How Voter Compass research works

What Voter Compass is (and is not)

Voter Compass produces **data-driven projections of alignment** between your stated priorities and what candidates have said and done. It is **not an endorsement engine** — no human editor ranks candidates, and no candidate can pay to change a score.

Candidate lists come from **official filing sources** (NH Secretary of State filing lists, enriched with Ballotpedia and campaign sites), imported through a human-reviewed pipeline — never manual entry, never model memory. Races included in this build list candidates from the official NH Secretary of State cumulative filing list. Where evidence is lacking, a candidate shows "not yet scored," never an invented number.

Where the data comes from

**Candidates and races.** The candidate list is imported from official filing sources: the NH Secretary of State's candidate filing list is the source of truth, enriched by Ballotpedia race pages and campaign websites. Every import is staged for human review before it replaces anything voters see, and a parsed record is only accepted when the candidate's name appears **verbatim** in the fetched source document — the import model cannot introduce names the document does not contain.

**Research sources.** For each candidate, the research pipeline fetches and retains the text of the candidate's **campaign website** plus **typed public sources**: candidate questionnaires (Vote 411), voting records (Congress.gov, Just Facts Votes), non-partisan reference profiles (Ballotpedia, Vote Smart), and news coverage. Database adapters attach these automatically when the candidate's name appears verbatim on the fetched page; operators can also attach sources manually. The source's type sets its evidence weight — **deeds (roll-call votes, self-authored questionnaire answers) count more than campaign-site claims**. Every published evidence item cites the URL of a source the pipeline actually fetched, and its excerpt must appear **verbatim** in that source's retained text; anything else is discarded before publication. No quote you see was written by a model. Each candidate profile lists the exact sources used — with type, origin (operator-attached vs auto-discovered), and evidence weight — in its **"Sources Used"** section, so you can audit which databases informed a score.

**Campaign finance.** For federal races, finance figures (total raised, spent, cash on hand, small-dollar share) come from the **Federal Election Commission's public records** and always link to the official FEC page for verification. Money is shown in a separate transparency panel only — it is **not an issue position and never affects issue scores or the radar chart**. Races without a keyless official tracker (state races) show no figures rather than unverifiable ones.

**Geography.** Town-to-district mappings start from a simplified dataset; each town is checked against Secretary of State district records and flagged **verified** by an operator. Unverified towns are labeled on the ballot page.

How issue scores are computed

Each candidate receives a **0–100 score per issue category**, produced by our research pipeline:

1. **Gather** — the pipeline fetches and retains the candidate's campaign website plus configured public sources (Ballotpedia, news coverage, voting records).

2. **Extract** — the retained texts are reduced to position statements tagged by category. Only excerpts that appear verbatim in the retained source they cite are published; anything else is discarded.

3. **Score** — positions are scored for depth and specificity in that category, with confidence capped by how little source material exists.

Each score carries a **confidence value (0–100)** reflecting how much evidence supports it. Low-evidence scores are visibly lower-confidence, and they count for less in matching.

**Coverage gate:** scores are only published when a research run clears a minimum bar (at least 3 verified evidence items and overall data confidence of 35+). Below the bar, the candidate shows **"insufficient evidence"** with whatever verified evidence was found — never partial or padded scores.

Trustworthy rating and consistency index

**Trustworthy rating (0–100):** measures whether a candidate's words match their deeds by pairing statements with actions (votes, contracts, administrative decisions) on the same topic across the gathered sources. Aligned pairs raise the rating; contradictions lower it. The rating is only published when the run clears the coverage gate; with thin sources it stays unpublished rather than guessed.

**Consistency index (0–100):** measures position stability over time — higher means fewer unexplained reversals. The same coverage gate applies.

Every contradiction counted is shown on the candidate's profile as a flagged evidence item with its source, so you can judge it yourself.

How your match percentage is computed

You **rank** your priority topics in order of importance (ties allowed). Ranks convert to weights deterministically:

- **weight = 0.8^(rank − 1)** — rank 1 → 1.00, rank 2 → 0.80, rank 3 → 0.64, and so on. Tied topics share the same rank and therefore the same weight.

