Methodology versions
| v3.0 · August 31, 2026 | Honesty is no longer numerically scored — a promise this site made and could not keep, retired in the open. Since v2.1, 25% of every legislator grade was reserved for promise-keeping, publishing “only with evidence trails and two agreeing human reviews.” This site has one reviewer. Nothing ever published through that gate, and nothing ever would have; every member’s card disclosed a reservation that was, in truth, permanent. Before changing anything we ran four competing automation designs through an adversarial review, and every route to a numeric honesty score failed the site’s own standards: the scores were either degenerate (nearly everyone at 100%), purchasable (stuffable with safe pledges, gameable by silence), or laundered a model’s judgment about a named official into print. What replaces it: the Integrity Record — three factual ledgers plus preregistered grade caps, with a published rulebook. (1) Checkable commitments: spoken floor pledges that name their own test (“I will vote against H.R. 25”) are compiled by deterministic code — no model anywhere — and tested against recorded votes; everything vaguer is excluded with a published reason code. Absence of evidence is never a penalty: an unkept commitment costs delivery credit, printed beside the verbatim quote, never a claimed verdict on intent. (2) Name-vote consistency: formal endorsements beside final-passage votes, counted only when the text at the vote is the text they signed — run against all 163 qualifying passage votes of the 119th Congress it yields 27 qualifying pairs and zero adverse rows, because on the modern floor the endorsed text and the voted text are almost never the same document. That finding publishes rather than being papered over. (3) Adjudicated record: findings by courts and evenly-bipartisan-by-rule bodies — the only tier that touches a grade, as a cap (conviction → F, censure-with-report → C−), never a bonus, with every panel printing exactly which sources were searched and which are not yet connected. The formula change. The retired 25% is not redistributed by judgment: the four surviving weights each scale by exactly 4/3, so every grade changes by the same constant and no rank moves except at the winsorization bound. This is the first sanctioned exception to “weights never renormalize” — that rule protects readers from partial grades masquerading as whole ones while a block waits for data; a block the methodology no longer scores at all is not waiting. The governors’ card slot now reads “Not numerically scored” instead of “In human review”; the governors’ edition of the Integrity Record is future work and is labeled as such. |
| v2.7 · August 28, 2026 | Attendance is now scored — and the numbers come from Congress itself. The participation block (10% of the legislator rubric) had been reserved for one stated reason: the only missed-vote figure available was a lifetime careerrate from a third party, which is the wrong clock for a per-Congress rubric. So we stopped relying on it. This site now counts missed votes itself — 645 House roll calls and 890 Senate votes of the 119th Congress, read from the records the House Clerk and the Senate publish about themselves, each one accepted only after our own tally reproduced the chamber’s published totals exactly. Every rate is shown with the numerator and denominator it came from. The rule. Participation is a one-sided hinge, not a z-score: perfect attendance earns nothing, and only absence beyond the member’s own chamber’s 75th percentile subtracts, capped at a quarter of a standard deviation. Three quarters of both chambers score exactly zero and are untouched. Two guards are part of the rule, not exceptions to it: attendance may never be the only block a member is graded on (82 members would otherwise have entered the rankings on attendance alone, most of them tied at zero, out-ranking members measured on five times as much of the rubric), and a rate is published at 25 recorded votes but only scored at 100, because with a denominator of a few dozen a single absence swings the rate by more than a point. What moved. Coverage rose from 40% to 50% of the rubric for a typical House member and 65% to 75% for a senator. The largest fall is Rep. Nancy Mace, No. 148 to No. 249, on 118 missed votes of 645. No member gained more than 14 places, because the block cannot reward anyone — those members rose only because others fell. Median missed votes are 2.20% Republican and 2.00% Democratic in the House, 2.10% and 2.05% in the Senate, so the block imports no partisan skew. The roll call records whether a member voted and never why, and that caveat travels with the figure everywhere it appears. |
