Political strategy and data, at human resolution
Brushworks is a political intelligence firm. We model how individual voters behave, determine what would move each of them, and put persuasion into the field with a standard of proof — across the United States and in the harder theaters abroad.
Most campaigns still buy the electorate the way television networks sold it in 1985: by age, sex, county and party file. Those labels are easy to purchase and easy to defend in a budget meeting. They are also the reason two neighbors with identical demographics receive identical mail and vote in opposite directions.
People do not decide as categories. They decide as individuals, under pressure, with whatever they happened to pay attention to that month. That attention leaves traces — what someone reads, buys, searches, cancels, renews, drives past on the way to work. Read carefully and modeled properly, those traces describe a person far better than the file does: what they are anxious about, what they take pride in, whom they trust, and what they have quietly stopped believing.
Brushworks was built to work at that resolution. We do the research, the modeling, the message architecture and the delivery ourselves, through our own media seats, and we measure every program against a held-out control — so that what we report at the end is what actually happened.
The order matters. Nothing is built before the race is mapped; nothing is written before the deciders are named; nothing is reported that was not measured against a control.
Where the race will actually be decided: the precincts, the margin, the money coming against you, and the clock.
Deciders identified by what they do rather than by label, and grouped by what would move them.
A narrative track for each group, built from the anxieties and hopes it already holds and sequenced for the persuasion arc.
Three registers: precision paid media, organic advocacy, and the ambient campaign. All of it disclosed, all of it through our own seats.
Randomized holdouts, daily measurement, and a publish gate no model passes without beating the baseline.
Our view of the American electorate is assembled from roughly sixty-two billion observed behaviors, resolved to more than 270 million adults and their households, organized into a ten-thousand-topic dictionary, and scored by models trained on real outcomes rather than proxies.
No single signal is decisive. The pattern is. A subscription cancelled, a commute lengthened, a local news habit, a sudden interest in prescription prices — each is noise on its own and evidence together. The graph is rebuilt monthly, digital activity refreshed daily, and none of it touches ethnicity, religion or ideology, which we do not model.
Every other vendor in politics starts from the voter file, because that is what there is to buy: age band, sex, party, precinct, vote history. It is enough to send mail. It is not enough to know what to say. We describe deciders as phenotypes — composite people assembled from thousands of behavioral neighbors — so that a strategist and a copywriter can hold an entire segment in mind as a single person, and so that the person receives a line that was written for her.
Cohorts are built from what people do — sixty-two billion observed events — never from ethnicity, religion or party as a proxy for what they will do.
The category default: age band, sex, party, precinct
Each cohort is a person with a name, a narrative track and a register mix. Copy is written to her; the studio produces the variants; every one carries a disclosure.
The category default: one message, three demographic cuts
Every track is judged against a randomized holdout. A line that moves the Steward and suppresses the Pharmacist is caught in week two, not guessed at in November.
The category default: a post-election poll
The Steward
n., first-ring suburb
Homeowner, twenty years at the same employer, runs the parish fish fry. Believes the deal has been broken by people who never had to keep one.
"The people who wrote the last budget never had to keep one."
The Splitter
n., late-deciding
Has voted for both parties in the last four cycles and resents being told what she is. Decides in the final ten days, mostly on character.
"Two votes, two dates, on the record. Judge them yourself."
The Night Nurse
n., shift worker
Mid-career, watches the news on a phone between rounds, has no patience for anyone who has not done a hard job under pressure.
"She ran a trauma floor for eleven years. You'll hear she has never run anything."
The Renter
n., downtown, under 35
Three employers on the résumé, a lease that resets every twelve months, and a commute that has quietly become the central fact of life.
"Your lease resets in March. Your rent didn't ask permission."
The Pharmacist
n., exurban professional
Meticulous about sources, allergic to slogans, and will read the footnote if there is one. Trusts the messenger who shows the working.
"Here is the ratio she wrote, the hospital that adopted it, and the audit that followed."
The Contractor
n., rural-exurban, self-employed
Runs a crew of four, bids every job against someone cheaper, and keeps his own books at the kitchen table on Sunday nights.
"Section 4 cost you a week in March. Here is who wrote it and who voted to repeal it."
