Published free. Not legal advice; the questions to bring to counsel, and the answers to expect.
The approval chain, signed at each step · illustrative
Reading the packs
What this is. Four regional packs and a set of cross-cutting analyses, written for the people who have to decide whether a program can run: compliance leads, counsel, campaign managers, ministry project officers and the data vendors that serve them. Each entry names the instrument, the trigger that brings a firm within it, the obligations that bite in political and data work, the enforcement record, and the practical consequence. Where the law is unsettled or litigated, the entry says so. Currency: the firm's review of September 2026, reviewed with each site release. Statutes, thresholds and guidance move monthly; treat every row as the question to ask, not the answer to rely on.
58jurisdictions across four regions
6kinds of rule: privacy, political advertising, communications, foreign influence, synthetic media, data brokers
14enforcement actions that changed how this work is done
Three definitions recur and decide most cases. Sensitive or special-category data, in most of the world, political opinions are on the list, so the act of scoring someone's likely vote is itself a regulated act; in the United States it is not, and the regulated acts are the sale, the call, the disclaimer and the foreign principal. Sale or sharing: the transfer of personal data to a third party for value or for cross-context advertising, which most U.S. state laws let residents opt out of and which several treat a browser signal as exercising. Political advertising: now defined in statute in the European Union, the United Kingdom, Canada and Australia, each with a labeling and record-keeping regime that platforms enforce before regulators do.
Ten insights an expert should take from the packs
The United States regulates the channel; the rest of the world regulates the data. American law is a patchwork of communications, campaign-finance and consumer-protection rules with no general privacy statute and no treatment of political opinion as sensitive. Almost everywhere else, a political-opinion score is special-category data whose mere processing needs a legal basis. A firm that models behavior rather than ideology is compliant by design in both worlds; a firm that models ideology needs consent it cannot get at scale outside the United States.
Platforms are now the first regulator. Meta and Google withdrew paid political and issue advertising from the European Union in October 2025 rather than meet Regulation 2024/900's targeting and transparency duties; their ad libraries, verification and disclaimer rules are enforced in hours, not months. A program's compliance review must read platform policy as law, because the practical sanction, account loss mid-flight, is faster than any fine.
"Register before you persuade" is now global. FARA and the LDA in the United States, FIRS in the United Kingdom (commenced July 2025), FITS in Australia, third-party regimes in Canada, the United Kingdom and New Zealand, transparency registers in the European Union: the filing date must precede the first activity, and the filings are public. A registered program has nothing to find; an unregistered one is a criminal exposure for the firm and a scandal for the client.
Consent for calls and texts is nearly universal, and political speech is not the exemption people think it is. The TCPA exempts nothing political; the Do Not Call registry does. Canada's CASL exempts donation solicitations but not much else; PECR treats party messages as direct marketing. The FCC's 2024 ruling that AI voices are "artificial" and the $6 million forfeiture for the New Hampshire primary robocall settled the question for synthetic voice.
Synthetic media is regulated by three different theories at once. Labeling (the EU AI Act, roughly half of U.S. states, Brazil, China), prohibition near elections (Korea's ninety-day ban, several U.S. states, Singapore, Brazil), and existing fraud and impersonation law (the FCC, the FEC's fraudulent-misrepresentation provision). Several U.S. state laws have been enjoined on First Amendment grounds; the theory that survives everywhere is disclosure.
The data-broker registries are the new perimeter. California's Delete Act builds a one-stop deletion mechanism that brokers must honor from August 2026; Texas, Oregon and Vermont maintain registries; the FTC has banned the sale of sensitive location data by order. A licensor that is not registered where it must be is a provenance failure for every downstream program.
Party exemptions protect parties, not vendors. Australia exempts registered parties from the Privacy Act; PIPEDA does not reach federal parties; most U.S. state laws exempt nonprofits and committees. None of these reach a vendor, which is covered in its own right, and British Columbia's regulator has held that its private-sector law reaches federal parties operating in the province.
Cross-border transfers are solved on paper and fragile in practice. The EU–U.S. Data Privacy Framework survived its first challenge in 2025; standard contractual clauses and the UK addendum carry the rest; Brazil, India, Saudi Arabia, China and Vietnam add localization or adequacy conditions of their own. Institutional mandates increasingly require in-country holding, which is a design decision, not a paperwork one.
Enforcement follows the scandal, and the scandal is always the same one. Cambridge Analytica produced the ICO's fine, the FTC's order, Canada's joint finding, Australia's settlement and the EDPB's statement; the New Hampshire robocall produced the FCC's rule. Regulators act on political data when a case is vivid, and the cases are vivid when consent was missing and the messenger was hidden.
The cheapest compliance is a method that does not need the exemption. Behavior instead of identity, disclosed messengers, one account of record, a sealed control and a written declination policy satisfy most of these regimes without a special legal basis. The compliance review then checks paperwork rather than redesigning the program.
United States
No comprehensive privacy statute; a federal layer for campaign finance, foreign agents, communications and consumer protection; twenty state consumer-privacy laws with different thresholds and exemptions; four state registries for data brokers; and a wave of state law on artificial intelligence in political advertising, some of it enjoined.
Federal instruments
Instrument
Trigger
Obligations that bite
Enforcement and record
In practice
Federal Election Campaign Act and FEC regulations (52 U.S.C. §30101 et seq.; 11 C.F.R.)
Any public communication for or against a federal candidate; any expenditure coordinated with a campaign; any contribution
Disclaimers with sponsor and "paid for by" language on public communications (§30120; 11 C.F.R. §110.11), including internet communications placed for a fee; coordination test with payment, content and conduct prongs (§109.21) and a firewall safe harbor for vendors serving both sides (§109.21(h)); prohibition on foreign-national contributions, expenditures and decision-making (§30121); fraudulent misrepresentation of campaign authority (§30124), which the FEC has said reaches deceptive AI content under existing law
FEC civil penalties and conciliation; DOJ criminal prosecution for knowing and willful violations; the FEC declined in 2024 to write AI-specific rules, relying on §30124
The compliance review confirms disclaimer wording per channel, documents the firewall between practices, and screens every funder for foreign-national status; vendor payments appear in the committee's reports
Acting within the United States as an agent of a foreign principal in political activities, public relations, publicity or political consulting; "foreign principal" includes governments, parties, and foreign persons and entities
Registration within ten days of agreeing to act and before acting; conspicuous statement on informational materials disseminated to two or more persons; supplemental statements every six months; retention of records; the LDA exemption (§613(h)) is unavailable where the principal is a foreign government or party
DOJ National Security Division FARA Unit; criminal penalties for willful failure; advisory opinions published since 2018; enforcement rose sharply after 2016 and continues
The firm registers before any activity that touches the United States, labels every informational material, and reports semiannually; the acceptance screen refuses any mandate aimed at a U.S. election
Lobbying Disclosure Act (2 U.S.C. §1601 et seq.)
Lobbying contacts with covered federal officials above the statutory time and income thresholds
Registration within 45 days of the first contact; quarterly activity reports (LD-2); semiannual contribution reports (LD-203); FARA applies instead where the client is a foreign government or party
Clerk of the House and Secretary of the Senate; civil and criminal penalties; DOJ enforcement
Institutional mandates that involve official contacts are registered under whichever regime applies before the first meeting
Any call or text to a mobile number made with an automatic telephone dialing system or an artificial or prerecorded voice; any prerecorded call to a residential line
Prior express consent for autodialed or prerecorded calls and texts to mobiles (political messages included); the Supreme Court's narrow reading of "autodialer" (Facebook v. Duguid, 2021) exempts human-initiated peer-to-peer texting but not prerecorded or AI voice; the FCC's February 2024 declaratory ruling treats AI-generated voices as "artificial"; revocation of consent by any reasonable means must be honored within ten business days (rules effective April 2025); accurate caller identification; time-of-day limits
FCC forfeitures and private actions at $500–$1,500 per call or text with class treatment; the FCC fined the consultant behind the January 2024 New Hampshire primary deepfake robocall $6 million and the carrier $1 million; the FCC's 2024 "one-to-one consent" rule was vacated by the Eleventh Circuit in January 2025
Autodialed texting needs documented consent; P2P texting runs on human-initiated platforms with a suppression list; no synthetic voice, ever; every opt-out honored across channels
Telemarketing Sales Rule and the National Do Not Call Registry (16 C.F.R. Part 310)
Telemarketing of goods, services or charitable solicitations
Political calls are outside the TSR and the national registry, but remain subject to the TCPA's consent rules and to state political-call statutes; caller-ID accuracy and abandonment rules apply to any seller
FTC; state attorneys general under the Telemarketing Act
A political program cannot rely on the DNC exemption to send autodialed texts
CAN-SPAM Act (15 U.S.C. §7701 et seq.)
Commercial electronic mail
Political email is largely outside the statute, but fundraising email with a commercial element, deceptive headers or subject lines, and failure to honor opt-outs can fall within it; mailbox providers enforce their own reputation rules regardless
FTC; state attorneys general; private ISP actions
Accurate headers, a working unsubscribe honored within ten days, and physical address on every message as a matter of practice
FTC Act §5 and data-broker enforcement (15 U.S.C. §45)
Unfair or deceptive acts affecting commerce
Privacy statements must be true; the FTC's 2019 order against Cambridge Analytica treated deceptive harvesting of profile data as unfair; 2024 orders against X-Mode/Outlogic and InMarket banned the sale of sensitive location data and required deletion; the Kochava litigation established that selling precise location revealing visits to sensitive sites can be an unfair practice
Consent orders with twenty-year terms, deletion and data-minimization requirements, civil penalties for order violations
Provenance records for every licensed source; no precise location; no data revealing visits to sensitive places; privacy policy statements audited against practice
Online services directed to children under thirteen, or actual knowledge of a child
Verifiable parental consent before collection; the 2025 rule amendments tighten third-party disclosure consent and retention
FTC civil penalties; state attorneys general
The firm holds no data on minors; age-gated sources are excluded at provenance review
Platform political-advertising policies
Any paid political or issue placement on Meta, Google, Microsoft, X, TikTok (which bans political ads) and the connected-television and programmatic platforms
Advertiser verification in the sponsor's name; "paid for by" disclaimers; public ad libraries retaining ads for years; AI-generated content disclosure (Meta, Google since 2023–2024); targeting restrictions on political ads in some platforms; issue-ad definitions broader than statute
Ad rejection, account suspension, library annotation: applied algorithmically and appealed slowly
Seats are verified in the sponsor's name before day twelve; disclaimer language is set per platform; a rejected creative is a program delay, so masters are checked against platform policy at the studio stage
State consumer-privacy laws
Twenty states have comprehensive privacy statutes in force. They share a structure (rights of access, deletion, correction and portability; opt-outs from targeted advertising, sale and consequential profiling; opt-in consent for sensitive data; data-protection assessments for high-risk processing; controller–processor contracts) and differ on four points that matter for political work: whether nonprofits and political committees are covered; whether a browser opt-out signal must be honored; what counts as sensitive data; and whether the attorney general must offer a cure period. No state lists political opinion as sensitive data. Most exempt nonprofits; a vendor to a nonprofit is covered in its own right.