| coverage · August 30, 2026 | The bill corpus went from 14 laws to 31, and the selection queue is now empty. Every law the preregistered screen finds — any enacted public law since 2013 whose text carries a single dollar figure of $50 billion or more — has now been read: 2,149 laws scanned, 44 candidates, 44 extracted. Twenty-six cleared the $100 billion entry test; sixteen did not; and two more are on the site because they were extracted before the rule existed and the corpus is append-only — the Fiscal Responsibility Act at $66.7 billion tracked and the CHIPS and Science Act at $54.2 billion. The sixteen refusals have their verdicts published with the reason, which is the half of the rule that had been running in private. It is where the famous laws are: three Bipartisan Budget Acts and two NDAAs that appropriate nothing, a debt-limit statute (a borrowing ceiling authorizes no spending), the FAST Act (highway contract authority, not appropriations), and the National Security Supplemental, a real appropriations act that misses the threshold by under $7 billion and is excluded anyway because the threshold applies identically to every law. Two entries record the screen’s own false positives: the $250 billion in a bank-regulation law is an asset threshold, and the $300 billion in a foreign-assistance law is a congressional finding about the world. An absence a reader has to notice on their own is indistinguishable from a choice. |
| correction & method · August 30, 2026 | A new class of error was found in our own extractions, and a check now exists for it. Every figure on this site is mechanically re-located in the source as an exact string — and that check is blind to an invented description sitting beside a correct number. An extraction of the FY2026 defence authorization carried eleven fabricated military-construction descriptions (“B-21”, “beddown”, “hospital replacement”) on line-items whose amounts were all right; the tables it read carry only State, Installation and Amount. The pipeline now checks every capitalised phrase in a description against the enacted text, and that check found 90 more across later laws — an invented NASA directorate breakdown, a “Post-9/11 GI Bill” label on an account the Act calls Readjustment Benefits, a “nuclear warhead stockpile” description the text never gives. Common names that genuinely help a reader are kept but marked as glosses; characterisations the text does not support are gone. None of these touched a dollar, a mechanism or a citation, which is why nothing caught them before. |
| new data · August 30, 2026 | Dollars from different years can now be compared, and the government’s own end-of-year breakdown is published. Every dollar figure on this site was nominal, which meant a 2013 law read about a third smaller than a 2025 law of identical size and no page could say otherwise — the charts carried a warning (“a longer bar can be higher prices rather than more spending”) that was honest and a dead end. Two datasets close it, both from OMB’s Historical Tables and both keyless: total outlays broken out into the 19 budget functions the government appropriates against, for every closed fiscal year since 1940, and the composite outlay deflator OMB deflates its own tables with — a spending-weighted index over what federal outlays actually buy, deliberately not a consumer price index, because the federal government does not buy a household’s shopping basket. Two new pages use them: Federal Spending by Year, which shows a closed year’s composition and puts any two years side by side, and Bills, Side by Side, which does the same for any two tracked laws. Both refuse the easy version: the breakdown must reconcile, function by function, to the outlays this site already published from a different OMB table (itself checked against the Treasury’s own statement) or it does not ship; a function OMB reports as unavailable prints as no figure rather than as zero; OMB’s estimate columns for future years are dropped at ingestion rather than greyed out on a page; and the bill comparison declines to line up spending categories across laws, because each law’s categories come out of its own text and a shared taxonomy would be an invention of ours sitting where a fact should be. |
| methodology & correction · August 29, 2026 | Bill selection is now preregistered, and our own audit is why. The stated rule (“enacted, at least $1 billion, recorded floor vote”) turned out to admit roughly 160 laws while this site extracts a handful — an unstated “landmark” judgment was doing the real selecting, and an unstated filter is where a preference can hide. The rule is now stated in full: a mechanical screen (any enacted law since 2013 with a single figure of $50 billion or more in its text) produces a published queue, a law enters the corpus when hand extraction verifies $100 billion of budget authority, thresholds move only downward and only for everyone, and a recorded vote in either chamber qualifies — a chamber that passed a law by voice vote gets that said on every member’s row rather than invented positions. Also corrected: every tracked roll call now cites the chamber’s own record (two had cited a third-party tracker; the positions matched the Clerk exactly on all 293 overlapping members, but a reader checking a federal vote should land on the federal record), and the one law the House enacted by agreeing to a resolution now says so beside its citation instead of leaving the discrepancy for a reader to find. |