Commerce Media Civic Household Health and work Bars show the signal domains that define each composite
Across more than five hundred contests — presidential, gubernatorial, Senate and House, ballot measures, and over a hundred bond and levy elections — the programs we have designed have won ninety-six percent of the time, for candidates and causes across the spectrum. Client names are held under agreement. The methods and the margins are not.
| Office | Method | Note |
|---|---|---|
| Presidential | Hyper-localization of a national narrative | A 2016 national campaign; the message rewritten county by county for the places that decided it. |
| Gubernatorial | Behavioral targeting across three distinct strategies | Three competitive states, three different theories of the race, one measurement standard. |
| U.S. Senate | The neglect algorithm | Persuadable voters both parties had stopped contacting, found and moved. |
| U.S. House | Grassroots amplification | Small-dollar and volunteer growth engineered from behavioral affinity, not lists. |
| Ballot measures | Simplifying complexity | A 2022 California measure: 65 percent support at launch, 17.7 percent at the ballot. |
| Bonds and levies | Concrete, not abstract | More than a hundred measures; more than five billion dollars approved by voters raising their own taxes. |
| Committees and NGOs | The permanent campaign | Trigger-based activation between cycles, so the audience is warm when it matters. |
A California measure whose text ran to dozens of pages, at 65 percent support, defeated with 17.7 percent of the vote — one true sentence per voter.
The persuadable, likely voters both parties had stopped calling, found by cross-referencing contact history with receptivity.
A diaspora treated as the constituency it is, across the United States, the United Kingdom, Canada, the UAE and Australia.
The same discipline travels. Our people have run engagement, security and narrative programs for governments and institutions in the Gulf, South Asia, Southeast Asia and West Africa, and diaspora programs in five countries at once. Seven offices sit on a single desk, so a mandate in Riyadh and a race in Michigan are watched by someone who is awake.
Diaspora as an electorate. A ten-million-person community spread across five countries is a constituency, whether or not anyone treats it as one. We 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.
So there is nothing to find.
Capabilities
Every one of them ends in a number we can defend — measured against a control we drew ourselves, on the fifth day.
I
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.
II
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.
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.
III
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.
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.
IV
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.
V
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.
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.
A score is not a rule. It is 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.
Sixty-two billion observed behaviors resolved to more than 270 million adults and their households, rebuilt in full every month; digital activity refreshed daily.
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.
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.
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.
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.
Seven offices, fourteen hours of the clock, one continuous watch on every program in the field; the daily loop runs wherever it is morning.
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.
Margin arithmetic, calendar, attack lines, money against you.
Deciders scored and grouped; the lexicon drafted.
Randomized control drawn and locked before any spend.
Narrative tracks and master briefs approved.
Disclosures, coordination protocols, registrations confirmed.
Single-origin accounts, audience matches, brand safety.
First flight live; the daily loop begins; weekly field reports thereafter.
Where the race is decided, when, and at what cost. Precinct arithmetic, calendar, opposition file.
Day 1Every decider segment scored, sized and described as a person, with the track assigned to each.
Day 3The narrative tracks, the seven-step sequencing and the master creative briefs, with disclosures.
Day 7Three registers across the calendar, by week, by segment, by seat, with the absentee drop marked.
Day 12What ran, what it reached, what moved against the holdout, and what changes next week — four pages, every Monday.
WeeklyThe difference the program made, measured against the control, with confidence intervals and the model kill log.
CloseFederal, statewide and local. We take a small number of races each cycle and staff each with a full team, from terrain through proof.
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.
A question whose text runs to dozens of pages, reduced to the one sentence each voter needs — without misrepresenting the law. For and against.
Asking taxpayers to raise their own taxes, precisely: the school, the roof, the bus route, the number. More than a hundred measures passed.
Engagement, security and narrative programs with local partners, registered where the law requires it, in theaters where the ordinary playbook does not survive contact.
For qualifying programs, Brushworks backs its fees with an underwritten performance assurance. It attaches only to media we place ourselves. Terms on request.
The human layer
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.
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.
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.
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.
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.
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.
An authored library of message sequences, each mapped to the seven-step arc and to the dispositions that trigger it.
Five families — behavioral enrichment, dispositions, occupational, interests, affinities — trained on real outcomes and kept small by the publish gate.
Seventeen thousand active signals against a ten-thousand-topic dictionary; curated toward what has predicted political decisions before.
Sixty-two billion resolved events over 270 million adults and their households. Household-attributed where the source is; rebuilt in full each month.
Individuals and households, matched on hashed identifiers; client files matched under written data-use agreements and never commingled.
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.
A campaign is not a launch; it is a loop. 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 worst week, by design.
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.
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 |
Casebook
Nine cases, by method. Names are held under agreement; the terrain, the move and the margin are in the open.
Every contest we have designed a program for, one dot each. Results are verified against public returns and, where a holdout existed, against the control. Written references are available to qualified prospects under agreement.
| Office | Method | Note |
|---|---|---|
| Presidential | Hyper-localization of a national narrative | 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 | Grassroots amplification | Small-dollar and volunteer growth engineered from behavioral affinity rather than purchased lists; national-profile members on both sides of the aisle. |
| Ballot measures | Simplifying complexity | 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 | Concrete, not abstract | 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. |
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 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.