State
Statute; in force
Applicability threshold
Notable for political and data work
California
CCPA as amended by the CPRA (2023); Delete Act (SB 362)
For-profit businesses with roughly $26 million revenue, or 100,000 consumers or households, or half of revenue from selling or sharing
"Share" for cross-context behavioral advertising is an opt-out right; Global Privacy Control must be honored; sensitive personal information includes racial or ethnic origin, religion and union membership with a right to limit use; risk assessments and cybersecurity audits under 2025 regulations; the California Privacy Protection Agency is a dedicated regulator with early enforcement against Honda and a retailer for defective opt-out flows; data brokers register with the CPPA and must process consumer deletion requests through the DROP mechanism from August 2026; political committees and nonprofits are outside "business"
Virginia
VCDPA (2023)
100,000 consumers, or 25,000 with half of revenue from sale
Opt-in for sensitive data; nonprofits exempt; no private right of action; cure period; AG enforcement
Colorado
CPA (2023); AG rules
100,000 consumers, or 25,000 with revenue from sale
Universal opt-out mechanisms binding since 2024; nonprofits are covered; detailed rules on profiling, dark patterns and consent refresh; amendments on minors and biometric data
Connecticut
CTDPA (2023, amended 2025)
Thresholds lowered by the 2025 amendments; sensitive-data processing alone can trigger coverage
Opt-out signals binding; expanded sensitive-data list; minors' protections; profiling rights
Utah
UCPA (2023)
$25 million revenue and consumer thresholds
Narrow rights, notice-and-opt-out for sensitive data, no assessments; the lightest of the twenty
Texas
TDPSA (2024); Data Broker Law (2023)
Any business that processes personal data and is not a small business under SBA definitions
No numeric threshold, so most vendors are covered; opt-out signals binding since 2025; data brokers register with the Secretary of State and post notices; the attorney general has been the most active state enforcer, with actions against automakers and brokers over driving and location data
Oregon
OCPA (2024); data-broker registry (2024)
100,000 consumers, or 25,000 with revenue from sale; nonprofits covered from July 2025
Consumers may obtain the list of specific third parties that received their data; opt-out signals binding; brokers register with the Department of Consumer and Business Services
Montana
MCDPA (2024, amended 2025)
50,000 consumers, lowered by amendment
Opt-out signals binding; amendments removed the cure period and added minors' protections
Florida
Digital Bill of Rights (2024)
Controllers with more than $1 billion in revenue and specified business models; some provisions apply more broadly
Sensitive-data and children's provisions reach smaller entities; treat as a large-platform law
Delaware
DPDPA (2025)
35,000 consumers, or 10,000 with revenue from sale
Nonprofits covered except narrow categories; opt-out signals binding from 2026
Iowa
ICDPA (2025)
100,000 consumers
Narrower rights; opt-out rather than opt-in for sensitive data; long cure period
Nebraska
NDPA (2025)
Texas-style: any non-small business
Broad reach; opt-out signals binding
New Hampshire
NHPA (2025)
35,000 consumers, or 10,000 with revenue from sale
Opt-out signals binding; AG rulemaking
New Jersey
NJDPA (2025)
100,000 consumers, or 25,000 with revenue from sale
Covers financial data; opt-out signals binding; regulations from the Division of Consumer Affairs; nonprofits not exempt
Tennessee
TIPA (2025)
$25 million revenue plus consumer thresholds
Affirmative defense for programs conforming to the NIST Privacy Framework
Minnesota
MCDPA (2025)
100,000 consumers, or 25,000 with revenue from sale
Right to question the result of profiling and learn what data was used; documented privacy program and inventory required; opt-out signals binding
Maryland
MODPA (2025)
35,000 consumers, or 10,000 with revenue from sale
The strictest minimization standard in the country: collection limited to what is reasonably necessary; sale of sensitive data prohibited outright; sensitive data includes national origin, citizenship and immigration status; targeted advertising to minors prohibited
Indiana
ICDPA (2026)
100,000 consumers
Virginia-style rights; long cure period
Kentucky
KCDPA (2026)
100,000 consumers
Virginia-style; cure period
Rhode Island
RIDTPPA (2026)
35,000 consumers, or 10,000 with revenue from sale
Requires listing third parties to which data is sold; no cure period
Political-specific state rules
Area
The rule
Where it is strictest
What to do
Voter registration files
Every state licenses its file on its own terms; most restrict use to election, political, scholarly or journalistic purposes and prohibit commercial use, resale or public posting; some require an applicant's sworn statement
California (certified purpose, no commercial use), Pennsylvania and Virginia (restricted purposes); a few states sell to anyone
Obtain the file under the state's terms, record the permitted purpose in the provenance record, never commingle with commercial licensing, and destroy at close where the license requires
Artificial intelligence in political advertising
By 2026 roughly half of the states require disclosure of AI-generated or materially altered content in political ads, or prohibit deceptive deepfakes of candidates within a window before an election, with candidate injunctive remedies
Texas (criminal, 2019), California (three 2024 statutes, one enjoined as viewpoint-based), Minnesota, Michigan, Washington, New York, Florida, Arizona, Utah, Wisconsin, Indiana, New Mexico, Oregon, Colorado and others
Label AI-assisted content where required in the statute's words; never produce synthetic depictions of real people; confirm each state's window, remedy and pending litigation in the compliance review
State disclaimers and reporting
State campaign-finance law adds "paid for by", top-funder and committee-name requirements, and independent-expenditure reports within 24 to 72 hours near elections
California (top three funders on committee ads), Washington, New York, Colorado, Alaska
The compliance review confirms exact wording per jurisdiction and channel before air
Political calls and texts
State statutes regulate political robocalls and texts beyond the TCPA: identification at the start, hours, opt-out, state registries, and outright bans in a few states
Florida (consent for automated political texts), Oklahoma (mini-TCPA), Washington, Indiana, Wyoming, Arkansas and Louisiana (bans or registries)
Consent for autodialed texts; state-specific identification; a suppression list that honors every opt-out everywhere
Data brokers
Registration and disclosure regimes in California, Vermont, Texas and Oregon; California's DROP deletion mechanism from 2026
California (CPPA registry, DROP, audits), Texas (registration and security), Oregon (registry with consumer lookup)
Every licensor is checked for registration; no license from a broker that is not registered where it must be
Canada
A federal private-sector privacy law that does not reach political parties, provincial laws that sometimes do, the strictest anti-spam statute in the world, and an election law that reaches third parties, foreign spenders and calling services.
Instrument
Trigger
Obligations that bite
Enforcement and record
In practice
Personal Information Protection and Electronic Documents Act (PIPEDA)
Collection, use or disclosure of personal information in the course of commercial activity, federally and in provinces without substantially similar law
Meaningful consent appropriate to sensitivity; purpose limitation; openness; access and correction; safeguards; mandatory breach reporting and record-keeping; accountability for transfers to processors; federal political parties are not "commercial activity" and are outside the Act, but their vendors are within it
Privacy Commissioner of Canada investigates and issues findings; damages through the Federal Court; the 2019 joint finding with British Columbia that Facebook contravened PIPEDA in the Cambridge Analytica–AggregateIQ matter; Bill C-27, which would have replaced PIPEDA and regulated AI, died in January 2025
Consent documented per purpose; profiling disclosed in plain language; processors bound by contract; breach clock started on discovery
Quebec: Act respecting the protection of personal information in the private sector, as amended by Law 25
Any enterprise collecting Quebec residents' data, wherever established
Person in charge of protection (a privacy officer) by default the CEO; privacy impact assessments before projects involving personal information and before transfers outside Quebec; consent by purpose in clear language; disclosure of automated decisions and profiling with the means to deactivate; portability since 2024; transparency about technology that identifies, locates or profiles; administrative penalties to CAD 10 million or 2 percent of worldwide turnover and penal fines to CAD 25 million or 4 percent
Commission d'accès à l'information; the strictest regime in North America and the closest to the GDPR
Programs touching Quebec run a PIA before the terrain brief, document the transfer assessment, and present profiling controls in French
Alberta and British Columbia Personal Information Protection Acts
Private-sector organizations in the province; BC's Act reaches non-commercial organizations
Consent and access regimes substantially similar to PIPEDA; BC's Information and Privacy Commissioner held in 2022–2024 that PIPA applies to federal political parties collecting personal information in the province, a position upheld by the BC Supreme Court
Provincial commissioners; orders and, in Alberta, offences
A party program in British Columbia is treated as covered regardless of the federal gap
Canada's Anti-Spam Legislation (CASL)
Any commercial electronic message sent to or from a Canadian address; installation of software; alteration of transmission data
Express or implied consent, sender identification and a working unsubscribe on every commercial message; messages by or on behalf of registered parties, candidates and nominees soliciting contributions are exempt, but other campaign messaging with a commercial element is not; administrative penalties to CAD 10 million per violation for organizations
CRTC investigates and levies penalties; the Competition Bureau and the Privacy Commissioner share enforcement; a private right of action remains suspended
Donation solicitations rely on the exemption; anything else is sent on documented consent with identification and unsubscribe
Canada Elections Act (as amended by Bill C-76, 2018)
Parties, candidates, third parties and calling-service providers in federal elections
Third parties spending above CAD 500 on regulated activities during the pre-election and election periods must register with Elections Canada, report and observe spending limits; foreign entities are prohibited from spending to influence electors; partisan and election advertising must carry a tagline identifying the sponsor; registered parties must publish a privacy policy (a weak requirement Bill C-65 proposed to expand before it died); calling-service providers and their clients register with the CRTC's Voter Contact Registry and retain scripts and recordings
Commissioner of Canada Elections; administrative monetary penalties and prosecution; CRTC for the Voter Contact Registry
Third-party status is determined at acceptance; taglines on every master; calling services registered before the first call
CRTC Unsolicited Telecommunications Rules
Telemarketing and automatic dialing-announcing devices
Political calls are exempt from the National Do Not Call List but must identify the caller and honor calling hours; ADAD calls must begin with identification and provide a contact number; internal do-not-call lists must be kept
CRTC administrative monetary penalties
Identification scripts and hours by time zone across seven offices
Provincial election acts
Provincial and municipal campaigns and third parties
Separate third-party registration, spending limits and advertising blackouts; Quebec's Election Act restricts third-party spending most tightly; Ontario's regime was litigated to the Supreme Court in 2025 over pre-election third-party limits
Provincial chief electoral officers
Provincial programs are reviewed against the province's own regime, not the federal one
Europe
The most developed regime in the world, in four layers: a general data-protection law in which political opinions are special-category data; an EU regulation on political advertising that the largest platforms answered by withdrawing the product; platform duties under the Digital Services Act and labeling duties under the AI Act; and national election and marketing law on top. The United Kingdom mirrors the structure with its own instruments.