| feature · August 26, 2026 (4) | Five new charts, and a page for the debt. The site now draws what it previously only tabulated: a new Debt page charting every fiscal year since 1790 against the annual deficit (with linear, log and per-dollar-of-revenue scales, because each hides something the others show); a ranked board of every graded member of Congress; a grade card that draws its own rubric, with the components we refuse to score rendered as hatched, explicitly-labelled gaps rather than zeroes; a clickable map of where an enacted bill’s dollars go, ending in the statute’s own words; and, on president pages, every completed term measured against the identical metric set. None of these is a score, a rank of people, or a verdict — they are the same published numbers, drawn. |
| correction & security · August 26, 2026 (3) | An adversarial review of this site, and what it found wrong. We ran a red-team pass against our own pipeline and fixed what it caught. The worst was ours: a new leadership measure ranked members against every leadership title-holder and called that group “the only fair peer cell… every one carrying the same structural discount.” That was false — the group mixed floor leaders with honorific posts whose holders legislate normally — so the comparison was narrowed to floor-leadership roles and the false premise removed. The Congress’s enacted-law count, which had appeared only on the two majority-party leaders’ pages, now appears on every floor leader’s page in both parties, because a fact shown to one side and withheld from the other is a slant. A suppressed district figure had been dropped from its table rather than labelled; it now reads “Not published.” A typed superlative left the shutdown share text. On the security side: the reader Q&A now verifies every number in an AI answer against the record before a reader sees it and refuses the answer otherwise, sends the model only page-published fields, and checks the request’s origin; the email auto-responder now authenticates senders before replying. Full detail in the commit history. |
| feature · August 26, 2026 (2) | Ask about this bill — a labeled AI answer box on bill pages. A reader can now ask a question and get an answer generated by Claude from that page’s verified record only — the cited line items, disclosures and CBO estimate — with instructions to refuse anything the record doesn’t cover. This is the one place a model writes on this site, and it is labeled as exactly that: AI-generated at the reader’s request, never the site’s editorial voice, and if an answer ever disagrees with the page, the page is right. The build-time rule is unchanged — no sentence the site itself asserts about a bill or an official is model-written. |
| coverage update · August 26, 2026 | The starting line, and leadership measured. Two member-page additions. First: House pages gain the district’s own Census figures (ACS one-year levels — unemployment, median household income, poverty, college attainment, uninsured), each beside the national number, as context that never enters any grade; crime and tax burden are not in the table because no official source publishes either at congressional-district level, and that gap is stated rather than filled. Second: the effectiveness score’s known blind spot for leaders — it counts only bills a member sponsors themselves, and leaders move other members’ — is now measured instead of merely disclosed: every current leadership-post holder is ranked inside the leaders-only cohort of their own chamber, and the Speaker’s and Senate Majority Leader’s pages report the Congress’s enacted public-law count (govinfo’s own listing) as context, never as a score. |
| coverage update · August 25, 2026 (2) | Public safety, monthly — city pages gain the city’s own hate/bias-crime record. One preregistered rule for all 50 cities: an official municipal or police open-data feed gets the identical module — monthly counts by the city’s own bias motives, same-months comparisons, both directions, never scored, never part of any grade; no qualifying feed, no module, no inferred number, and the feed must be currently maintained. All 50 roster cities’ portals were surveyed with live fetches; launch coverage is the seven that qualify — New York, Austin, Dallas, San Diego, San Francisco, Washington and Louisville — with the near-misses excluded for stated reasons (Los Angeles publishes no bias motive and migrated records systems mid-series; Tampa keeps only a rolling year; Tucson’s dataset has been dormant since 2018; forty cities publish no official machine-readable hate/bias data at all, several offering dashboards, which are not data). The module was built after a claim about one city’s numbers failed verification against the primary source — the fix for a contested number is the number, published under a rule that existed before anyone asked. |