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.
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 are gone. Many are persuadable people whom the machinery has quietly written off.
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.
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.
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 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.
More than five billion dollars in community investment approved across more than a hundred measures, with margins up to fifty-seven points.
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 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.
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.
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.
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.
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 |
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.
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.
Small-dollar and volunteer bases that compounded across a cycle instead of decaying, for members on both sides of the aisle.
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.
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.
Audiences that stay warm, and activation curves that rise with each trigger instead of resetting to zero.
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 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.
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 |
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.
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.
A security posture that protected the crowd without profiling the person, and a framework the ministry runs each year.
Abroad
Twelve countries, eight domains, seven offices on one desk, and a specialty in the constituencies nobody else counts.
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.
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, 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.
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.
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.
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.
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. |
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 do. The same behavioral discipline that finds the Night Nurse in a Midwestern state 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.
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. |
WP-20 — Measurement
The models we did not ship, and why publishing that list is the most credible thing a data firm can do.
Abstract. A model catalog is only as honest as the list of what it rejected. We publish, in aggregate, a year of models retired at the publish gate or in field — the reasons, the baselines they failed, the drift that killed them — and argue that the kill log, not the win rate, is the number a buyer should ask for.
A model catalog is only as honest as the list of what it rejected. Every data firm shows a buyer the models that worked; almost none will show the ones that did not, and the omission is not a small one, because the rejected list is where the firm's standards are visible. This paper publishes, in aggregate, a year of models retired at the publish gate or in field — the reasons, the baselines they failed, the drift that killed them — and argues that the kill log, not the win rate, is the number a buyer should ask for.
The methodological sciences learned the lesson the hard way. The file-drawer problem — Rosenthal's phrase for the studies that were run, failed and never reported — inflates every published effect; pre-registration and the reporting of null results were the remedies, and the fields that adopted them became more reliable. A model catalog has the same problem in a commercial form: a vendor that trains a hundred models and reports the twelve that worked is describing a selection, not a method, and a buyer who sees only the twelve cannot tell a good method from a lucky one. The kill log is the model catalog's pre-registration. It records what was tried, what the test was, and what failed.
Every candidate model faces the same test: 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. In a representative year of the kind the annual summary reports, eighty-four candidates were tested. Forty-nine beat the file; forty-one beat the look-alike; thirty-four beat both and shipped. Seven of the thirty-four were later killed in field on drift. The gate is not a formality; it rejects most of what is put through it, and the models that ship are the survivors of a test designed to fail them.
The log records why. Most deaths are at the gate: the candidate could not beat the file, which means the behavior it used added nothing the public record did not already say for that outcome — a finding we report to the client as such, and use the file. Some beat the file and not the platforms' targeting, which is a humbling result and a useful one. A smaller number beat both and drifted: the world the model was trained on stopped resembling the world it was scoring, usually because a program's own activity changed the behavior it was watching. Three in the year were killed for reconstructing an excluded attribute — the model had found a path to ethnicity, religion or ideology through behavior, which the policy forbids and the performance did not justify. And a handful were killed for predicting only in the program they were trained on: the cardinal sin of a model that has memorized rather than learned.
A model that ships is not finished; it is watched. Discrimination in field is measured daily against the holdout, and a kill threshold is set before the program starts — a level below which the model is retired and the cohort is scored by the next-best model or by the file. Drift is not failure in the sense that the gate measures failure; it is the ordinary consequence of a program changing the world it is scoring, and the discipline is to notice it in week six rather than at close.
The firm
A political practice built around that one idea, with strategy, research, modeling, creative and delivery in a single team under a single standard of proof.
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.
Federal, statewide and local campaigns, from terrain through proof.
Donor-funded committees and issue programs, separated from the candidate practice by structure, systems and seats — not by memo.
Institutional and diaspora mandates for governments and institutions, run with local partners across seven offices.
The human layer: signals, models, the publish gate, the holdout. Serves all three practices without crossing the wall.
Master briefs, narrative tracks and variant production across three registers, every piece disclosed.
Seven offices, fourteen hours of the clock, one continuous watch on every program in the field.
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.
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.
So there is nothing to find.
Political work is trusted or it is nothing. 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.
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
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
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
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
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
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.
Verified by: audit rights; the proof report
The founder
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.
Adam Syed founded Brushworks on a single conviction: that elections are decided by people, not demographics, and that a firm willing to work at that scale — one voter, one household, one true sentence at a time — would win where the conventional campaign could not.