European Union
Instrument
Trigger
Obligations that bite
Enforcement and record
In practice
General Data Protection Regulation (Regulation (EU) 2016/679)
Processing by a controller or processor established in the EU, or of data subjects in the EU when offering them goods or services or monitoring their behavior (Art. 3)
A lawful basis for every purpose (Art. 6); political opinions, religious beliefs, ethnic origin and health are special-category data whose processing is prohibited unless explicit consent or a narrow condition applies (Art. 9): scoring likely vote is processing political opinions; transparency to the person (Arts. 13–14) including the logic of profiling; rights of access, erasure, restriction, objection to direct marketing (absolute) and to profiling, and not to be subject to solely automated decisions with significant effects (Arts. 15–22); data-protection impact assessments before large-scale profiling (Art. 35); a DPO where core activities involve large-scale special-category processing (Art. 37); a representative for non-EU controllers (Art. 27); records of processing (Art. 30); breach notice to the authority within 72 hours (Art. 33); transfers only with adequacy, standard clauses or another Chapter V mechanism
National authorities coordinated by the European Data Protection Board; fines to €20 million or 4 percent of worldwide turnover; EDPB Statement 2/2019 on personal data in political campaigns; national fines on political profiling (Italy's Rousseau platform, 2019); the EU–U.S. Data Privacy Framework (2023) survived the Latombe challenge in the General Court in September 2025
The firm does not process political opinions as data: behavior only, no ideology scores, which removes the Art. 9 problem for the program's data; DPIAs before every EU program; the objection right honored as a permanent suppression; transfers under the DPF or standard clauses with a transfer assessment
ePrivacy Directive (2002/58/EC) as implemented nationally
Electronic communications, cookies and direct marketing by electronic means
Consent for cookies and trackers beyond the strictly necessary; consent for automated calls, SMS and email marketing to individuals (with a narrow existing-customer exception); national variations in the treatment of political messaging, generally as direct marketing
National data-protection and telecom authorities; cookie fines have been the most frequent enforcement in several states
No tracking on the firm's site; SMS and email programs run on consent held by the client or partner of record, verified before use
Regulation (EU) 2024/900 on the transparency and targeting of political advertising
Political advertising (preparation, placement, promotion, publication or dissemination of a message by, for or on behalf of a political actor, or liable to influence an election, referendum, legislative process or voting behavior) in the EU; applies from 10 October 2025
Every political ad labeled as such with the sponsor, the entity ultimately controlling it, the amounts, the election or referendum, and a link to a transparency notice; a European repository of political advertisements; targeting and ad delivery only with explicit consent collected separately for that purpose, never on special-category data (Art. 9 GDPR) and never to minors; sponsors from outside the EU barred from paid political advertising in the three months before an election or referendum; due diligence and record-keeping for publishers and intermediaries
National authorities, with data-protection authorities enforcing the targeting rules under GDPR penalties; Meta and Google announced that they would cease political, electoral and social-issue advertising in the EU from October 2025 rather than comply
Paid political advertising in the EU is now largely a matter of publisher inventory, television and out-of-home under national law; EU programs run through organic advocacy, earned media and disclosed placements with the transparency notice, and never on behavioral targeting of political opinion
Digital Services Act (Regulation (EU) 2022/2065)
Intermediaries and platforms serving EU users; heightened duties for very large online platforms and search engines
No advertising based on profiling with special-category data (Art. 26); no profiling-based ads to minors (Art. 28); ad repositories with sponsor, payer, targeting parameters and reach for very large platforms (Art. 39); systemic-risk assessment and mitigation including electoral processes (Arts. 34–35); the Commission's 2024 election guidelines on integrity measures
European Commission for very large platforms, with fines to 6 percent of turnover; national Digital Services Coordinators
Platform ad repositories are the public record of a program; the compliance review checks that placements appear correctly there
Providers and deployers of AI systems in the EU market; general-purpose model providers
Transparency duties (Art. 50): deployers of systems generating deepfakes must disclose that content is artificially generated or manipulated, with a narrower duty for text published to inform the public; AI systems intended to influence the outcome of elections or voting behavior are listed as high-risk (Annex III, point 8(b)), triggering risk management, data governance, human oversight and registration duties on providers; general-purpose model obligations since August 2025; most Art. 50 and high-risk obligations apply from August 2026
National market-surveillance authorities and the AI Office; fines to €35 million or 7 percent for prohibited practices, €15 million or 3 percent for most other breaches
AI-assisted content is labeled where the Act requires; no synthetic depictions of real people; the firm's models predict responsiveness to a message, not voting behavior, and the compliance review documents why the Annex III listing does not attach
United Kingdom
Instrument
Trigger
Obligations that bite
Enforcement and record
In practice
UK GDPR and the Data Protection Act 2018, as amended by the Data (Use and Access) Act 2025
Processing by UK-established organizations or of UK data subjects when offering services or monitoring behavior
Same structure as the EU regulation; political opinions are special-category data, with a specific condition in Schedule 1 permitting registered parties to process political opinions as part of their legitimate activities, subject to the right to object; DPIAs, DPOs and records as in the EU; the 2025 Act eases some research and legitimate-interest rules, restructures the ICO as the Information Commission, and lifts PECR penalties to UK GDPR levels
Information Commissioner's Office: the 2018 investigation into data analytics in political campaigns, the £500,000 fine against Facebook under the previous law, the audits of all major parties in 2020 with published recommendations, and fines against Leave.EU and an insurer for cross-use of data; fines to £17.5 million or 4 percent
Parties rely on the Schedule 1 condition; vendors do not: the firm's UK work processes behavior, not opinion, and honors objection as permanent suppression
Privacy and Electronic Communications Regulations 2003 (PECR)
Electronic direct marketing: calls, texts, emails, cookies
Consent for marketing calls to TPS-registered numbers, for automated calls, and for texts and emails to individuals; the ICO treats political parties' promotional messages as direct marketing; penalties raised by the 2025 Act
ICO monetary penalties; a substantial record of fines for texting and calling without consent
Texting and email on documented consent only; TPS screening; no automated calls
Elections Act 2022, Part 6: digital imprints
Paid digital political material anywhere; unpaid material from parties, candidates, campaigners and certain others
Since November 2023 an imprint with the promoter's name and address on all paid digital political material and on unpaid material from regulated actors; the Electoral Commission's guidance covers where the imprint sits when the platform cannot carry it
Electoral Commission investigation and civil sanctions; criminal offence for failure
The imprint is set on every master before it leaves the studio
Political Parties, Elections and Referendums Act 2000 (non-party campaigners)
Spending on regulated campaign activity above the registration threshold during the regulated period
Registration with the Electoral Commission, spending limits, donation controls and reporting; foreign-source funding restrictions
Electoral Commission
Third-party status confirmed at acceptance; spending tracked against the regulated period
National Security Act 2023, Part 4: Foreign Influence Registration Scheme (FIRS)
Arrangements to carry out political-influence activities in the UK at the direction of a foreign power; an enhanced tier for specified states
Registration within ten days of the arrangement and before activity, since the scheme commenced on 1 July 2025; the enhanced tier (initially covering Russia and Iran) registers a wider range of activity; public register for the political-influence tier
Home Office; criminal offences for failure to register
Any institutional mandate from a foreign government touching UK political influence is registered before the brief is written
National layers in the EU
Member state
Instruments
What matters for political and data work
Germany
Bundesdatenschutzgesetz; Gesetz gegen den unlauteren Wettbewerb §7; Bundesmeldegesetz §50; state election laws
Consent required for telephone and email advertising to consumers; parties may obtain registry extracts (name, address, doctorate) in the six months before an election, for that election only; state data-protection authorities have examined party microtargeting on social platforms; the Federal Constitutional Court's tradition on informational self-determination shapes strict readings
France
Loi Informatique et Libertés; CNIL guidance on political communication; Code électoral (L.52-1, L.49); Loi du 22 décembre 2018 on information manipulation
Paid political advertising on television and radio is prohibited; commercial advertising for electoral purposes is banned in the six months before an election, including online; campaigning is silent from the eve of the vote; CNIL's guidance governs prospecting, consent and the use of the electoral roll; platforms must publish sponsor and payment information for issue content during campaign periods under the 2018 law
Spain
LOPDGDD (2018); LOREG Art. 58 bis
The 2018 law permitted parties to compile political opinions from public web sources within the electoral framework; the Constitutional Court struck the provision down in 2019 (STC 76/2019) as an unjustified interference with fundamental rights: the clearest judicial statement in Europe that opinion profiling by parties needs a precise legal basis
Italy
Codice privacy; the Garante's decisions
The Garante fined the Five Star Movement's Rousseau platform operator in 2019 for security and profiling failures, and has issued guidance on political communication and the use of membership data
Netherlands
UAVG; the Authority for Personal Data's guidance; the DDMA code
The regulator has warned parties against political microtargeting without consent and audited campaign apps; Dutch practice is consent-heavy for email and SMS
Ireland
Data Protection Act 2018; Electoral Reform Act 2022
The Data Protection Commission supervises most platforms' EU operations, so platform ad-library and targeting behavior is shaped by Irish enforcement; the 2022 Act created an Electoral Commission and provides a framework for online political-advertising transparency
Poland, Czechia, Nordic and Baltic states
GDPR with national election codes
Consent-based marketing regimes; election silence periods; sponsor identification on political material; several authorities have issued campaign-period guidance on data use
The rest of the world
Most major markets have a comprehensive privacy law modeled loosely on the GDPR; the local variations that matter are whether political opinion is sensitive, whether parties are exempt, whether the law localizes data, and whether the election authority regulates online advertising and synthetic media directly.