| coverage update · August 25, 2026 | The October 2025 shutdown record joins every member page. The 43-day lapse — the longest on record — was a failure of the one duty that is Congress’s alone, so all 531 member pages now carry the same section: the member’s recorded votes on H.R. 5371, the funding bill, on both sides of the lapse, each linked to the chamber’s own roll call and staged only after the parsed tallies matched the officially published results (House 217–212 and 222–209; Senate 55–45 rejected and 60–40 passed; the Senate’s 14 rejected funding-advance votes counted from its own vote menu). No grade or rank changed: a penalty applied identically to all 531 members cancels out of a rank-based system, and apportioning blame beyond the votes is a causal judgment this site does not make. The record is the accountability — every member wears their own votes, both parties, forever. |
| v2.6 · August 24, 2026 | The Senate tenure ramps are removed — one formula for every senator. Through v2.5, a senator’s state-outcomes block blended trend and current condition by a tenure ramp (condition share 0 for a first term, rising to 40% at 24 years), and a second tenure ramp scaled the block’s weight (25% → 100% over 12 years). Both scored different senators with different formulas and printed them in one ranked list — the exact device v2.4 removed from the governor board, with the reasoning published: a fixed weight is the only honest basis for an ordinal list. An outside methodological review pointed the v2.4 argument back at the Senate board, and it was right; this entry is the correction. The rule now. The condition share is fixed at 60% for every senator — the same trend/condition blend the governor board uses for the same state outcomes, pinned equal by a test so the two boards cannot drift apart — and the attribution ramp is gone: attribution is the block’s 25% weight cap, stated once, identical for everyone. Tenure still bounds the data — the trend is fitted only inside a senator’s own Senate service, and short windows carry less weight — never the rule. The effect, reported not tuned. 72 of 91 ranked senators moved (20 by five or more places); the House is untouched (its state block is reserved). Biggest movers: Tina Smith rose 40 → 25, James Lankford fell 28 → 42. The D–R gap in mean senator grade moved from +0.19σ to +0.21σ. The senator the review used as its exhibit, Mike Rounds, stays No. 2 — the formula was corrected because it was indefensible, not because of where it put anyone. The mean measured share of a senator’s grade rises from 58% to 60%, because the removed ramp had been discounting real data for junior senators. Also in v2.6 — one rulebook for the shutdown year. Through v2.5, the 2025 unemployment annual average — an eleven-month figure, because October 2025 was never collected during the federal shutdown — was scored for governors with a caveat but excluded from city scoring: same statistic, two rulebooks. Cities now follow the governor rule: the value enters the series, flagged with the verbatim BLS footnote wherever it appears. The effect: city series extend to 2025, four cities whose windows had been too short to rank — Fort Worth, Boston, Milwaukee and Atlanta — join the ranked board (17 → 21 of 21 scored), and 14 of the 17 previously ranked cities moved (largest: Dallas rose 8 → 4, Baltimore fell 12 → 19). |
| v2.5 · coverage update · August 24, 2026 | Public Law 119-21 — the 2025 reconciliation law, popularly the “One Big Beautiful Bill Act” — enters the tracked bills, the vote records, and the fiscal block. No rule changed; the law passed the standing numeric entry test the day it was signed (July 4, 2025) and joins thirteen months late because this site’s hand-verified extraction takes real labor per law — a coverage gap the methodology and bills pages now disclose in writing rather than leave implied. The extraction: 92 line items, every quote verbatim from the enrolled text and adversarially re-audited; CBO’s estimate is quoted on three bases (+$3.4T vs. the January 2025 baseline, the figure the grade uses like every other law; −$366B vs. the Senate’s current-policy baseline; $4.1T with debt service), because quoting one CBO number without its basis is how a true number misleads. The effect, reported not tuned. The fiscal block now covers seven scored laws, dollar-dominated by the largest two — one from each party’s trifecta ($1.8T American Rescue Plan, no Republican Yeas; $3.4T P.L. 119-21, no Democratic Yeas). The party means of the fiscal score flipped: Democrats moved from −0.74σ to +0.52σ and Republicans from +0.76σ to −0.55σ, the mirror image of what the changelog reported when the block launched with the rescue act dominant. 445 of 449 ranked members moved (309 by ten or more places), and the D–R gap in mean member grade moved from −0.06σ to +0.19σ in the Senate and −0.12σ to +0.08σ in the House. Same rule, both directions, published either way — at this coverage the block measures the 2021–2025 fiscal era, and that limitation is stated on every page that renders it. |
