How PoliTracker scores

This page is the public mirror of the scoring rules in CLAUDE.md §5 and §6 at METHODOLOGY_VERSION 2026.10.1. Every number on a profile links to the score run that produced it; runs are reproducible bit for bit from their recorded inputs, and any change to a weight, threshold or rule bumps this version and re-runs every politician.

What is scored today

Four indices are defined. Effectiveness (E) is computed for every member of Congress as a legislative record, and for governors and mayors as an outcomes index over public metrics a reviewer has recorded. Promise adherence (A) is computed per term once reviewers have published its promises and the floors below are met. Alignment with app respondents in the district (C) needs published opinion cells and is shown as "Impact History (coming soon)" until then. Approval rating (I) is computed wherever a public opinion poll about the officeholder has been recorded. Each index renders "Not enough evidence (n of k)" until its floor is met, never a zero. The composite requires A and at least two indices.

Effectiveness (E), weight 0.25

Computed per term. For legislators it is the voting record below; for governors and mayors it is the executive effectiveness index that follows it.

• Productivity: for each bill (House or Senate bill, not a resolution) the member sponsored in the term, the furthest stage reached counts 1 (introduced), 2 (reported by committee or discharged), 4 (passed a chamber) or 8 (enacted), plus 0.3 for each cosponsored bill that was enacted. Commemorative bills (naming, designations, awareness periods) are excluded by a keyword rule reviewers can override.
• Participation: roll calls the member cast a vote in (yea, nay or present) divided by roll calls held while seated. "Not voting" is never labeled excused or unexcused. A leave of absence recorded in the Congressional Record (a Senate "necessarily absent" announcement, a House "leave of absence" grant) is subtracted from the denominator; a vote that two official sources disagree on is excluded from both numerator and denominator until a reviewer decides.
• Cohort: members of the same chamber of the same legislature, same session and same chamber-control side (majority or minority, the only party-derived input). For Congress the session is the Congress; for a state legislature it is the two-year session in force; for a local body the calendar year. A cohort never crosses a state line. Each member's productivity and participation become percentiles within the cohort: (members strictly below + half of the ties) ÷ cohort size. A cohort under 10 members widens to the whole body; a body under 10 members uses an absolute scale (productivity relative to the body's best, participation used directly) and is labeled so. Members whose caucus the roster does not tell (independents, fusion-line labels) carry no control side and widen to the whole body when fewer than ten share that state.
• Chamber control is data, not a rule: data/chamber-control.json records the majority party per Congress (House and Senate history) and per state chamber for the current session (NCSL's partisan-composition table, with the roster's own seat count beside it and the date it was read). Nebraska's nonpartisan Legislature records no majority. State legislators carry no recorded leaves of absence (the Congressional Record is a federal source), so their participation denominator is every roll call held while seated until a Tier A source exists.
• E = 0.7 × productivity percentile + 0.3 × participation percentile, shown as "ranked n of N in the House minority" with the underlying counts.
• Publishes after 180 days in office. Reference method: the Center for Effective Lawmaking's Legislative Effectiveness Score, with these documented deviations: no "significant" tier, cosponsor credit, percentile within cohort instead of mean-1 normalization, coverage of state and local bodies, and no committee-attendance term (no official source publishes it).

Executive effectiveness (governors and mayors)

A chief executive neither sponsors bills nor casts roll calls, so their E measures how the jurisdiction moved while they held office, across six groups of public statistics: economic health and shared prosperity (real wages against inflation, unemployment and underemployment, housing affordability); service delivery and infrastructure (condition of roads, bridges, transit and water mains; 311 and permit response times; fire, EMS and police response times); public safety and health outcomes (violent and property crime rates and clearance rates, the share of residents who feel safe at night, infant mortality, opioid overdose deaths, absenteeism from chronic illness); fiscal management (credit rating, rainy day fund, cost-effectiveness of a named program); institutional trust (resident satisfaction surveys, retention of civil service staff); and equity, civil rights and community protection (civil rights agency funding, disparities in capital investment, safety of cultural and religious sites, hate crime clearance, disparities in stops and arrests, alternative crisis response, access to services regardless of immigration status, the firewall with federal immigration enforcement, language access, the maternal mortality gap, and protection of reproductive and gender-affirming care). The full catalog, with what each metric measures and which direction counts as better, is src/domain/effectiveness/executive.ts.

• Each metric is recorded by a reviewer from the publishing agency's own release: its value at or just before the term began, its latest value, and, where one exists, the same statistic for a benchmark jurisdiction over the same window (the nation for a governor, the state for a mayor). The metric scores the change over the term minus the benchmark's change, adjusted so that a better outcome is positive, on a bounded scale: 50 means the jurisdiction moved with its benchmark, 100 or 0 means it beat or trailed the benchmark by the metric's full span (for example 3 points of unemployment, 20 % of a crime rate, 2 credit-rating notches). A metric with no benchmark is scored on its own change and is marked as unbenchmarked.
• Categories average their metrics without weighting, and E is their mean shrunk toward 50 with two phantom categories: E = 50 + Σ(category − 50) ÷ (n + 2). Scores start at 50 and move with evidence; one category can never decide the value alone.
• Publishes after 180 days in office and once at least 3 of the 6 categories have a recorded metric; otherwise "Not enough evidence (n of 3)". The tile expands to every category and metric used, with baseline, latest and benchmark values and a link to each source, and while it is withheld it lists which categories are measured so far.
• It is an index of outcomes, not of acts: it describes the territory under the executive, benchmarked, and does not claim that the executive caused the change. Party never enters.