Americas
Jurisdiction
Instruments
Obligations that bite
Enforcement and record
Brazil
Lei Geral de Proteção de Dados (2020); Superior Electoral Court (TSE) Resolution 23.610/2019 as amended in 2024
Political opinion and party membership are sensitive data requiring specific, highlighted consent; legal bases, DPIA-like reports on request, DPO appointment, international transfer rules (ANPD regulation 2024, with standard clauses); the TSE's 2024 electoral rules require labeling of AI-generated content, prohibit deepfakes, bar chatbots simulating candidates, restrict paid boosting to candidates and parties, and hold platforms liable for failing to remove disinformation promptly
ANPD administrative sanctions since 2021 (warnings, fines to 2 percent of Brazilian revenue capped at R$50 million per infraction); the TSE removed content and sanctioned campaigns in the 2024 municipal elections
Mexico
Ley Federal de Protección de Datos Personales en Posesión de los Particulares (new law, 2025); INE regulations on the electoral roll and propaganda
Consent and privacy notices; ARCO rights; the 2025 reform dissolved the independent regulator INAI and moved oversight to a government secretariat, a change international counsel flag for adequacy and independence; the electoral authority controls the padrón electoral and regulates campaign propaganda and the use of citizen data
Administrative fines; electoral sanctions by INE and the electoral tribunal
Argentina
Ley 25.326 (2000); reform bill pending
Database registration with the Agencia de Acceso a la Información Pública; political opinion is sensitive; EU adequacy since 2003; the pending reform would align with the GDPR
Administrative sanctions; habeas data actions
Chile
Ley 21.719 (published December 2024; in force December 2026)
GDPR-style rights and bases, sensitive data including political opinion, DPO-like roles, transfer rules, and a new Personal Data Protection Agency with fines to 4 percent of turnover
The new Agency from 2026
Colombia
Ley 1581 (2012); Decree 1377
Registration of databases with the National Database Registry; sensitive data includes political affiliation and requires explicit consent; the electoral authority regulates propaganda
Superintendencia de Industria y Comercio fines and orders
Peru, Uruguay, Costa Rica, Panama, Ecuador
Peru Law 29733; Uruguay Law 18.331 (EU-adequate); Costa Rica Law 8968; Panama Law 81; Ecuador LOPDP (2021)
Consent-based regimes; political opinion sensitive in each; registration duties in Peru and Costa Rica; Ecuador's law is GDPR-modeled with a superintendency
National authorities of varying activity
Middle East and Africa
Jurisdiction
Instruments
Obligations that bite
Enforcement and record
United Arab Emirates
Federal Decree-Law 45/2021 (PDPL); DIFC Data Protection Law 2020; ADGM Data Protection Regulations 2021
Consent-based federal regime with sensitive-data rules and cross-border restrictions, executive regulations still developing; the free-zone laws are GDPR-close with DPO and DPIA duties and active commissioners; political campaigning is not a lawful activity: institutional and public-communication work only, with government counterparties, under the relevant emirate's media rules
Federal Data Office; DIFC and ADGM commissioners with published fines
Saudi Arabia
Personal Data Protection Law (Royal Decree M/19, 2021, amended 2023; enforced from September 2024) and implementing regulations
Consent or another basis; data-subject rights; records; DPO where required; transfer rules with adequacy and safeguards; registration of certain controllers with SDAIA; no political campaigning: institutional communication only
SDAIA; fines and, for sensitive-data disclosure, criminal penalties
Israel
Protection of Privacy Law, Amendment 13 (in force August 2025); Elections Law
Modernized regime with a strengthened Privacy Protection Authority, DPO duties for large processors, database registration narrowed, and administrative fines; political opinion is sensitive; after the 2020 leak of the voter registry through a party's campaign app, use of the registry is tightly conditioned
Privacy Protection Authority; fines under Amendment 13
Turkey
Law 6698 (KVKK), amended March 2024
Explicit consent or a listed condition for sensitive data including political opinion; controllers' registry (VERBIS); the 2024 amendments aligned cross-border transfer rules with the GDPR (adequacy, standard clauses, binding rules)
Personal Data Protection Authority fines
South Africa
Protection of Personal Information Act (in force 2021); Electoral Act
Political persuasion is special personal information (s.26) with a specific authorization for political parties (s.31) that does not extend to vendors; direct marketing by electronic means requires consent or an existing-customer relationship (s.69); prior authorization for certain processing; cross-border rules (s.72)
Information Regulator with enforcement notices and administrative fines to R10 million; the Regulator's first fines fell on public bodies in 2023
Nigeria
Nigeria Data Protection Act 2023
Registration of data controllers and processors of major importance; political opinion is sensitive; DPO duties; transfer rules; the Commission has issued directives and begun enforcement
Nigeria Data Protection Commission; fines to 2 percent of annual revenue
Kenya
Data Protection Act 2019
Registration of controllers and processors; political opinion is sensitive; the Office of the Data Protection Commissioner has penalized organizations, including a political party sanctioned for registering members without consent
ODPC penalties to KES 5 million or 1 percent of turnover
Ghana
Data Protection Act 843 of 2012
Registration with the Data Protection Commission; consent; sensitive data includes political opinion; the Electoral Commission's rules govern the voters' register
Data Protection Commission
Egypt, Morocco, Tunisia
Egypt Law 151/2020 (regulations pending); Morocco Law 09-08; Tunisia Organic Law 2004-63 with reform pending
Licensing and registration regimes; sensitive data including political opinion; enforcement developing
National authorities
Asia-Pacific
Jurisdiction
Instruments
Obligations that bite
Enforcement and record
Japan
Act on the Protection of Personal Information (amended 2022); Public Offices Election Act
Purpose specification; opt-out third-party provision with notification to the PPC; "special care-required" data (including creed) needs consent; records of third-party transfers; the election law permits internet campaigning but restricts paid online advertising to parties and bans it for candidates, with strict timing rules
Personal Information Protection Commission orders and fines
South Korea
Personal Information Protection Act (amended 2023); Public Official Election Act (amended December 2023)
Consent-centric regime with separate consent for sensitive data including political opinion; large administrative fines (to 3 percent of related revenue); the election law bans deepfake campaign content in the ninety days before an election and requires labeling outside that window, with the National Election Commission prosecuting violations in the 2024 election
Personal Information Protection Commission; NEC
China
Personal Information Protection Law (2021); Data Security Law; Deep Synthesis Provisions (2023); AI labeling measures (2025)
Consent and separate consent for sensitive data; localization and security assessment for transfers by important operators; synthetic media must be labeled visibly and in metadata; no political campaigning exists; institutional communication under state media rules only
Cyberspace Administration of China; fines to 5 percent of turnover
India
Digital Personal Data Protection Act 2023 (rules notified 2025 with phased implementation); Election Commission advisories
Consent or legitimate use; notice; consent managers; significant data fiduciaries with DPO and audit duties; children's data with verifiable consent; the Election Commission's 2024 advisories require labeling of AI content and removal of deepfakes within hours during campaigns
Data Protection Board with penalties to INR 250 crore per breach
Singapore
Personal Data Protection Act 2012 (amended 2020); Parliamentary Elections Act; Elections (Integrity of Online Advertising) (Amendment) Act 2024; POFMA
Consent and Do Not Call registry; mandatory breach notification; online election advertising must identify the publisher and be paid for by or on behalf of registered actors; the 2024 Act bans realistic manipulated content depicting candidates during the election period; falsehood correction directions under POFMA
PDPC fines to S$1 million or 10 percent of turnover; Returning Officer directions
Indonesia
Personal Data Protection Law 27/2022 (fully effective October 2024)
GDPR-style bases and rights; political views are specific (sensitive) data; DPO duties; transfer rules; a supervisory agency still being stood up; election commission rules on campaign content
Administrative fines to 2 percent of revenue; criminal provisions
Malaysia
Personal Data Protection Act 2010, amended 2024 (phased in 2025)
Consent; breach notification and DPO appointment added in 2024; cross-border transfer whitelist replaced by adequacy assessment; Election Commission oversight of campaign material
Personal Data Protection Commissioner
Vietnam
Decree 13/2023 on personal data protection; Personal Data Protection Law (passed June 2025, in force January 2026)
Consent and impact assessments filed with the Ministry of Public Security; sensitive data includes political views; localization for certain operators; no political campaigning
Ministry of Public Security; fines under the 2025 law
Thailand
Personal Data Protection Act 2019 (in force 2022); Election Commission regulations
GDPR-modeled bases; political opinion sensitive; DPO for large-scale processing; the Election Commission regulates campaign spending and online material
PDPC administrative fines
Philippines
Data Privacy Act 2012; COMELEC rules on social media and AI (2025)
Consent; registration of data-processing systems; political affiliation is sensitive; COMELEC requires registration of official campaign accounts and platforms, and regulates AI-generated content in campaigns
National Privacy Commission; COMELEC
Australia
Privacy Act 1988 as amended by the Privacy and Other Legislation Amendment Act 2024; Foreign Influence Transparency Scheme Act 2018; Commonwealth Electoral Act 1918
Registered political parties are exempt from the Privacy Act (s.6C) and members of parliament for political acts (s.7C), but their vendors are not; political opinion is sensitive information; the 2024 reforms add a statutory tort for serious invasions of privacy, a children's online privacy code, automated-decision transparency and higher penalties; electoral matter must carry an authorisation statement; foreign-influence arrangements, including on behalf of foreign governments and parties, must be registered under FITS; foreign donations are restricted
Office of the Australian Information Commissioner (which settled with Meta for A$50 million in 2024 over Cambridge Analytica); Attorney-General's Department for FITS; Australian Electoral Commission
New Zealand
Privacy Act 2020; Electoral Act 1993
Principles-based regime with breach notification; election advertisements must carry a promoter statement; third-party promoters register above spending thresholds; foreign donations capped
Privacy Commissioner; Electoral Commission
Comparison matrix
Twelve jurisdictions, eight questions a compliance lead asks first.
Jurisdiction
Political opinion sensitive?
Parties exempt?
Vendor covered?
Ad labels required?
Foreign-influence register?
Synthetic-media rule?
Consent for texts?
Browser opt-out signal?
United States (federal)
No
n/a
Sectoral
Yes (FEC)
Yes (FARA)
Fraud law; FCC on voice
Yes (TCPA)
No
California
No
Yes
Yes
Yes (state)
–
Yes (partly enjoined)
Yes
Yes
Canada (federal)
No
Yes (PIPEDA)
Yes
Yes (tagline)
Third-party regime
No
Yes (CASL, partial)
No
Quebec
Yes (sensitive)
Partly
Yes
Yes
Provincial
No
Yes
No
European Union
Yes (Art. 9)
No
Yes
Yes (2024/900)
Transparency register
Yes (AI Act)
Yes (ePrivacy)
No
United Kingdom
Yes
Sch. 1 condition
Yes
Yes (imprints)
Yes (FIRS)
No (fraud law)
Yes (PECR)
No
Germany
Yes
No
Yes
Yes
EU register
Yes (AI Act)
Yes
No
Brazil
Yes
No
Yes
Yes (TSE)
No
Yes (ban and label)
Consent-based
No
India
No list
No
Yes
ECI rules
No
ECI advisories
Consent-based
No
Japan
Creed
No
Yes
Election law
No
No
Opt-out
No
South Korea
Yes
No
Yes
Yes
No
Yes (90-day ban)
Yes
No
Australia
Yes
Yes (s.6C)
Yes
Yes (authorisation)
Yes (FITS)
No (SA/ACT truth laws)
Spam Act consent
No
Cross-cutting analyses
Consent standards, compared
Regime
Standard for ordinary data
Standard for sensitive data
Standard for texts and email
GDPR / UK GDPR
A lawful basis; consent must be freely given, specific, informed, unambiguous, withdrawable
Explicit consent or a narrow condition; profiling on political opinion needs the former
Programs run from seven offices move data across borders by design, and the mechanisms are settled but fragile. Transfers out of the EU and the UK rest on adequacy (the EU–U.S. Data Privacy Framework for certified U.S. importers, which survived the General Court's judgment in Latombe in September 2025; the UK extension; adequacy for Japan, Korea, Canada and others), on standard contractual clauses with a transfer impact assessment, or on binding corporate rules. Brazil's ANPD regulation (2024) adds standard clauses of its own; Saudi Arabia, Turkey and China require adequacy or safeguards and, in China's case, security assessment; India permits transfers except to blacklisted countries; Vietnam and Indonesia impose filing or localization duties on certain operators. The practical rule for institutional mandates is in-country holding with replication attestations, decided in the terrain brief rather than discovered at close.