| v2.5 · August 2026 | Fiscal conduct enters the legislator grade at 10%. A member’s fiscal record is the deficit footprint of the laws they voted for: each tracked law carrying a verified CBO estimate contributes its signed amount to a Yea voter’s footprint — a Nay adds nothing — and lower footprints rank higher, Blom rank-normalized within chamber. This scores conduct, deliberately not the national debt, which rose under every member of the era and belongs to no one of 535 votes. Weights move from 30/35/25/10 to 25/30/10/25/10 (state outcomes / bills / fiscal / promises / attendance); the two reserved blocks are untouched. Limitations, stated plainly. Coverage at cutover is six scored laws, dollar-dominated by the largest (the $1.8 trillion American Rescue Plan), so this measures the pandemic-era fiscal record more than a career. Members without recorded votes on at least three scored laws are reserved, not zero-scored — 371 of 530 members are measured at cutover. Two of the six laws score as deficit decreases, so the rule credits deficit-cutting votes wherever they occurred. Who moved, and the symmetry number. At cutover, 207 of 358 ranked representatives and 12 of 91 ranked senators moved ten or more places. The rule is identical for every member, but the outcome is not neutral at this coverage: the D–R gap in mean member grade moved from +0.02σ (indistinguishable from zero) to −0.11σ, driven mostly by the rescue-act vote. That gap is reported here rather than tuned away — adjusting a symmetric rule to force a symmetric outcome would be putting a thumb on the scale in the other direction — and it will move as coverage broadens beyond the pandemic-era laws. |
| v2.4 · July 2026 | Where a state stands is now the larger half — 60/40, the same for every governor. v2.3 introduced standing at up to 35% on a tenure ramp. Two things were wrong. It was too small: Oklahoma stands below the national average on 11 of its 16 scored categories — 19.6 teen births per 1,000 against a 12.98 average, 14.1% of adults uninsured against 10.8%, 4th-grade reading and 8th-grade math both trailing — and still ranked No. 2, because it is cheap and improving. A board that calls that second-best in the country fails the reader. And the ramp scored a ten-year incumbent (35% standing) and a two-year governor (0%) with different formulas, then printed them in one ranked list. Oklahoma now ranks 11th. The trade-off, stated plainly: this board now ranks state outcomes with the sitting governor named, not a governor’s personal performance — a governor two years in inherited nearly all of their state’s standing. Nobody is scored before two years in office, movement still carries 40%, and both halves are published separately on every card. Party symmetry re-checked after the change and holds: governors D–R gap +0.039σ, p = 0.69. |
| v2.3 · July 2026 | Where a state stands now counts — this one changed every governor score, and some ranks. Through v2.2 a category score was pure trend: the state’s slope minus the peer median slope. Level was excluded on purpose so nobody was punished for what they inherited. That principle is kept — but alone it produced a result we could not defend: being genuinely good earned nothing. On the v2.2 data the correlation between a state’s obesity level and its obesity score was r = −0.06, statistically zero. Vermont, the second-leanest state in the country at 29.0%, scored −0.56; Alabama at 38.9% scored +1.33. Each category score is now (1−φ) × trend + φ × level, with level measured at the end of the scoring window and φ ramping 0 → 0.35 with tenure (nothing for the first two years). Trend stays the majority partner at 65% or more, so inheriting a mess still costs almost nothing and improving it is still the fastest way to score — but holding the best outcome in America is no longer worth zero. Prompted by a reader question we could not answer honestly under the old rule. |
| v2.2 · July 2026 | Additive context; no grade changed. Every president’s and member’s page now shows total federal spending by fiscal year during their time in office (OMB Historical Tables), labelled plainly as the whole federal budget of those years — not money they personally directed. We also built a presidential grade and then chose not to publish it: on running it, the only honest whole-country baseline turned out to be dominated by who inherited a crisis, not governance, so any ranking would measure timing and read as partisan. Instead each president gets the data without a verdict — how each national outcome moved during the term, with OECD context, and no overall score. Plus a “jump to a state” picker on the Congress page. |
| v2.1 · July 2026 | Two additions to legislator accountability; governor rules unchanged. First, a blended legislator grade: one grade combining the bills a member moves (Legislative Effectiveness Score, rank-normalized within chamber and majority status), how the region they serve has fared on their watch (senators, capped by Senate tenure), attendance, and promise-alignment. Parts we can’t yet measure are reserved at zero weight, never redistributed, and a “measured” figure says how much of the grade is real. Second, bill-spending vote pies: each member’s recorded votes on a rule-selected set of major money bills are joined to spending line-items pulled verbatim from the enacted bill text, showing what they voted to fund and to reject, every dollar cited to its section. Bills enter by a mechanical rule (enacted or floor-voted, ≥$1B, verifiable), never by hand; only appropriations feed a pie — authorizations, guarantee ceilings and pay-fors are disclosed but never summed in. |
