Open seat, R+3. Macomb County - Warren, Sterling Heights, Eastpointe, Roseville. John James left it for the governor's race; both parties call it one of the most competitive House seats in America. This is the playing field before anyone spends a dollar.
REGISTERED VOTERS REAL
625,152
state voter file with full participation history - every general, every voter, back years
MATCHED + ON THE BOARD REAL
376,279
resolved through the IDGraph to reachable identities - email, phone, address, device. Every one appears on the Control Center board.
Step 2 - Archetypes Emerge
The full electorate, grouped bottom-up - not stereotyped top-down
Before a single interview, the whole file gets structure. Every matched voter is matched to 82 real, verifiable facts about their life - age, income, education, homeownership, net worth, household, geography. Then a grouping method lets natural groups form on their own from similarity - like sorting a crowd by who resembles whom, without naming any groups first. (For the technical reader: BIRCH-Ward hierarchical clustering.) Nobody draws "soccer moms" or "NASCAR dads" and sorts voters in - the machine proposes the groupings, the data decides how many, and the names come last, as cosmetic labels on structure that already exists.
HOW K WAS CHOSEN REAL - silhouette sweep
We never decide the number of groups - we test every count from 2 to 21 and score how cleanly voters separate at each. The score climbs, peaks, and the winning count (14) is read off the chart. The data answers; we just ask the question.
WHAT IS FORBIDDEN AS INPUT discipline
Purchase history: never used for grouping. Voting history: never used for grouping. Race: never a primary lens. The groups are built only on who people verifiably ARE - everything else is kept out on purpose, so a stereotype cannot sneak in through the side door.
3,600 REAL MI-10 VOTERS - each dot is one person, and people who resemble each other sit close together REAL
Voters who resemble each other on real attributes sit near each other. The structure is already visible as density - the machine has not been told anything yet.
HOW MANY GROUPS? THE DATA DECIDES REAL - silhouette sweep
Plain English: we ask "how many groups should there be?" by testing every answer from 2 to 21. For each count, a silhouette score measures how cleanly voters separate - how much closer each voter is to their own group than to the nearest other group. Higher = crisper groups. We read the answer off the data; we never assert it.
Result: six archetypes covering all 376,279 matched voters, every voter in exactly one REAL. These are the quantitative WHO. They say nothing yet about WHY anyone votes - that requires listening, which is the next step.
Step 3 - Listen
Long conversations, decomposed into atoms
AI-moderated interviews - conversations, not questionnaires. No issue list is handed to the voter; no categories exist yet. Each transcript is decomposed into propositions - every distinct belief in the voter's own words, carrying its full structure: the claim, who they blame, what they'd do about it, and how hard they hold it.
WHO GETS INTERVIEWED sampling design
Not a random blob, and not volunteers from the internet. We sample inside each of the 14 groups from Step 2 - every group gets its own conversations, weighted so the small groups are heard, not drowned out by the big ones. Every participant is verified against the actual voter file before the interview starts: a real MI-10 voter, in the group the file says they are in - not a professional survey-taker.
HOW MANY IS ENOUGH the stopping rule
There is no magic number - there is a rule. We keep interviewing a group until new conversations stop revealing new themes (researchers call this saturation), then we re-run the whole analysis from scratch - if the same structure does not re-emerge, we keep going. In practice that means roughly 25-50 deep conversations per group - several hundred for a district like MI-10. For comparison: a classic poll buys 500 five-minute answers to our questions. This buys hundreds of 20-40 minute conversations in their words.
And it never stops: panels are re-interviewed through the race, which is how the system sees priorities actually moving - not just where they started. MI-10 program figures shown are the sampling design; panel recruitment runs on this plan.
AI-MODERATED INTERVIEW - WARREN, MIILLUSTRATIVE - reconstructed example
Thanks for making time. To start - how are things going for you and your family these days?
Honestly? Tired. Working a lot. The kids are good - that's what matters.
voter
You said you're working a lot - what does a normal week look like?
I do billing at a medical office, and most weekends I pick up shifts at my sister's salon. It didn't used to take two jobs to stay even.
voter
What changed?
Everything costs more and nobody's paycheck moved. Groceries are stupid. Daycare for my youngest costs more than my rent did five years ago.
voter
Earlier you said the kids are what matters. What do you worry about for them?
School, mostly. And their phones - what they see on there. My mom helps with pickup; she's on Medicare and I'm the one reading her bills every month. It's a lot.
voter
If you could fix only ONE of the things you've mentioned - costs, the kids' school, your mom's healthcare - which comes first?
Costs. If the math worked, half the rest gets easier. Then school. Mom's stuff scares me, but we manage.
voter
Who do you trust to fix any of this?