Synthetic media, three theories
Theory
Where
What it requires
Litigation and limits
Disclosure
EU AI Act Art. 50; roughly half of U.S. states; Brazil TSE; China; Korea outside the ban window; platform policies
A label that content is AI-generated or materially altered, in the statute's wording, at the point of display
The most durable theory; a U.S. court enjoined a California statute whose disclosure format was found to burden speech, so format matters
Prohibition near elections
Korea (ninety days), Singapore (election period), Brazil (campaign period), several U.S. states (candidate deepfakes within 60–120 days)
No realistic manipulated depiction of a candidate in the window, with takedown and injunctive remedies
U.S. bans face First Amendment challenges; outside the U.S. they are enforced
Fraud and impersonation
FCC (artificial voice), FEC §30124, state consumer-fraud statutes, tort law
No deceptive impersonation of a candidate or official; consent rules for synthetic voice calls
The New Hampshire robocall forfeiture and the parallel criminal case show the theory works without new statute
Data brokers and provenance
Four U.S. states register brokers (Vermont since 2019, California since 2020 with the CPPA's Delete Act regime, Texas since 2023, Oregon since 2024); California's DROP mechanism requires registered brokers to process consumer deletion requests from August 2026 and to be audited from 2028; the FTC's 2024 orders bar the sale of sensitive location data by specific brokers and signal the theory for others; the EU and UK reach the same activity through records of processing, legitimate-interest assessments and the objection right. A provenance record (source, license, permitted purpose, registration status, refresh cadence, retention) is the document every regime asks for first, and the one most programs cannot produce.
Enforcement that changed the work
Year
Action
What it established
2018
ICO fines Facebook £500,000 (Cambridge Analytica); ICO report "Democracy disrupted?"
Harvested profile data used for political profiling is a data-protection breach; regulators will audit parties and vendors
2019
FTC order against Cambridge Analytica; Canadian and BC commissioners' joint finding against Facebook; Spain's Constitutional Court strikes LOREG Art. 58 bis; Italy's Garante fines the Rousseau platform
Deceptive harvesting is an unfair practice; platforms are accountable for third-party apps; parties cannot compile opinions without a precise legal basis; party platforms must secure and limit profiling
2020
ICO audits of UK political parties; Israel's voter-registry leak through a campaign app
Published recommendations on lawful basis, transparency and vendor contracts; registry access conditioned after a breach
2021
Supreme Court, Facebook v. Duguid
The TCPA autodialer definition is narrow; peer-to-peer texting is outside it, prerecorded and artificial voice are not
2023–2024
FTC orders against X-Mode/Outlogic and InMarket; FTC v. Kochava proceeds
Sale of sensitive location data can be an unfair practice; deletion and minimization by order
2024
FCC declaratory ruling on AI voices; $6 million forfeiture over the New Hampshire primary robocall; Australia's OAIC settles with Meta for A$50 million; Brazil's TSE applies its AI rules in municipal elections; Korea prosecutes deepfake campaign content
Synthetic voice needs consent like any artificial voice; the Cambridge Analytica exposure is priced; election authorities will remove and sanction synthetic content
2025
Eleventh Circuit vacates the FCC's one-to-one consent rule; FIRS commences in the UK; the General Court upholds the EU–U.S. Data Privacy Framework; the CPPA's first enforcement actions on opt-out mechanics; Texas actions on driving-data brokers; Meta and Google withdraw political ads from the EU ahead of Regulation 2024/900
Consent rules will be litigated at the margin; foreign-influence registration is live in the UK; transfers rest on the DPF; opt-out flows are audited; platform withdrawal is a compliance outcome to plan around
2026
California DROP deletion processing from August; EU AI Act Art. 50 and high-risk duties from August; Chile's law in force in December
Broker deletion at scale; labeling of synthetic content as a legal duty in the EU; another GDPR-modeled regime in Latin America
The day-ten compliance review
What the firm's compliance lead and counsel of record sign before any program goes on air, and what any client can ask its vendor to produce.
Item
Question answered
Evidence
Acceptance screen
Who is the client, where do the funds come from, what is the purpose, which jurisdictions does the work touch, is any party sanctioned
Screening record; declination letter if any
Registration
Does FARA, the LDA, FIRS, FITS, a third-party regime or a transparency register apply, and was the filing made before activity
Filing receipts with dates preceding the first activity
Practice wall
Is this candidate or independent-expenditure work, and is the firewall documented for the cycle
Wall certification; team and system assignments
Provenance
For every data source: license, permitted purpose, broker registration, refresh, retention; no sensitive categories; no minors; voter-file terms honored
Provenance record per source
Legal basis and consent
For each jurisdiction: the basis for processing; consent held for texts, email and calls; objection and opt-out honored as suppression
Consent logs; suppression list; DPIA where required
Models
No model reconstructs an excluded attribute; model cards on file; the program predicts responsiveness, not vote
Proxy-reconstruction test results; model cards
Disclaimers and imprints
Exact language per jurisdiction and channel on every master; sponsor verification on every seat
No synthetic depictions of real people; AI-assisted content labeled where required
Studio attestation; labeling log
Calls and texts
Consent basis by channel; identification scripts; hours by time zone; no artificial voice
Platform configuration; scripts
Transfers and residency
Mechanism for each cross-border flow; in-country holding where required
Transfer assessment; residency attestation
Retention and destruction
Program data destroyed within ninety days of close with certification; holdout assignment retained per the client's instruction
Destruction certificate at close
Sign-off
Compliance lead and counsel of record sign; delivered to the client
Signed review
Questions to bring to counsel
Which of the jurisdictions the program touches treat a persuadability score as political-opinion data, and what is our basis there?
Which registrations must be filed before the first activity, and who is the local partner of record in each market?
Are any of our data licensors required to be registered as brokers where they operate, and are they?
What consent do we hold for texts and email in each jurisdiction, who holds it, and how is revocation propagated across channels within ten business days?
Which state or national synthetic-media rules apply to our creative, what wording do they require, and which are currently enjoined?
What is the transfer mechanism for each cross-border flow, and does any mandate require in-country holding?
Which platform policies, as of this week, would reject or restrict the placements in the media plan?
What does the client's own counsel need from us to certify the wall, the disclaimers and the media trail?
Download and citation
The packs are published as a single PDF, updated with the site: Brushworks compliance packs, second edition, September 2026 (PDF). Cite as: Brushworks (2026), The compliance packs: what data and political firms must obey, country by country, second edition, September 2026. Corrections and updates to research@brwrks.com; we publish them with a dated note.
Second edition, reviewed September 2026. These summaries are general information, not legal advice, and do not create a professional relationship; thresholds, dates and case outcomes should be verified against primary sources before any decision, and counsel consulted in each jurisdiction.
The commitments we are bound by, in plain language.
Written for the people whose behavior we model, the clients we work for, the regulators who oversee both, the journalists who will ask, and anyone who reads this site. Effective September 18, 2026. Each notice below is a commitment the firm can be held to, and How to verify says where each one can be checked.
A policy is a claim. This ledger says where each claim can be checked, by a client's counsel, a committee treasurer, a ministry's compliance office, a regulator, or a reporter.
Claim
Where to check it
Who can
We register before foreign-principal work touches the United States
The Department of Justice's public FARA registry and the House and Senate Lobbying Disclosure databases
Anyone
Every paid political communication carries its disclaimer and names its sponsor
The public ad libraries of the major platforms, under the sponsor's verified name; the client's own filings
Anyone
Candidate and independent-expenditure work are walled
The firm's written wall certification, delivered on request to any committee, counsel or regulator
Clients, committees, counsel
Every program is measured against a sealed control
The client's holdout report: the control is reported to the client directly and never enters the delivery path
Clients and their auditors
No model uses ethnicity, religion or ideology
The registry exclusion and the proxy-reconstruction test, documented in the program's model cards and kill log
Clients, auditors; summarized annually in public
Every data source has a documented lawful basis
The provenance record for the client's program, delivered with the engagement
Clients, auditors, regulators
The performance assurance is underwritten
The carrier's certificate, provided to qualifying clients under the engagement letter
Qualifying clients
Security controls are independently assessed
The independent assessment summary, under non-disclosure agreement
Clients and prospective clients under NDA
Concerns are answered and nobody is retaliated against
The semiannual transparency report: concerns received, resolved and outstanding, in aggregate
Anyone
Press and public inquiries: press@brwrks.com, answered within one business day. We do not disclose client identities held under agreement, and we do not decline to confirm a registration, a disclaimer or a policy.
Reports
What the firm publishes, on what cadence, and what is available under agreement.
The record for the cycle: contests, methods, aggregate lift against holdouts, losses and post-mortems, the kill-log summary, the governance ledger and the independent-review summaries.
Registrations filed, disclaimers carried, concerns received and resolved, mandates declined, incidents, rights requests received, and the practice-wall certification.
For prospective clients under non-disclosure agreement: wall certification, a redacted sample proof report and kill log, the processor list, a sample provenance record, a sample model card, the independent-review summaries, and the carrier's certificate for the performance assurance.
What we collect, from whom and on what basis, how long we keep it, and how to use your rights. Effective September 18, 2026.
A person as a set of recorded acts · illustrative
Applies to this website, to the people whose data appears in our programs, and to clients, counterparties and correspondents.
In short. We use behavioral data, what people do, to design and measure lawful political and institutional communication. We never model who people are; we never sell information; we never run undisclosed accounts; we never keep a client's data after the engagement; and we never build a profile of anyone from this website. You can ask what we hold about you, have it corrected or deleted, and opt out. The addresses are at the end.
1. Who we are and what this policy covers
Brushworks Strategies Inc. ("Brushworks," "we") is a Wyoming corporation providing political strategy, behavioral research, creative and delivery services to campaigns, committees, institutions and governments, with offices in Washington, New York, Ottawa, London, Munich, Dubai and Tokyo. This policy covers three groups: visitors to this site; people whose information is in the data we license, receive or produce for a program; and the clients, vendors and correspondents we deal with. Brushworks is the controller of the personal information described here except where it processes a client's data on the client's instructions under a written data-use agreement, in which case it acts as a processor and this policy applies alongside that agreement.
2. Definitions
Behavioral data: recorded acts with a time and a source: a subscription, a search, a renewal, a visit, a donation, a returned ballot. Opinions, attributes and categories are excluded.
Signal: a defined, versioned pattern over behavioral data that has been shown to predict a later act. The registry of active signals is curated and pruned monthly.
Cohort: a group of people defined by what they attend to and act on. Cohorts are composites; no cohort is a person, and no program contacts a cohort of one.
Program: a single engagement for a single client with a single purpose, from terrain to proof report.
Holdout: a share of every cohort drawn at random before any spend and never contacted, against which the program is measured.