| v2.0 · July 2026 | A core rule changed, so the major version moved. v1.0 gave Performance Scores only to executives and showed legislators’ outcomes as unscored context. That blanket exemption is replaced: an official is scored on the constituency they actually answer to, ranked only against the same office. Senators now get the same state-outcome card and OVR as governors, ranked among senators, over their current term — with influence-sharing disclosed and the state’s data-transparency penalty left with the governor, who actually controls those collections. House members are still not handed state numbers (they represent a district, not a state); district scoring awaits district data. All other rules — trend not level, lags, peer-differencing, the ±2.5σ cap, citizen weights, thin-data flags, human review for promise ratings — are unchanged. |
| v1.2 · July 2026 | Additive. Tax burden now measured as a share of personal income (Census ÷ BEA), replacing dollars-per-resident — the per-capita metric is retired to context. New social-services domain: SNAP food-aid access (USDA) and unemployment-insurance access (DOL). The overall rating can now be weighted by citizen priorities (Pew Jan. 2024 survey, cited on the methodology), published beside the equal-weight version. New data-transparency penalty for states that skip data collections they run. Directions, lags, weights and the penalty were fixed before any v1.2 score was computed. |
| v1.1 · July 2026 | Basket expansion, additive only: 13 verified categories live at v1.1 — unemployment, carried from v1.0, plus NAEP 8th-grade math, median air-quality index, drinking-water violations, adults without health coverage, teen births, poor mental-health days, homicide, state taxes per resident, residential electricity price, 4th-graders below basic reading, adult obesity and firearm deaths, each shipping only after two-stage verification — scored under the unchanged trend-vs-peers rules; overall rating (OVR) gates at 3+ scored categories. School-shooting casualties load as unscored context. Directions and lags fixed at registration, before scoring. The amendment text is in the published rulebook, dated. |
| v1.0 · July 2026 | Initial public release — Performance Score, Economy domain, one preregistered metric (unemployment rate, BLS LAUS). Full rulebook published verbatim at /methodology/v1.0/. |
Scores last recomputed: Sept. 3, 2026 — that is a recompute under the same rules, not a source refresh. Each source publishes on its own schedule, and every category carries the year its data describes and the state of its last source check — including when that check is failing and older values are being served. Oldest live series: Food-aid access, through 2023. Recomputes and roster updates are not logged as changes; new immutable score rows are written each run, so history is never overwritten.
Site & policy changes
| August 10, 2026 | Exogenous events are now marked on every time chart and named in every rundown. Recessions as dated by the NBER, the COVID-19 emergency on the WHO’s own declaration and end dates, and hurricanes NOAA records at $100 billion or more (Katrina, Harvey, Maria, Ian) appear as shaded bands on every chart whose years they touch, and each rundown names the ones inside its measured window — because 2020 unemployment belongs to the pandemic and 2009 revenue to the Great Recession, and a chart that draws the spike without naming the event invites blaming it on whoever held office. The list is mechanical, not editorial: three criteria, three dating authorities, and policy acts are never listed — a law is an officeholder’s own doing, not an exogenous shock. The same marks and the same note for every officeholder of both parties. |
| August 9, 2026 | President pages: the debt on their watch, drawn — and the promises beyond it. Every president’s page now pins the federal debt at the last fiscal year closed before their inauguration and at the last one closed on their watch (Treasury figures, daily for the sitting president), with the arc drawn between them — zero-based axis, nominal dollars, the fiscal-year boundary disclosed. Below the terms, a new section states what sits beyond the borrowing number: the audited balance sheet’s federal employee and veteran benefits payable, and the 75-year social-insurance shortfall — each figure on its own stated basis, never summed. The economy outcomes are also now drawn year by year, because endpoints alone can hide the shape of a term, and the card gained a federal receipts (% of GDP) row — the slice a president’s own law can actually move, from OMB’s Historical Tables, cross-checked against Treasury dollar receipts. Like the combined tax ratio it is reported, not judged: a tax level has no better direction. And the rundowns — presidents’ and governors’ — now name everything that worsened on the officeholder’s watch, alongside what improved, judged by each metric’s own printed endpoints. The same beats, the same rules, for every officeholder of both parties. |
