State of the Platform

ev-accounts v2.22

Auth, tiers, XP and gems, politician database, TIGER geofencing, elections. This cycle it became the one engine. Civic Trivia and Validation Quests folded their backends in, and the internal HTTP calls the platform used to make to itself — award XP, award gems, adjust a rating, ingest a consensus — became in-process calls, with the retired service keys deleted rather than left lying about. Security moved with it: the HS256 path is gone, so a stray environment variable can no longer switch off signature verification; secrets are no longer derived from Supabase API keys; and a voter’s Compass answers are now scoped to the caller and gated by their own visibility setting. A second piece of housekeeping earns its place here: concurrent sessions now reserve migration numbers from an allocator, because two collisions in six days is what counting by hand costs. Standing from earlier cycles: the ballot a voter actually receives, seeded from each Secretary of State’s own list, which exposed federal races carrying only the major-party field — 16 U.S. House candidates where the state listed 38; a surname-only dedupe key that shows every match to a human; Puerto Rico end to end, parties stored verbatim in Spanish because PNP and PPD each contain both US national parties; and three corrections kept on the record — five North Carolina members described as primary sponsors when the state’s own page had folded cosponsors into that line, eight Washington canvass rows left open inside a recount margin, and the occupancy-evidence view that decides “seated official or discovered name?” structurally rather than by provenance. 8,335 districts mapped, 121,178 legislative votes and 19,622 bills on record.

On the Record

Speaker-attributed, searchable transcripts of public meetings, synced to the original video. Live in the pipeline: 154 meetings, 16,852 transcript segments, 942 identified speakers, including LA City Council with 7,200 recorded council votes. Built by volunteer Chris Andrews; it feeds the platform's meetings system. It was the fastest-growing dataset on the platform through July and has barely moved for five cycles while the stance work had the floor — one meeting last cycle, and none at all this one. The segment count fell again — 17,077 to 16,852, a second consecutive drop with no new ingest, after 583 last cycle — and two speakers went with it. A reprocessing remains the most likely explanation and it is still unconfirmed. Printed unexplained rather than smoothed over, and it is now the oldest open question on this page. The Fallacy Finders hand-off is no longer a plan: a written transcript contract, and a validator that fails loudly on a bad one.

Sourced stances, by where they serve

35,877 stance positions across 4,078 politicians and 63 topics. Every stance carries reasoning and cited sources — checked today, 0 of them have an empty source list. 1,665 of the 2026 candidates already have researched stances. 2,722 sources were independently re-checked in an April sweep, and that figure has not moved since: every row in it was written in April 2026. It is a dated achievement, not a running total, and this page had been printing it beside live counts as though it were one. The checking since has taken a different form — the source-supports-claim audit below, which retires what it cannot evidence rather than certifying what it can.

Why the total moves in both directions. A source-supports-claim check re-reads every cited source and asks whether it supports the stance we wrote. It swept the map — Utah fell from 1,358 stances to 531 — and closed a fabricated-source retirement inside it: 298 stances retired, 160 URLs denylisted, one cluster sourced to a newspaper that does not exist. Four cycles of falling totals were that check doing its job. This cycle the total rose 2,839, and every named row above rose with it, because retirement was never the end state: the re-research worklist feeds those seats back through sourcing, and the Senate and Miami-Dade waves added new ones. A number that only ever goes up is not measuring anything. The Season 2 cutover has a cost worth naming: 1,678 of these answers — about 5% — sit on topics the open season no longer asks, so they are history a voter can read rather than a position a voter can compare against today.

How a stance gets seated. A member is placed on a topic because of a bill they sponsored, read in its session-law text — not a sentence someone else wrote about them. Three rules travel with every wave: a unanimous bill cannot evidence a position; cite the session law, not the introduced bill, because four North Carolina bills had been gutted and replaced by a committee substitute long after members filed them, and a sponsor list never changes; and where a mechanism matches no chair clause by clause the answer is blank. A blank is only as strong as the search behind it: Seattle’s ten “researched blanks” rested on a single source class, and re-research seated 48 rows across 11 of 11 offices.

Two corrections to this table, both kept. Per-state counts once walked a single database link and missed everyone seeded after occupancy moved onto dated office_terms: Wisconsin read 5 researched officials where it holds 208. Reading all three links fixed that and over-corrected in the same motion, sweeping in names from campaign-finance filings that sit on placeholder offices with no geography; migration 1702 settles it — a real seat carries geography, a placeholder carries none. The durable lesson: the sum check this table relies on could not see either error, because a misattribution moves a row between buckets without changing the total. What does catch it is agreement across cycles: rebuilt today, thirteen of the fourteen rows reproduced their politician counts exactly, and the ten new seats landed where the work happened — eight in Florida and two in Georgia, both inside the remainder row.

Budgets you can follow

Treasury Tracker's database now holds budget data for 7,507 government entities across all 50 states plus DC, spanning fiscal years 1976 to 2028:

• 2,789 cities, 624 counties, 50 states, 38 towns, and the federal budget
• 2.6 million budget line items and 1.3 million individual transactions
• 282,457 public salary records
• State general funds upgraded to audited ACFR sources across all 50 states, each tie-verified to the penny; an independent blind re-derivation confirmed 149 of 151 exact matches, with the two exceptions documented
• Los Angeles’ severed history is repaired — 22 continuous years (FY2003–2024) of all-funds actuals from one source; the defect was a label, not a number. FY2025 and FY2026 were withdrawn rather than classified, and an earlier verdict on the same data was retracted, because the tie that appeared to confirm it had been fitted rather than found
• 231,754 of 270,227 rows carry an audit grade — read from an audited report (798), compiled from one (19,590), or self-reported by the government that spends the money (211,366). 38,473 still read unknown, which means nobody has checked yet. The grade is not a score: most statewide files are self-reported by design, and saying so on the row is the point
• A fiscal year is a factual claim, and 36,311 rows were making one nobody had checked. One line of SQL wrote a literal July start onto every row it created, whatever the entity’s real calendar. No dollar figure moved, which is exactly why nothing downstream noticed: the earlier gate validated conformity to the hardcode rather than correctness
• Four statewide sweeps landed in eleven days — Pennsylvania’s 2,619 approved governments, Michigan’s 1,493 townships and villages, Florida’s 479 filers, and Indiana’s counties read one audited ACFR at a time. Every one of them arrived with a refusal list: Florida’s 14 entities whose publisher contradicts itself, Michigan’s 58 declared unloadable entity-years, and a legacy Indiana vintage deleted outright after a township was found publishing a school corporation’s figures
• Every figure durably sourced. The core promise: no best-guess data wearing a real-looking label

Does the tool get used as a tool?

Everyone who arrived, against the people who did the one thing that app exists for. Bars are the 90-day rate; counts are people, 30 days and the full quarter.

Two of these are start problems, not finish problems. Only 14 people started a trivia game in a quarter and 9 finished — and that is on 106 arrivals, the most Civic Trivia has had. Read & Rank is the same shape: 23 races started, 10 ballots revealed. The Compass is the opposite — 111 of its 129 visitors opened the quiz and 36 saw it through, so six in seven start and better than a quarter finish.

How many arrive

Unique visitors per app, 30 days and the full quarter. Bars scale to the 90-day count. Each app is counted once across every hostname that serves it — Treasury draws traffic on three, the main site on three.

Treasury Tracker is still the second most-visited thing we run and holds the platform’s largest dataset. Civic Trivia moved most this quarter, 87 to 106. Whether Treasury’s 159 reached a budget is the question the note below is about — and it now has a partial answer.