Methodology

Qualitative intelligence, treated as a discipline.

A repeatable method. An auditable standard of evidence. A tiering system that distinguishes what is known from what is inferred. An honest account of what the method cannot do.

The premise

The voter file tells you who to contact. It does not tell you what they believe, who they trust, or why your message is not landing.

The dominant tools of American political data are descriptive at industrial scale and explanatory not at all. They give every operator the same vocabulary — turnout scores, persuasion scores, issue salience — and that vocabulary, repeated long enough, comes to feel like understanding.

It is not understanding. It is a sketch of a population with no account of what the population believes. Every consultant in the country knows this, privately. None of them have built a discipline for the missing layer.

We have. What follows is the protocol, the tiering model, and the guardrails — written down with enough specificity that you can hold us to it.

The Four-Step Protocol

How a community profile actually gets made.

Site Entry & Rapport

Researchers establish trust with civic anchors — faith leaders, mutual-aid networks, longtime residents, the people communities already turn to.

Site entry is the part of qualitative work the rest of the industry skips. We don't. Before any interview is conducted, our researchers spend time on the ground identifying and earning the trust of the people whose endorsement makes the rest of the protocol possible. This is unglamorous, slow, and non-substitutable. It is also the difference between intelligence and noise.

Inputs

District scope · existing relationships

Outputs

Anchor list · entry permissions

Ethnographic Interviewing

Semi-structured, open-ended interviews and participant observation in the community's own setting, on the community's own terms.

Interviews are conducted in person where possible, by researchers trained in ethnographic method. Question protocols are semi-structured: a spine of comparable questions across sites, with room for the conversation to follow what the respondent actually wants to say. Sessions are recorded with consent and transcribed verbatim. No leading. No script-reading. No agenda imported from headquarters.

Inputs

Anchor introductions · interview protocol

Outputs

Transcripts · field notes · observation memos

Thematic Coding

Transcripts coded inductively against emergent themes — the discipline of qualitative method, not the assembly of anecdote.

Transcripts are read, re-read, and coded by analysts using inductive thematic analysis. Themes emerge from the material rather than being imposed on it. A second analyst reviews the codebook. Disagreements are resolved by adjudication, not averaging. The output is a structured account of what the community itself raised, ranked by salience and supported by direct quotation. AI-assisted first-pass coding is permitted and logged; final codebook adjudication is human.

Inputs

Transcripts · field notes

Outputs

Coded themes · ranked salience · supporting quotation

Evidentiary Tiering

Every profile is tagged by evidence maturity. Consumers of the intelligence see exactly which tier any given claim is operating at.

This is the part that lets the work move at the speed of a cycle without collapsing into hand-waving. Every claim that leaves our platform carries an evidence tag — AI Baseline, Field-Refined, or Vetted Intelligence — and a campaign manager opening a community profile sees, on the same screen, what is grounded in primary research and what is not. The tiering is the discipline. The tiering is also the guardrail.

Inputs

Coded themes · review chain

Outputs

Tagged community profile · published artifact

The Evidentiary Tiering Model

Every claim carries a tier. The tier is the discipline.

Operators consuming our intelligence see the evidence maturity of every claim on the same screen as the claim itself. The tiering is what lets us move at cycle speed without confusing synthesis for primary research.

AI Baseline

Evidence: Synthesis of public sources, news archives, and demographic context by a frontier LLM.
Review: Automated cross-check against source provenance. No human adjudication.
Use for: Background context. First-pass orientation. Onboarding a new district at speed.
Don't use for: Decision-grade targeting. Public claims. Anything attributed as primary research.

Field-Refined

Evidence: AI Baseline corrected against at least one round of primary interviews or on-site observation.
Review: Analyst review with named reviewer logged. Source quotations attached to load-bearing claims.
Use for: Internal campaign strategy. Message development hypotheses. Field briefings.
Don't use for: Long-form attribution as definitive community position without further interviewing.

Vetted Intelligence

Evidence: Full protocol applied: site entry, multiple interviews, thematic coding, second-analyst adjudication.
Review: Two-analyst adjudication, full chain logged, prompt and model versions stored per artifact.
Use for: Decision-grade targeting, persuasion strategy, principal-level briefings, published reporting.
Don't use for: Generalization beyond the studied community without explicit scope notes.

AI Posture

We use frontier models aggressively. We do not let them launder inference as fact.

Large language models are the only reason this kind of work is now feasible at the cadence campaigns require. We use them for synthesis, classification, and first-pass drafting. We do not use them as the source of truth.

Every community profile carries one of the three evidence tags above. Every published claim is reviewed by a human analyst. No decision-grade output is generated from voter file inference alone. Prompts and model versions are logged per artifact, so an audit of any claim can reconstruct the exact chain that produced it.

AI does not solve the thin-data problem. It accelerates whatever you point it at. Pointed at thin data, it gives you thin description at machine speed. Pointed at the output of a disciplined qualitative process, it gives the field its first real chance to make thick description scale.

What this isn't

The things we refuse to do, written down.

No voter file inference alone

We will not produce a community profile from quantitative scoring data without primary research. The two are complements; substituting one for the other is the failure mode we exist to correct.

No synthetic personas marketed as research

AI-generated 'representative voters' are useful for internal hypothesis testing and nothing else. We never publish them as findings, and we never let a synthetic composite stand in for a real interview.

No cross-client data sharing

Community intelligence developed for one engagement does not flow to another. Separate teams, separate data isolation, disclosed roster on request. The integrity of the work depends on this and there are no exceptions.

See the methodology applied to a community you care about.

Every engagement starts with a conversation. Tell us the district, the question, or the program — we'll show you the protocol against your real problem.

Request a Demo Read the field notes →