[laughs] Nobody? I don't trust either party to actually do it, honestly. Show me somebody who's lived it.
voter
Notice what the moderator never did: named an issue first, mentioned a party or a candidate, or suggested an answer. It followed her words. And the ranking question at the end matters most - she just told us her priority order in her own voice. That ordering is exactly what the next step is built from.
FROM THAT CONVERSATION, THE SYSTEM EXTRACTS THE BELIEFS - each with its full structure:
"Everything costs more and nobody's paycheck moved. My mom's on Medicare and I'm the one reading her bills every month. And honestly? I don't trust either party to actually fix any of it." ILLUSTRATIVE transcript excerpt
claim: essentials outrunning wages · blame: corporations + both parties · prescription: price accountability · held: hot - high heat, high energy
claim: carrying a parent's healthcare burden · blame: system complexity · prescription: protect Medicare, simplify · held: personally - lived daily, high stakes
claim: neither party will deliver · blame: political class · prescription: none offered · held: with resignation - firm, but expects no fix
In testing, the interview system correctly recovers the district's real themes 79% of the time. It measures what is actually there - it never plants topics and finds them back. REAL - instrument validation
Step 4 - Stack Identities Emerge
A stack identity = a trunk, decomposed and rank-ordered
Propositions organize upward into trunks - the big themes the district actually talks about, weighted by how much. But the trunk alone is not the insight. The insight is the stack identity: the trunk at the top, decomposed into its child frames, rank-ordered by how the voter holds them. Same trunk, different order = a different person to persuade.
ECONOMIC SECURITY - stack identity A
1Cost of essentials
2Law & order / neighborhood decline
3Government corruption
4Wages & job security
ECONOMIC SECURITY - stack identity B
1Cost of essentials
2Education & childcare
3Healthcare burden
4Housing
Same trunk, two different stack identities - you would never message these two groups the same way, and no demographic file can tell them apart. The hierarchies shown illustrate the method. Our standing rule: categories come OUT of the conversations - we never feed them in. And any claim that a pattern emerged must survive being re-run from scratch before we believe it.
THE FIVE MI-10 TRUNKS - district weight REAL - instrument
Step 5 - Cells
A cell = an archetype × a stack identity
Two different machines have now read the district. The voter file gave us archetypes - quantitative clusters of who people are on paper (the six from Step 2, REAL). The interviews gave us stack identities - the hierarchies of why they decide. The expansion engine crosses them: it assigns a stack identity to every voter inside each archetype, expanding what the interview panels measured onto the full file. The intersection is a cell - and the cell is the unit everything downstream runs on.
ARCHETYPE REAL - voter file, k=6 × STACK IDENTITY from interviews = CELL
Warren Families (192,813 voters) × stack A (essentials → law & order → corruption → wages) = cell, est. 64,000
Warren Families (192,813 voters) × stack B (essentials → childcare → healthcare → housing) = cell, est. 79,000
Affluent Senior Homeowners (18,579 voters) × stack A = cell, est. 7,400 - same stack identity as the first row, entirely different people to reach
Why the cross matters: the same archetype splits into different stack identities - two Warren families with identical census profiles can carry opposite hierarchies. And the same stack identity appears across archetypes - the message travels, the channel and voice change. Neither machine alone can see this. And there is still no distance between stacks - you cannot compute how far stack A sits from stack B. Segmentation, not clustering.
Archetypes and their populations: REAL. Stack identities and the expansion assignments: MODELED v1 until the discovery engine and VoterVerify-recruited interview panels run - the expansion engine assigns from shared signals, panels validate, post-race results grade it. The seeds are labeled seeds.
Step 6 - The Atom
One voter
The resolution the system runs at. A real voter on the board: Warren, 40s, votes most generals (turnout band T3). Her archetype: Warren Families (REAL, voter file). Her assigned stack identity: B. Her cell is that cross - and everything the campaign will ever say to her is generated from this hierarchy and aimed only at its moveable layers.
HER STACK IDENTITY · cell-level, MODELED v1
1Cost of essentialsMOVEABLE
2Education & childcareMOVEABLE
3Healthcare burdenFIXED
4HousingFIXED
Individual stack identities come from interview panels; until hers is measured, she carries her cell's structure. No consultant, no poll, no file has ever shown a campaign this object. This is the atom.
Step 7 - The Handoff
The READ is complete before day one
Race → conversations → propositions → stack identities → cells → a single voter. All of it exists when you turn the key - and none of it mentions a candidate yet. Screen 2 is FIT: the Hines proposition set meets these hierarchies, layer by layer, and the creative engine cuts keys for the locks that already exist.