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.
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.
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.
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 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.
He writes for the Brushworks Review, where his papers set out the firm's method and its doctrine.
Press kit
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.
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 founded on a single conviction: that elections are decided by people, not demographics. 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.
| 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 |
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
The Brushworks wordmark is the word set in Newsreader 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.
Contact
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.
Who we work with
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.
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.
One conviction you can hold your team to: elections are decided by people, not demographics, and a program that says it moved voters should be able to prove it.
| 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.
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.
Documentation
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.
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.
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.
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.
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.
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.
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.
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.
| 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 |
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"]
}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.
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.
| 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 |
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.
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.
| 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) | A | Modeling, cohort assignment, activation, measurement |
| Household | 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.
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.
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.
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.
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.
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 |
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.
{
"signal_id": "SIG-AFF-02291",
"family": "affinity",
"topics": ["T-CIV.0182", "T-CIV.0190"],
"window_days": 540,
"aggregate": "recency_weighted_count",
"half_life_days": 210,
"state": "active",
"since": "2025-11-03",
"entered_models": 14, // passing models this signal has entered, lifetime
"last_entered": "2026-09-01",
"exclusion_check": {"edition": "2026.09", "max_auc_vs_excluded": 0.53, "pass": true},
"definition_version": 3
}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.
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.
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.
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.
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.
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.
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.
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.
model_id: M-DISP-0142
family: dispositions
task: response_to_argument
program: BW-C-2026-0417
registry: REG-2026-08
training_cutoff: 2025-12-31
label: recorded_act:returned_ballot
holdout: client · sealed 2026-09-05 · hash bw:hold:9e1c…7a40
baselines:
file_used_well: {auc: 0.671, top_decile_lift: 1.84}
platform_lookalike: {auc: 0.702, top_decile_lift: 2.11}
candidate: {auc: 0.748, top_decile_lift: 2.47, brier: 0.187}
gate_margin: {auc: 0.030, top_decile_rel: 0.10} # pre-registered
gate_result: pass (Δauc +0.046 · Δtop-decile +17%)
exclusion_check: pass (max auc vs excluded 0.54)
field_kill: 7-day rolling auc < 0.702 for 3 consecutive days
owner: head of research
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.
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.
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.
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.
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
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.
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.
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.
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.
| 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 |
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.
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.
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
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.
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.
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.
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.
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.
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.
| 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 |
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.
| 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"
}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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
| 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 |
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.
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.
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.
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.
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 keyWhere 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 |
GET …/segments | Cohorts: composite description, size, holdout size, minimum detectable effect, assigned track and registers | Never members |
GET …/cards | Model cards in force, with history | As delivered on day three and on any change |
GET …/kill-log | Every kill-log entry for the program | Gate failures, field retirements, pruned signals, withdrawn variants |
GET …/flights | Flighting plan by week, cohort, register and seat; actual against plan | Reconciled nightly |
GET …/readouts?week= | Lift by cohort, register and arm, adjusted and unadjusted, with intervals and the FDR family | Same tables as the field report |
GET …/media-trail | One row per placement: seat, platform, variant manifest, spend, reconciled invoice | Ad-library references where the platform publishes them |
GET …/ledger?since= | Ledger entries in chain order, with hashes | Chain head matches the bundle |
{
"program": "BW-C-2026-0417", "week": 3, "cohort": "S-03", "register": "precision_paid", "arm": "paid_only",
"outcome": "leading:returned_ballot", "unit": "household",
"n_treated": 18420, "n_holdout": 2046,
"lift_pts": 2.9, "ci95": [0.8, 5.0], "lift_adj_pts": 2.7, "ci95_adj": [0.9, 4.5],
"fdr_family": "W03-registers", "q": 0.031,
"balance_ok": true, "contamination": 0,
"edition": {"graph": "2026-09", "registry": "REG-2026-08", "catalog": "CAT-0417@2026-09-12"}
}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.
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.
| 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 |
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.
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.
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.
| Object | Format | Example |
|---|---|---|
| Program | BW-<practice>-<year>-<seq>, practice C (candidate), I (independent expenditure), N (international and institutional) | BW-C-2026-0417 |
| Cohort | S-<nn> within a program, with a composite name | S-03 "The Night Nurse" |
| Model | M-<family>-<seq>: BEH, DISP, OCC, INT, AFF | M-DISP-0142 |
| Signal | SIG-<family>-<seq> | SIG-AFF-02291 |
| Topic | T-<domain>.<seq> | T-CIV.0182 |
| Variant | V-<program seq>-<cohort>-<register>-<step>-<n>: PP, OA, AE | V-0417-S03-PP-04-02 |
| Ledger entry | <type>-<program seq>-<seq>: DR, KL, SL, AC, HO, DX | KL-0417-0002 |
| Edition | Graph YYYY-MM; registry REG-YYYY-MM; dictionary YYYY.MM; catalog CAT-<program seq>@YYYY-MM-DD | REG-2026-08 |
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.