Hashed identifier: a one-way transformation of an identifier such as an email address, used to match an audience inside a delivery platform without transmitting the identifier itself.
Client data: data a client lawfully holds and provides to us under a data-use agreement. Licensed data: data licensed from a provider under written terms that document its lawful basis and source.
3. Visitors to this site
We collect the minimum needed to operate the site: server logs (IP address, browser and device type, pages requested, timestamps), the information you choose to send when you brief us, and, only if you consent, the analytics described in the cookie notice. We do not sell visitor information, we do not use this site to build behavioral profiles, we do not match visitors to the behavioral graph, and we do not use advertising cookies or tracking pixels from any third party.
4. The people whose behavior we model: sources
Programs use data of four kinds. Licensed behavioral data from established providers, under written agreements that document the lawful basis on which the data was collected and may be used for political communication and that warrant its provenance. Client data a client lawfully holds (a supporter file, a member list, a donor record) provided under a data-use agreement and used for that client's program alone. Public records, including voter registration and participation files, obtained and used only as the law of the issuing state permits, for the political purposes the state allows. Our own measurement of the programs we run: the holdout comparison, response records and sentiment instruments. Information is resolved to individuals and households for research, audience design, delivery and measurement, and for nothing else.
5. What we never do
We do not model, license, infer, purchase or accept as a targeting filter any data describing ethnicity, race, religion, or political ideology, even where the law would permit it. The exclusion is enforced where signals are defined, and every model is tested for reconstructing an excluded attribute by proxy; models that do are retired and logged.
We do not process the sensitive categories data-protection law protects: health status or records, genetic or biometric data, sexual orientation, immigration status, criminal history, precise geolocation, or financial account data. Attention to a topic in the media, a person who reads about hospital staffing, is a media-behavior signal, not a health record, and is treated as such.
We do not sell personal information, and we do not share it for cross-context behavioral advertising, in any jurisdiction, under any definition.
We do not operate undisclosed personas or accounts on any platform, in any language, for any client.
We do not combine one client's data with another's, and we do not retain a client's data after the engagement and its wind-down period.
We do not knowingly collect or model data about anyone under eighteen. Voter files by construction contain only adults; licensed sources are contractually restricted to adults; a record found to belong to a minor is deleted on discovery.
We do not scrape websites or social platforms, and we do not license from brokers who cannot document how their data was collected.
We do not produce synthetic depictions of real people (voice, likeness or fabricated statements) for any purpose.
6. Lawful basis, by jurisdiction
In the United States, our use of behavioral data for political communication rests on the licenses and agreements above and on the state laws governing voter files and consumer data; the California Consumer Privacy Act as amended by the California Privacy Rights Act, and comparable statutes in other states, govern the rights described in section 8, and we treat those rights as available to everyone regardless of residence. In the European Union, the United Kingdom and Germany, processing rests on legitimate interests assessed and documented for each program, with consent where the law requires it and with the safeguards the GDPR, the UK GDPR and the Bundesdatenschutzgesetz provide for processing that may reveal political opinions, which our behavior-only design is built to avoid. In Canada, on PIPEDA and Quebec's Law 25. In the United Arab Emirates, on the Personal Data Protection Law. In Japan, on the Act on the Protection of Personal Information. Where a program touches more than one of these, the stricter rule governs.
7. Sharing
Audiences leave the platform only as hashed identifiers, matched inside the delivery platforms that carry a program; raw personal data does not travel. We share information with processors (hosting and infrastructure, security and monitoring, delivery platforms, measurement) bound by written terms at least as protective as this policy; with clients as their data-use agreement provides and only for their program; and with regulators, courts or law enforcement where the law requires it, in which case we notify the affected client unless prohibited. We do not share personal information with data brokers, with other clients, or with any political party, committee or campaign other than the client for whom a program is run.
8. Your rights, and how to use them
Wherever you live, you may ask us what personal information we hold about you and receive a copy; have it corrected or deleted; opt out of any use for targeted advertising; object to or restrict processing; limit the use of any sensitive information (we hold none); and, where the law provides it, receive a portable copy and appeal a decision we make about your request. Write to privacy@brwrks.com with enough information for us to find your record and confirm it is yours; we may ask for verification proportionate to the sensitivity of the request, and we accept requests from an authorized agent with written authority. We answer within thirty days, or within forty-five where the applicable law allows it and the request is complex, and we tell you if we need longer. We do not discriminate against anyone for exercising a right. A Global Privacy Control signal from your browser is treated as an opt-out request where the law so provides. If you are unsatisfied, section 13 lists the regulators you may contact.
Where you live
Rights we honor
Our response time
California and other U.S. states with consumer privacy statutes
Know, access, correct, delete, opt out of sale or sharing and of targeted advertising, limit sensitive use, non-discrimination, appeal
45 days, extendable once by 45 with notice
Other U.S. states
The same rights, extended voluntarily
45 days
European Union and United Kingdom
Access, rectification, erasure, restriction, portability, objection, and rights in relation to profiling under Articles 15–22
One month, extendable by two for complex requests
Canada
Access, correction, withdrawal of consent, and complaint to the Commissioner
30 days
United Arab Emirates and Japan
Access, correction, deletion, restriction and objection as the PDPL and APPI provide
As the law requires; never longer than one month
9. Retention
Client program data is retained for the life of the engagement and a wind-down period of not more than ninety days, then destroyed with certification. Licensed data is refreshed and superseded on its providers' cycles; the behavioral graph is rebuilt in full each month, and superseded data is not archived. The schedule in section 12 governs everything else.
10. Security
Access to program data is limited to personnel working on the relevant program under written obligations, with named-user access, multi-factor authentication, encryption in transit and at rest, logged access to client files reviewed monthly, and separate systems and seats for the candidate and independent-expenditure practices. Every employee and contractor completes data-protection and political-compliance training before access and annually thereafter, and loses access on the day their role ends. Security controls are independently assessed each year and the assessment summary is available under non-disclosure agreement. Where a mandate requires it, program data is held in-country with residency and replication attestations delivered to the client.
Control area
What is in place
Evidence available
Identity and access
Named-user accounts, multi-factor authentication, least-privilege roles scoped to a program, quarterly access recertification, same-day revocation on role change
Access matrix; recertification records
Encryption
TLS 1.2 or higher in transit; AES-256 at rest; keys managed in a dedicated key service with rotation and separation of duties
Configuration attestation
Segregation
Separate environments per practice; program data isolated by client; no shared credentials; delivery seats separated by practice
Wall certification; environment diagram
Logging and monitoring
Immutable access and change logs retained one year, reviewed monthly; alerting on anomalous access; export and bulk-read events flagged
Log samples; review records
Vulnerability management
Continuous dependency scanning; patching within defined windows by severity; annual independent penetration test with remediation tracking
Test summary under NDA
Secure development
Code review for every change; secrets kept out of repositories; production changes logged and reversible
Change records
Business continuity
Encrypted backups with restoration tested quarterly; documented recovery objectives per system; the desk's seven-office handoff for operational continuity
Restoration test records
Incident response
Written runbook; on-call across the desk; containment, notification and post-incident review under counsel
Runbook summary; incident log in aggregate
Vendors
Due diligence before onboarding; written terms; annual review; a maintained processor list
Processor list; diligence records
People
Background screening where lawful; training before access and annually; written obligations that survive employment
Training records
11. Automated decisions, profiling and international transfers
Our models score the likelihood that a person will act; they do not make decisions with legal or similarly significant effects about anyone, and no score is ever used to deny anyone a service, a benefit, an opportunity or the right to vote. Where the GDPR or UK GDPR applies, this is profiling for a legitimate interest and not automated decision-making within Article 22; a data-protection impact assessment is completed for every program in those jurisdictions and is available to the supervisory authority on request. We operate from seven offices in six countries. Personal information may be processed in the United States and, for programs abroad, in the country of the mandate. Transfers out of the European Economic Area or the United Kingdom rely on standard contractual clauses or an adequacy decision; transfers out of other jurisdictions rely on the mechanisms their law provides.
12. Retention schedule
Data
Kept for
Then
Client program data
Life of the engagement, plus a wind-down of not more than ninety days
Destroyed, with certification on request
Hashed audiences in delivery platforms
Life of the program
Deleted from the seats at close
Licensed behavioral data
Refreshed and superseded on providers' cycles; the graph is rebuilt monthly
Superseded; not archived
Models trained on a client's outcomes
Life of the program
Retired with the program; recorded in the kill log
Measurement and proof reports
Delivered to the client; the client's property
Firm copy retained as an engagement record for seven years
Compliance records: registrations, wall certifications, reviews, conflict and gift registers
Seven years, as the law and our auditors require
Destroyed
Site logs
Ninety days
Deleted
Correspondence about a brief
Two years, unless an engagement follows
Deleted
Rights requests and concerns
Three years, as a compliance record
Deleted
13. Processors, regulators and contact
We use processors in four categories (hosting and infrastructure, security and monitoring, the delivery platforms that carry a program, and measurement) each bound by written terms at least as protective as this policy; a current list is available on request and is delivered to every client. If you are not satisfied with our answer, you may contact your regulator: in the United States your state attorney general or, in California, the California Privacy Protection Agency; in the European Union your national data protection authority; in Germany the federal or state commissioner with jurisdiction; in the United Kingdom the Information Commissioner's Office; in Canada the Office of the Privacy Commissioner or, in Quebec, the Commission d'accès à l'information; in the United Arab Emirates the UAE Data Office; in Japan the Personal Information Protection Commission. We post revisions here with a new effective date and, for material changes affecting people in our programs, notify affected clients. Questions, requests and complaints: privacy@brwrks.com, or by post to Brushworks Strategies Inc., attention Data Protection Lead, at our registered agent's address in Wyoming.
Version 1.2, September 18, 2026. Versions 1.0 and 1.1 published September 2026. Prior versions available on request.
The sources we accept and refuse, how data is classified and handled, and the uses we prohibit. Effective September 18, 2026.
The registry by family · illustrative
Governs every dataset Brushworks licenses, receives or produces, and every model built from it. Incorporated by reference into every client, vendor and employment agreement.
In short. One program, one purpose, one client; behavior only; hashed on the way out; destroyed with certification at close. Every source has a documented lawful basis or it is not used. Every model faces the same gate, every failure is logged, and every result is measured against a control drawn before any money was spent. A request to do otherwise ends the engagement.
1. Principles
Purpose limitation. Data is used for the program it was licensed or provided for, and for no other client, program or purpose. A client's file is matched to the graph for that client's program alone and unmatched at close.
Behavior only. We do not model or infer ethnicity, race, religion or political ideology, and we do not accept such attributes from clients or vendors as targeting criteria, even where they are lawful to use. We do not process the sensitive categories listed in the privacy policy.
No commingling. One client's data is never combined with another's. Models trained on a client's outcomes belong to that program and are retired with it. Learning that improves the firm's methods is carried forward only in aggregate and never as data.