| August 9, 2026 | Member pages moved to readable addresses. /congress/m001169/ became /congress/ct-murphy/ — state, then surname, for every member of Congress. Every old address redirects permanently to the new one, so links shared before today keep working. Where two members of one state share a surname, the longer-serving member holds the short form and the other carries their first name — and an address, once assigned, is never renamed or given to a different person, even after a member leaves office. |
| August 2, 2026 | This site started counting its readers, and published a page saying exactly what it counts. Until today nothing measured traffic at all — there was no way to tell “nobody read it” from “nobody could find it,” which are opposite problems. Netlify’s server-side Web Analytics is now on. It reads the host’s own request logs: no code was added to any page, no cookie is set, and no third party is involved. It counts pages served, and tells unique visitors apart by IP address within a single day. The full description — including the part that is not nothing — is on the new privacy page, which is linked from the footer of every page on the site. |
Corrections
| August 31, 2026 | A second adversarial review — 33 reviewers across the whole site — found one number that was one-sided, one badge that could not fire, and a promise made on 31 pages that this site has never kept. The substantive fixes: (1) Senator pages reported that the Senate rejected funding votes on the shutdown bill 14 times. True — and the same Senate vote file shows the competing continuing resolution rejected 7 times in the same weeks on the same question. Two bills, each blocked by the opposite side, and this site counted one. Both counts now publish, derived by one rule. (2) The homepage map’s thin-data dot required every category to be thin, so New York sat 3rd with 9 of its 12 categories resting on fewer than three data points and no mark at all; eleven states were majority-thin and unmarked. The dot now follows the majority and the tooltip prints the count. (3) Every bill page promised “three layers” of state impact and then told the reader the middle one was missing — it has never shipped in any build. The count is now derived from what actually staged. (4) On CHIPS, “every tracked dollar carries its year” came from rounding 99.96% up to 100; “every” now requires exactly 100%. (5) 71 district pages ranked “the widest gap” among a set of one; they now count instead of ranking. And the words. A baseline — the term that decides whether the same law costs $3.4 trillion or saves $366 billion — went undefined on 523 member pages; it is now defined once, above the CBO figures, without touching CBO’s own wording. Division is defined on first use across the UK edition, and the shutdown’s procedural questions now carry a plain-English line beneath the chamber’s own. Rubric was doing three different jobs: the weighted scoring scheme (kept — it is the right word, and it is defined where a reader first meets it), an unwritten per-metric rule (now “scoring rule in development”), and a plain metric set on the presidential pages (now “the same measures”). No score or rank changed in any of this. |
| August 25, 2026 | A 48-reviewer audit of every page type found and fixed seven claims that were flatly wrong, alongside dozens of gaps. The wrong claims, owned plainly: (1) the two House party leaders’ sponsored-bill counts included procedural number reservations — the Speaker’s “8 bills” was 7 placeholders and one resolution; both counts are corrected at the source. (2) City and state rundowns said “Below the average” where the level was above it on lower-is-better metrics; the standing phrase now follows the metric’s own direction. (3) Chamber-switching senators’ vote tables cited Senate roll calls for votes they cast in the House; those rows now cite the record that actually contains them. (4) The archived v1.0 rulebook titled itself v2.6; its title is now derived from its own text. (5) The homepage described the score by a retired formula and an internal project codename; both corrected. (6) Senator pages said “no vote of theirs is in this grade” — false since v2.5’s fiscal block. (7) The retracted preregistration claim survived on three overlooked surfaces; all now carry the corrected form. Every fix is display or copy; no score or rank changed. |