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.
| 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
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.
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.
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.
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.
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.
| 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.
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.
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.
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.
| 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 |
Kill-log summary, launch edition, September 2026. Next edition: with the annual report, spring 2027.
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.
| 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.
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 | Test summary |
| Continuity | Encrypted backups; quarterly restoration tests; recovery objectives per system; seven-office operational handoff | Restoration records |
| Residency | 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.
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 |
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.
The compliance packs · Second edition
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.
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.
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.
| 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 |
| Foreign Agents Registration Act (22 U.S.C. §611–621) | 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 |
| Telephone Consumer Protection Act and FCC rules (47 U.S.C. §227; 47 C.F.R. §64.1200) | 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 |
| Children's Online Privacy Protection Act (15 U.S.C. §6501) | 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 |
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 |
| 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 |
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 |
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.
| 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 |
| Artificial Intelligence Act (Regulation (EU) 2024/1689) | 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 |
| 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 |
| 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 |
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.
| 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 |
| 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 |
| 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 |
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 |
| 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 | Consent (ePrivacy / PECR), narrow existing-customer exception |
| U.S. state laws | Notice; opt-out of sale, sharing, targeted advertising and profiling | Opt-in (most states); Utah and Iowa opt-out; Maryland bans sale | TCPA prior express consent for autodialed or prerecorded messages |
| Canada (PIPEDA / Law 25) | Meaningful consent proportionate to sensitivity; Law 25 consent by purpose | Express consent | CASL express or implied consent; donation solicitations exempt |
| Brazil LGPD | Ten legal bases including legitimate interest | Specific, highlighted consent | Consent or legitimate interest with opt-out; electoral rules on messaging |
| Australia / New Zealand | Notice and reasonable expectation; consent for secondary use | Consent | Spam Act / Unsolicited Electronic Messages Act consent; electoral messages partly exempt |
| Japan / Korea | Purpose notice (Japan); consent (Korea) | Consent (both) | Opt-out (Japan); consent (Korea) |
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.
| 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 |
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.
| 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 |
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 | Master approval records; platform verification screenshots |
| Synthetic media | 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 |
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.
Policies, disclosures and governance
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. Every commitment on this page is one the firm can be held to, and the section that follows says how.
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.
Applies to this website, to the people whose data appears in our programs, and to clients, counterparties and correspondents.
Brushworks, 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.
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.
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.
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.
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.
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 |
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.
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 |
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.
| 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 |
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, 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.
Governs every dataset Brushworks licenses, receives or produces, and every model built from it. Incorporated by reference into every client, vendor and employment agreement.
| 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 |
| Class | Examples | Handling |
|---|---|---|
| Client-confidential | Client files, strategy, creative masters, plans, results | 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 |
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.
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.
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.
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.
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.
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.
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.
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.
Version 1.2, September 18, 2026.
How the firm is run, who is accountable for what, what we will not take on, and where to take a concern.
Brushworks, 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 |
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.
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.
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.
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.
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.
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.
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 |
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.
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.
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.
Version 1.2, September 18, 2026.
What the law requires of us in the United States and abroad, what we do beyond it, and where the public can see it.
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 |
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.
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.
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.
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.
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.
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.
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.
Version 1.2, September 18, 2026.
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 | Thirteen months |
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, 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.
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.
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.
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.
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.
Version 1.2, September 18, 2026.
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.
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.
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.
Spring 2027, in preparationRegistrations filed, disclaimers carried, concerns received and resolved, mandates declined, incidents, rights requests received, and the practice-wall certification.
Semiannual, from January 2027Every model retired at the publish gate or in field during the year, in aggregate, with the reasons.
AnnualIndependent security reporting, residency and replication attestations, restoration tests.
Under agreementHow the holdout is drawn, how lift is estimated, how the baselines are constructed — the standard of proof, written down.
Published with the ReviewFor 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.
On request, under NDAFree summaries of the privacy, political-advertising and communications rules in the United States, Canada, Europe and the rest of the world.
September 2026; reviewed with each releasePrivacy, data use, governance, political and foreign-agent disclosures, cookies, terms and accessibility — this page.
Effective September 18, 2026Tell 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.
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.
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.