Provenance. We license only from providers who document the lawful basis and source of their data and warrant it in writing. We do not scrape, and we do not buy from brokers who cannot answer the question.
Minimization. Programs use the signals the models need and nothing else; the registry is pruned monthly of signals that have stopped predicting.
Hashed activation. Audiences are transmitted to delivery platforms as hashed identifiers matched inside those platforms. Raw personal data is never uploaded to a platform.
Retention. Program data is destroyed on a written schedule after close, with certification to the client on request and as standard for institutional mandates.
Onward transfer. No data is transferred to a third party except delivery platforms and processors bound by written terms at least as protective as this policy.
Human review. Every creative master is reviewed and approved by a named person before any variant is produced from it; no communication is published without a person having read it.
2. Sources we accept, and sources we refuse
Accepted, with a provenance record
Refused, without exception
Licensed behavioral data from providers who warrant lawful collection and permitted political use
Data obtained by scraping websites or social platforms
Client files the client lawfully holds and may use for political communication
Hacked, leaked, breached or otherwise unlawfully obtained data, however it is offered
Voter registration and participation files, within the issuing state's permitted uses
Health, genetic, biometric or financial-account records; precise geolocation; immigration or criminal-history data
Our own program measurement: holdout comparisons, responses, sentiment instruments
Any dataset describing ethnicity, race, religion or ideology, or built to infer them
Public statistical and administrative data at the aggregate level
Data about minors, and data from sources that cannot exclude them
Aggregate, non-identifying platform reporting
Facial recognition or other biometric identification outputs
Program-scoped access; encrypted; never commingled; destroyed at close with certification
Licensed behavioral
Events, signals, the graph
Provider terms enforced; refreshed and superseded; not exported; used only for the program a license permits
Derived
Scores, cohorts, models, holdout assignments
Labeled with confidence and attribution level; models retired with the program; holdouts sealed and reported to the client only
Activation
Hashed audiences in delivery platforms
Hashed before transmission; matched in-platform; deleted from seats at close
Public
This site, the Review, published reports
No personal data; corrections dated
4. Prohibited uses
No program run by Brushworks may use data to suppress turnout or discourage any eligible person from voting; to deceive voters about the time, place, manner or eligibility rules of voting; to impersonate a person, organization or official source; to operate undisclosed accounts or personas; to target on identity; to re-identify a hashed audience or a holdout; to harass or intimidate; to produce synthetic depictions of real people; or to resell or repurpose data licensed or provided for a program. A request to do any of these ends the engagement, and the request is recorded.
5. Model governance
Every model faces the publish gate: on the client's own sealed holdout it must beat a demographic baseline, the voter file used as well as a good consultant would use it, and a look-alike baseline, the delivery platforms' own targeting, before it is used. Candidate models are tested for reconstructing an excluded attribute by proxy and retired if they do. Models that fail any test are logged with the reason and killed; the kill log is delivered to the client with the proof report and summarized in aggregate each year. Every shipped model carries a model card recording its training data, baselines, discrimination, known limits and kill threshold, and is watched in field daily against that threshold. Every score is labeled with its confidence and with the attribution level, person or household, of its weakest input.
6. Measurement and reporting
Every program carries a randomized holdout drawn before any spend and sealed; the holdout reports to the client directly and never enters the delivery path. Results are reported as the difference against the control with confidence intervals, never as raw outcomes alone, and null results are reported as prominently as positive ones. Weekly field reports, the proof report and the kill log are the client's property. The methodology (holdout size, estimand, interval construction, baselines) is published with the Review.
7. Provenance standard
Before a dataset is licensed or a client file is accepted, the firm records, for that source: the lawful basis on which it was collected and may be used for political communication; the notice given to the people in it; the provider's written warranties and audit rights; the refresh cycle; whether records are person- or household-attributed; the categories present, confirming that the excluded and sensitive categories are absent; and the jurisdictions in which it may be used. A source that cannot answer every question is not used. The record is available to the client for its program and to a regulator on request.
8. People and vendors
Every employee and contractor signs this policy, completes training before access and annually, may not use program data for any personal purpose including personal political activity, and loses access on the day their role ends. Every vendor that touches program data signs terms at least as protective as this policy, accepts audit, and is listed on the processor list delivered to clients. A breach of this policy by an employee is a disciplinary matter up to dismissal; by a vendor, grounds for termination.
9. Incidents
We notify affected clients within seventy-two hours of confirming a security incident involving their data, and regulators and individuals as the applicable law requires, with a written account of what happened, what was affected, what has been done and what will change. Incidents are recorded, reviewed under counsel, and reported in aggregate in the transparency report.
10. Audit and retention
Clients may audit our compliance with this policy on reasonable notice under their agreement, including the single-origin media trail from invoice to placement to impression and the provenance records for their program. Independent security reporting is available under agreement. The retention schedule in the privacy policy applies to every dataset covered here.
11. Use of AI tools
AI tools assist the studio in producing variants from an approved master and the research practice in modeling; they do not originate claims, do not depict real people, and do not decide anything about a person. Every AI-assisted variant traces to a source the campaign supplied and is read by a person before it exists as a creative; every model is gated and logged as section 5 describes. AI-generated or materially altered content is disclosed wherever the platform or the law requires it. No client or program data is used to train any model that serves another client, and no program data is submitted to a third-party AI service that retains it.
How the firm is structured and checked: practice walls, reviews, sign-off gates, conflicts and independent review. Effective September 18, 2026.
The daily loop, reviewed · illustrative
How the firm is run, who is accountable for what, what we will not take on, and where to take a concern.
In short. Separate practices behind a documented wall; a signed compliance review before anything goes on air; written client sign-off gates with a ten-percent reallocation threshold; no opposing sides of the same contest; clients of both parties and none; a written list of work we decline; a concerns channel that answers in writing and never retaliates; and an annual independent review.
1. Structure and accountability
Brushworks Strategies Inc. is a Wyoming corporation whose whole business is political and institutional work. It is organized as a candidate practice, an independent-expenditure practice and an international practice, served by research and modeling, a creative studio and a seven-office desk. Political work in the United States is contracted through Brushworks and no other entity. The chief executive is accountable to the board for every policy on this page. A compliance lead owns the reviews, registrations and registers; a data protection lead owns the data use policy, the retention schedule and rights requests; the head of research owns model governance and measurement; counsel of record is named in every engagement letter. Each reports on the policies to the board at least quarterly.
Role
Accountable for
Reports to
Cadence
Chief executive
Every policy on this page; client acceptance; the conflict register
The board
Quarterly, and after any incident
Compliance lead
Compliance reviews, registrations, disclaimers, the gift register, training
Chief executive and counsel of record
Every program at day ten; quarterly summary
Data protection lead
Data use policy, retention, rights requests, impact assessments, processor list
Chief executive; regulators on request
Monthly review; annual report
Head of research
Publish gate, model cards, kill log, measurement methodology
Chief executive; clients through the proof report
Every program; annual kill-log summary
Head of the desk
Seven-office handoff, crisis architecture, business continuity
Chief executive
Quarterly rehearsal
Counsel of record
Named in every engagement letter; signs the compliance review; supervises investigations
The client and the board
Every program
2. Practice walls
The candidate practice and the independent-expenditure practice are separate teams with separate systems, separate media seats and separate access to data. No person works in both during a cycle; no data, creative or plan crosses between them; shared functions (research, the studio, the desk) serve each practice through separate, logged channels under the firewall policy. The wall is documented, is reviewed by counsel each cycle, and a certification of it is available to any client, committee, counsel or regulator that requires one. The international practice is separately walled from domestic political work where a mandate touches the United States.
3. Compliance review
Every program passes a compliance review on the tenth day of the engagement and before any material change: disclaimers and disclosures, coordination protocols, registration under the laws governing foreign representation and lobbying, sanctions and end-use screening, platform political-advertising policies and verification, synthetic-media and AI-disclosure requirements, and brand safety. The review is signed by the compliance lead and counsel of record. A program does not go on air without a signed review, and the review is delivered to the client.
4. Client sign-off gates
Strategy, creative masters and the media plan are approved in writing by the client (or, for donor-funded committees, by the contributors' advisory board) before execution. Reallocations of more than ten percent of any budget line require the same approval. In-flight optimization within approved lines is Brushworks' responsibility and is reported weekly. Clients hold audit rights over the single-origin media trail and receive the invoice-to-impression chain in the proof report.
5. Client acceptance, neutrality and what we decline
We work for candidates and causes across the political spectrum and take no position as a firm on partisan questions; the record spans both major parties at every level. Every prospective engagement is screened before acceptance for the identity of the client and the source of its funds, its purpose, the jurisdiction, sanctions exposure, and conflicts. We decline, and record the declination of: work on opposing sides of the same contest or question; mandates whose client, funds or purpose cannot be described in public; work for parties subject to sanctions; work for organizations that advocate violence or the denial of civil rights to any group; programs whose object is to suppress or discourage lawful voting; foreign-principal work intended to influence a United States election; and any program that would require undisclosed messengers, synthetic depictions of real people, or targeting on identity. Mandates declined are reported in aggregate in the transparency report.
6. Conflicts, gifts and personal political activity
A conflict register is reviewed before every engagement and quarterly thereafter. A gift and hospitality register records anything offered to or by a public official. Employees may engage in personal political activity on their own time and at their own expense, may not use the firm's data, systems or name in doing so, and disclose personal political contributions and any campaign roles to the compliance lead so that conflicts can be identified before they arise.
7. Data governance
The data use policy above is binding on every employee, contractor and vendor. The data protection lead is responsible for its application, for the retention schedule, for data-protection impact assessments where required, and for responding to rights requests. Model governance (the publish gate, model cards and the kill log) is owned by the head of research and reported to clients with every proof report.
8. Independent review
Each year an independent assessor reviews the firm's security controls and the operation of the practice wall, and an independent reviewer with the relevant expertise reviews the measurement methodology (holdout construction, estimands, interval reporting) against the standard published with the Review. Summaries of both are available under non-disclosure agreement to clients and prospective clients and are referenced in the annual report. This policy is reviewed annually by counsel and after any incident, and revisions are posted with a new effective date.
Review
By whom
How often
Output
Security controls
Independent assessor
Annual
Assessment summary under NDA
Practice wall
Independent assessor and counsel
Annual and each cycle
Wall certification
Measurement methodology
Independent reviewer with domain expertise
Annual
Methodology note published with the Review
Policies on this page
Counsel
Annual and after any incident
Revision with new effective date
Training
Compliance lead
Before access; annually; on material change
Training records
9. Reporting to clients and the public
Every client receives the weekly field report and, at close, the proof report, the kill log and the compliance review. The firm publishes an annual report, a semiannual transparency report and an annual kill-log summary, listed under Reports below.
10. Raising a concern
Anyone (an employee, a client, a vendor, a member of the public) can raise a concern about our conduct at concerns@brwrks.com. Concerns may be raised anonymously. Reports are acknowledged within two business days, investigated under the supervision of counsel, and answered in writing. Retaliation against anyone who raises a concern in good faith is prohibited and is itself a disciplinary matter; a concern about a member of leadership is investigated by counsel alone and reported to the board. Concerns received, resolved and outstanding are reported in aggregate in the transparency report.