- For every ranked topic where the candidate has an effective score (direct or rolled up — see the topic tree section), we compute **weighted points** = weight × score × (0.5 + 0.5 × confidence/100). The confidence factor means a thinly-evidenced 90 counts for less than a well-evidenced 80.

- **Max points** for that topic = weight × 100.

- **Match % = 100 × Σ weighted points ÷ Σ max points**, over your ranked topics the candidate has an effective score in.

Candidates who have not been researched yet show **"not yet scored"** rather than a fake number. The formula is deterministic: the same inputs always produce the same ranking, and the full per-topic breakdown — including each topic's rank, derived weight, and whether the score was a roll-up — is shown under every result.

The topic tree: subtopics, roll-ups, and emerging issues

Topics form a **tree**: broad categories (Taxes & Budget) contain subtopics (Property Taxes, State Spending). You can rank a broad category or drill into its subtopics — whichever matches how you think about the issue.

**Scores live at the most specific topic evidence supports.** When a candidate has no direct score at a topic you ranked, but does have scores at its subtopics, we compute a **roll-up**: the confidence-weighted average of subtopic scores (each subtopic counts in proportion to its confidence), with the roll-up's confidence set to the average of contributing confidences — aggregation never inflates certainty. Roll-ups are labeled in every breakdown. If neither the topic nor any subtopic is scored, the topic simply contributes nothing — we never fill gaps with guesses.

**Geographic scope.** Some topics only matter in part of the state (e.g. data-center siting in one county). Scoped topics appear only for voters whose selected town is in the relevant area.

**Emerging topics.** Anyone can suggest a topic. Suggestions are human-reviewed; approved ones join the tree flagged **"emerging — research pending."** Candidates show as **not yet scored** on emerging topics until real evidence is gathered — an emerging topic never invents scores.

Behavioral reliability — how often words become action

Beyond issue positions, Voter Compass reports a **behavioral reliability layer**: how reliably a candidate's documented behavior predicts their behavior. It is **never a personality, temperament, or likability judgment** — every input is a documented, citable behavior of the candidate, and third-party characterizations or opinion pieces are never inputs.

**Transparency behaviors (all candidates).** Two observable facts are recorded: whether the candidate answered the Vote 411 questionnaire — counted only when a **verified verbatim excerpt from the response itself** is on record, never merely because a questionnaire page is configured — and whether a verifiable campaign financial disclosure exists. Non-response is itself recorded as a citable data point — it is shown honestly, never guessed around. Finance figures remain transparency-only: the disclosure signal records only that a verifiable record **exists**, and money never affects any score.

**Follow-through rate (incumbents only).** For candidates with an official record, the research pipeline pairs campaign statements with subsequent documented actions (roll-call votes, sponsorships) from voting-record sources. Aligned pairs and contradictions are aggregated into a **follow-through rate** — the share of documented statement-to-action pairs where the stated position carried through — and **every pair shows both sides**: the statement (with a verbatim excerpt and link to its source) and the action record. A pair only counts when both legs are deterministically verified — the action must cite a deed-type document the pipeline actually fetched, the paired statement excerpt must appear verbatim **in full** in a fetched statement-type source, and a "followed through" pair additionally requires the statement's cited publication date to predate the action — pairs with missing or out-of-order dates are discarded rather than guessed. Unverified pair claims can never move the rate. Challengers show "not applicable" rather than a rate, because absence of a record is never held for or against anyone.

The rate is published only above the same coverage gate as issue scores; below it, the profile shows honest absence. The indicator is displayed separately from issue scores and the radar — it is a **reliability weighting on issue projections** ("stated positions historically translate to action"), not a virtue score.

Current data coverage

This build targets the **NH State Primary, September 8, 2026**. Included races list filed candidates from the **NH Secretary of State cumulative filing list (07/16/2026)** across federal, statewide, Executive Council, State Senate, State House, and county offices for both party primaries. Included races where no one filed for a party's primary are shown as having no candidates.

Research coverage arrives candidate by candidate: candidates without verified evidence show **"not yet scored"** rather than an invented number. A candidate's profile lists the sources actually used.

Town-to-district mappings begin from a simplified dataset; towns verified against Secretary of State records are marked accordingly, and unverified towns carry a notice on their ballot page.