| August 24, 2026 | The Senate roster was 41 days stale, and for all of them this site reported South Carolina’s seat as vacant while a sitting senator served. Sen. Darline Graham was sworn in July 14, 2026 — the very day the roster was last fetched, before the upstream dataset had caught up — and no re-fetch happened until this correction. Her page now exists (unranked, honestly: nothing in the rubric is measurable for a member seated weeks ago), every senator count reads 100, and the pipeline now refuses to export on a roster more than 30 days old, warning at seven. A staleness the site disclosed (“roster as of”) but should never have allowed. Two claims were also corrected the same day: the methodology deck said every rule “was fixed before any politician was scored,” which this changelog itself refutes (v2.3 and v2.4 changed published scores) — it now says what the site actually keeps: rules published before they are applied, changed only in public. And the Congress index still described the retired v2.4 weights a month after v2.5 shipped; the percentages there are now read from the export rather than typed, so they cannot drift again. |
| August 2, 2026 | An editorial audit of the generated sentences found wording that contradicted the numbers printed beside it. No score, rank or underlying number changed — the words did. The rundown’s movement verdicts were computed against peers but worded absolutely, so on 18 of 96 checked movement beats the verdict said the opposite of its own figures: Texas read “moving the right way” on obesity that rose 33% → 35.61%, and Mississippi read “losing ground” on a homicide rate that fell 23.7 → 19.7. Verdict words now describe what the two printed endpoints did; the peer comparison stays, printed and checkable, in the clause that follows. Four category units were also readable as their own opposite — “Health coverage (% of adults 18-64)” is the uninsured share, and a falling number was praised in a way that parsed as coverage collapsing. Units now name their numerator (“% of adults 18-64 uninsured”, “% of 4th-graders below basic”, “% of systems with a violation”, “mean poor-mental-health days of 30”), and the category “Crime” — which carried a homicide-only metric — is renamed Homicide so the label claims no more than it measures. On 23 state pages the headline asserted “a middle-of-the-pack card with two clear outliers” — a hard-coded string that fired identically on ranks 4 through 26, with a count the template never computed. It now states a counted fact: “below the national average on N of its M scored categories.” And on member pages, the rundown said scores were benchmarked “against a chamber average of exactly 1.00” while the rank is computed within chamber and majority — as the methodology page has always said. The sentence now matches the methodology, a redundant multiplier that printed “0.0× the typical member’s” on 14 pages is gone, and pages for the 118th Congress’s leaders and committee chairs now say that the score counts only bills a member sponsors themselves — which is why leaders who move other members’ bills score low on it. |
| July 27, 2026 | The senate board was wrong, and it was wrong in a way that penalised long service. 59 of the 91 ranked senators moved ten places or more. A senator is scored partly on how their state is doing. That figure is meant to be a standard score — the state’s position measured against the average of all states, so half come out above the line and half below. The code divided by the spread but never subtracted the average. The result was not a score at all: across all 84 senators it averaged −1.66 with a spread of 0.07, and not one state came out above average. The best in the country, Massachusetts, read −1.49. The same figure computed for governors, which was correct, averaged −0.04 with a spread of 0.93. A near-constant is harmless until you multiply it by something. This one was multiplied by a factor that grows with years served — the longer someone had been in the Senate, the more of their grade came from a number that was the same for everyone and always negative. It had become a seniority penalty wearing a statistic’s name. A first-term senator lost nothing; a twenty-four-year one lost about 0.66 on that portion of their grade. This is a correction, not a methodology change. Nothing about what the site measures or how it weights anything has changed. The old numbers were arithmetically wrong and the new ones are right. The corrected figure averages +0.01 with a spread of 0.39, and 43 of 84 states now sit above average — which is what “average” means. Who moved. Largest gains: John Thune (R) +44, Jack Reed (D) +41, Chuck Grassley (R) +41, Patty Murray (D) +38, Susan Collins (R) +32. Largest falls: Adam Schiff (D) −24, Jon Ossoff (D) −23, Eric Schmitt (R) −22, John Kennedy (R) −22. Of the 44 senators who rose, 23 are Democrats, 19 Republicans and 2 independents; of the 46 who fell, 20 are Democrats and 26 Republicans. The error was not partisan and neither is the fix — it tracked how long someone had been in office, nothing else. How it was missed, and what now stops it. Every automated check this site had pointed at the inputs — the raw government numbers going in. Nothing read a published grade or rank and asked whether it was plausible. Two checks now do, and both run before anything can publish: one asserts that any figure called a standard score actually behaves like one, and one refuses to publish a board where every single unit is below average. Either would have caught this on the day it shipped. |
Think something on this site is wrong? See disputes & corrections on the About page.