11. Continuity and incident command
A program in the field is watched continuously by the office whose business day it is, with a written handoff between offices and a rehearsed crisis clock. Systems carry encrypted backups restored on a tested schedule, and every program's data can be reconstituted from the previous day's state. An incident (security, compliance, reputational or operational) is commanded under a written runbook by a named lead, with counsel engaged from the first hour, the client notified as the data use policy requires, and a post-incident review completed within thirty days and reported to the board.
Federal and state political work, independent expenditures, foreign principals, lobbying, sanctions and synthetic media. Effective September 18, 2026.
Every placement, disclosed · illustrative
What the law requires of us in the United States and abroad, what we do beyond it, and where the public can see it.
In short. Every communication carries its disclaimer and names its sponsor, and every paid placement on the major platforms is visible in their public ad libraries. Independent expenditures do not coordinate. Foreign mandates that touch the United States are registered before work begins, and no foreign money reaches United States political work. AI-assisted content is disclosed wherever the law or a platform requires it. Counterparties are screened. The performance assurance is underwritten and applies only to media we place ourselves.
1. Federal and state political work
Communications we produce for candidates, committees and ballot campaigns carry the disclaimers required by federal law (52 U.S.C. §30120 and 11 C.F.R. §110.11) and by the states and localities in which they run, including top-funder and committee-name requirements where they apply. We support clients' reporting obligations, including 24- and 48-hour independent-expenditure reports, with the placement records the single-origin media trail produces, and our vendor payments appear in clients' public filings. We do not make contributions to candidates or committees, and we do not solicit them.
Regime
When it applies to us
What we do
Where the public can see it
Federal disclaimers and coordination (FECA; 11 C.F.R. Parts 109, 110)
Every federal candidate, committee and independent-expenditure communication
Disclaimers on every communication; firewall policy; no common vendor across the wall
The communication itself; clients' FEC reports
State and local campaign finance
State, local and ballot-measure work
State disclaimers, top-funder and committee-name language; vendor reporting through clients
State and local filings
Foreign Agents Registration Act
Work for a foreign principal that involves activity in the United States
Registration before work begins; labeling of informational materials; semiannual reports
The Department of Justice's FARA registry
Lobbying Disclosure Act; state lobbying laws
Where our activity constitutes lobbying
Registration and periodic reports
House and Senate LDA databases; state registries
Foreign-influence and third-party regimes abroad
Political or institutional work in jurisdictions with such regimes
Registration and reporting with local partners of record
The relevant national registry
Platform political-advertising policies
Every paid political placement on the major platforms
Advertiser verification in the sponsor's name; disclaimers as the platform requires
The platforms' public ad libraries
Sanctions and export controls
Every counterparty, before and during an engagement
Screening of parties, owners and funds; a match ends the engagement
Certification available to counsel
2. Independent expenditures
Independent-expenditure work is performed by a separate practice that does not coordinate with any candidate, campaign or party committee. We do not act as a common vendor across the wall within the periods the regulations define, do not republish campaign materials, and maintain a documented firewall that satisfies 11 C.F.R. §109.21(h). Where a client requires it we provide a certification of the firm's posture with respect to a given contest, and we will provide the same certification to the Commission or a court.
3. Foreign governments and foreign principals
Where a mandate for a foreign government, political party or other foreign principal involves activity in the United States, Brushworks registers under the Foreign Agents Registration Act (22 U.S.C. §611 et seq.), or the Lobbying Disclosure Act where that statute applies, before work begins, and files the informational materials and periodic reports the law requires; those filings are public. We do not accept foreign funds for United States political work of any kind, and we do not accept a mandate whose object is to influence a United States election. Work abroad that does not touch the United States is conducted under the laws of the relevant jurisdiction, including third-party and foreign-influence registration regimes where they exist, and with local partners of record, and every placement carries its sponsor in the language of the market.
4. State lobbying and gifts
Where our work constitutes lobbying under a state's or a foreign jurisdiction's law, we register and report there. We do not offer gifts, hospitality or anything of value to public officials beyond what the applicable rules permit, and we keep a register of what is offered.
5. Sanctions and end use
Counterparties, their beneficial owners and the sources of their funds are screened against United States, United Kingdom, European Union and United Nations sanctions lists and against export and end-use restrictions before an engagement and periodically during it. A match ends the engagement.
6. Platforms, AI and synthetic media
Paid political placements on the major platforms run from accounts verified in the sponsor's name, so that they appear in the platforms' public ad libraries with their disclaimers and spend. Where content is generated or materially altered with the assistance of AI tools, we disclose it as the platform and the growing number of states that require disclosure of such content in political communication demand, and we do not produce synthetic depictions of real people (voice, likeness or fabricated statements) under any circumstances. Variants produced by the studio carry their disclosure as part of the master, so that it cannot be dropped in production.
7. What a disclaimer looks like
Every paid communication states who paid for it and, where required, whether it was authorized by a candidate; every organic and ambient placement we arrange identifies its sponsor; validators say who they are and who asked them. Where a jurisdiction requires additional language (a top-funder disclosure, a ballot-measure committee's name, an "independent expenditure" statement, an AI-content notice) the compliance review confirms it before the program goes on air.
8. Performance assurance
Where Brushworks offers a performance assurance, it is underwritten by a third-party carrier, applies only to political programs and only to media placed through Brushworks' own seats, and is described in full in the engagement letter, with the carrier's certificate provided to the client. It is not offered on institutional or commercial work. Nothing on this site is an offer of the assurance to any particular client.
What this site keeps in your browser, and what it does not. Effective September 18, 2026.
What this site keeps, and what it does not · illustrative
In short. No advertising cookies, no cross-site tracking, no third-party pixels, and nothing set without your consent except what the site needs to remember a display preference on your own device.
This site sets only what it needs to work. Analytics, if enabled, run only after you consent, truncate IP addresses, and are configured not to share data with any third party for advertising. We honor Global Privacy Control. You may withdraw consent at any time in your browser settings; every part of the site continues to work without cookies.
What is set
Purpose
Lasts
Display preference
Remembers a theme or tab choice on your device; never sent to us
Until you clear it
Privacy choices
Records your Do Not Sell or Share opt-out and your analytics choice on this device, from the "Your privacy choices" panel in the footer; never sent to us
Until you clear it
Analytics, if allowed
Aggregate page-level usage with truncated IP; no advertising use
The terms on which this site may be used. Effective September 18, 2026.
Counted, not claimed · illustrative
In short. The site is information, not an offer; engagements exist only in writing. Quote with attribution; do not republish exhibits or papers in full without permission. Figures marked illustrative are the method, not a client's result.
By using this site you agree to these terms. The material on the site is published for information; it is not an offer of services, not legal or financial advice, and not a guarantee of any result. Engagements are governed exclusively by written agreements. Case descriptions are accurate to the firm's records and are anonymized under client agreements; official election results are stated from the public record; figures marked illustrative or pooled depict the method rather than a specific program and should not be read as a specific client's result. Content, exhibits and the Brushworks name and marks are the property of Brushworks Strategies Inc.; you may quote from the site with attribution and may not reproduce exhibits or papers in full without permission. The Brushworks Review is published for general readers and practitioners; papers state their sources and their own limits, and corrections are published with the paper and dated. These terms are governed by the laws of the State of Wyoming.
Acceptable use
You may read, link to and quote from the site with attribution. You may not scrape it, reproduce it in substantial part, present its exhibits as your own, use its contents to train a model without permission, misrepresent the firm's positions, or use the site to attempt access to any system behind it.
No advice; no reliance
Nothing on the site is legal, financial, electoral or compliance advice, and nothing on it describes what a program would do for you; that is stated only in a written engagement, after a terrain brief, with the measurement design and the assurance terms set out in full. Case descriptions are accurate to the firm's records and anonymized under agreement. The Review's papers are the firm's own analysis and are corrected, with a dated note, when a reader shows us an error.
Liability
The site is provided as it is. To the extent the law allows, Brushworks accepts no liability for reliance on it, for interruptions to it, or for third-party sites it links to; nothing in these terms limits liability that cannot be limited by law.
Changes
We revise these terms by posting a new version with a new effective date. Material changes are noted in the version log. Disputes are subject to the exclusive jurisdiction of the courts of the State of Wyoming, and these terms are governed by its laws.
The standard this site is built to, how it is tested, and how to tell us where it falls short. Effective September 18, 2026.
Readable at every size · illustrative
This site is built to WCAG 2.2 AA: it is keyboard navigable, labeled for screen readers, and offers reduced-motion and dark-mode variants. Every animated exhibit has a text equivalent in its caption and label; charts are described in words; color is never the only carrier of meaning. Known limitations: some data exhibits are dense at the smallest screen sizes and can be scrolled sideways; live clocks read as text. If anything on the site is not accessible to you, write to access@brwrks.com and we will fix it or provide the content in another form within five business days.
Conformance and testing
The site is built to the Web Content Accessibility Guidelines 2.2 at level AA and is tested before each release with automated checks and manual keyboard and screen-reader passes at desktop and phone widths, in light and dark modes, and with reduced motion enabled. Exhibits that animate have text equivalents; interactive exhibits have keyboard-operable controls and do not rely on drag alone. Where an exhibit cannot be made fully accessible, a dense data visualization at the smallest screen sizes, its caption states what it shows, and the same content is available in words on request. This statement is reviewed with every release.
Tell us the race, the calendar and the margin. Within a week we will tell you whether we can move it and what it would take. Candidate work and independent-expenditure work are taken by separate teams; say which you are. Governments and institutions: tell us the mandate and the jurisdiction, and we will tell you what registration it requires before we say anything else.
Brushworks Strategies Inc., a Wyoming corporation · Washington, New York, Ottawa, London, Munich, Dubai, Tokyo
Your privacy choices. This site sets no advertising cookies and does not sell or share personal information. You can record an opt-out anyway; it is honored on this device and, where the law provides, by a Global Privacy Control signal from your browser.
Brushworks does not sell personal information and does not share it for cross-context behavioral advertising, in any jurisdiction, under any definition. This site sets no advertising or tracking cookies. The choices below are recorded on this device only and are honored here; a Global Privacy Control signal from your browser is treated as an opt-out where the law provides.
Do Not Sell or Share My Personal InformationRecords an opt-out from any sale or sharing, present or future, for this device.
Limit the Use of My Sensitive Personal InformationWe collect none; the choice is recorded so that it remains in force if that ever changes.
AnalyticsOff by default. If enabled, aggregate page-level usage with truncated IP and no advertising use.
To exercise any other right (access, correction, deletion, a copy of your information, an appeal) write to privacy@brwrks.com or use the contact page. Authorized agents may submit requests with written authority. We answer within the time the applicable law requires and never discriminate against anyone for exercising a right.