Promise adherence (A), weight 0.40

Only specific (weight 2) and directional (weight 1) promises with a decided status count, each multiplied by its prominence (1–3); aspirational promises are displayed but never scored. Status weights: fulfilled +1, in progress +0.5, compromised +0.25, stalled 0, broken −1; "not yet due" and "unverifiable" are excluded. With shrinkage w₀ = 4, A = 100 × ((Σ w·s ÷ (Σ w + 4)) + 1) ÷ 2: scores start at 50 and move with evidence. In-office promises never exceed half of the total weight (their weights are scaled down when they would); when no campaign promises exist, A is computed from in-office promises alone and labeled "In-office promises only". Needs at least 5 scored promises, 180 days in office and complete capture coverage.

• What the profile shows. The tile carries the value, its confidence and "n promises from m sources"; it expands to the promise-by-promise table the run used (context, status scored, weight w, status weight s, w × s), the formula with its totals filled in, and the capture checklist. While A is withheld the same expansion lists the three floors with the counts behind each, so a reader sees what would change it; it never shows a zero.
• Confidence comes from the total weight of scored promises: low under 8, medium from 8 to 24, high at 25 and over. Campaign and in-office sub-scores are computed unscaled over their own promises and shown only when each has at least five scored promises.
• Capture coverage (the checklist) is the set of source kinds searched for a term's promises: the campaign issues page, the announcement or acceptance speech, a debate transcript or candidate questionnaire, and, for a term with in-office promises, the inaugural or State-of address. Each must be covered, or marked unavailable with a reason, before A publishes, so the score never rests on whichever promises happened to be collected.

How a promise gets its status

A reviewer confirms each promise before it is published: the exact quote from the archived source, the date it was made, whether it was made in a campaign or in office, whether keeping it means something passes or is blocked, how specific it is and how prominent. A promise starts as "not yet due" and a reviewer decides its status from the recorded actions of the politician linked to it as evidence. Each link says whether the action supports or contradicts the promise and how strong it is: direct (a final-passage vote, primary sponsorship, a signature or a veto), omnibus or procedural (a vote on a package, cloture, cosponsorship) or a statement. A purely procedural action can never make a promise "broken". "Fulfilled" needs a direct supporting action on a policy that passed; "stalled" needs twelve months with no action. "Broken" and "compromised" are proposed by one reviewer and take effect only when a different reviewer approves them, and so does any decision that would move a politician's adherence score by five points or more or publish it for the first time. Every change is recorded with its reason and the evidence it rested on, and is visible in the promise ledger on the politician's profile.

Approval rating (I), weight 0.15

The share of respondents in favor of the officeholder, averaged without weighting over the most recent public opinion polls recorded for them, at most five, whose fieldwork ended on or before the run's date. Job approval ("approve of the job") and favorability ("favorable opinion") are both recorded, each labeled as what it asked; the share in favor is the number averaged. Polls are read from their pollster's own release by a reviewer into a committed file with the pollster, fieldwork window, population, sample size and margin of error, and are listed with those details on the profile. Needs at least one poll. The tile shows the percentage in favor. Before 2026.10 this index was defined as measured policy impact (third-party evaluations of enacted policy); that definition is retained in ADR-0016 for a future index and was never computed.

Composite

Weighted mean of the available indices (A 0.40, E 0.25, C 0.20, I 0.15) with the weights renormalized; published only when A exists and at least two indices do, labeled "Partial score (n of 4)" when fewer than four. A composite computed from in-office promises alone carries the same "In-office promises only" label as its A.

Reproducibility and holds

Each nightly run records the methodology version, the deployed commit and a hash of its canonical inputs. A term whose composite moves more than 5 points without an explaining event is held and keeps showing its previous published value; a run is held as a whole when more than 10 % of terms are held.

Changelog

• 2026.10.1: E for governors and mayors becomes an outcomes index over reviewer-recorded public metrics in six categories, benchmarked and shrunk toward 50 (ADR-0018); the tile is labeled "Executive Effectiveness". Legislators' E is unchanged. Prior runs keep "Not applicable" for executives.
• 2026.10: the fourth index becomes the Approval rating from public opinion polls (ADR-0016); A publishable for reviewed promises. Prior runs keep their 2026.09 definition.
• 2026.09: initial version; E live for the 119th Congress.

Data sources

Federal bills, actions and sponsorships come from GovInfo Bill Status and Congress.gov (Library of Congress); House roll calls from Congress.gov and the House Clerk; Senate roll calls from senate.gov; rosters from unitedstates/congress-legislators. State bills, actions, sponsorships and roll calls come from LegiScan, with bill text fetched from each state legislature's own site; state and local rosters from Open States; city matters, votes and council rosters from each pilot city's own Legistar site (a public record), with every council member linked by a reviewer before any vote is attributed. Every stored row carries its source, the time it was fetched and a content hash. Legislative data provided by LegiScan (legiscan.com).