How we count

Window. Meetings listed from October 24, 2025 through October 23, 2026, transcribed through October 9, 2026, the newest meeting we have a transcript of: talk time, issues, quotes and every count of transcribed meetings stop there. A count that includes meetings not yet transcribed says "listed", and a topic chart from the clerk's agenda times, which cover those meetings too, says how far it runs.

Meetings and hours. A meeting is one body on one day. A second video that day (part 2, a recess) adds to that meeting's hours, not to the meeting count. If the same video is listed under two bodies, we count it once. Hours are the length of each meeting's video; a body's hours show only when we know the length of at least 8 in 10 of its meetings. Videos under 10 minutes or over 12 hours are left out of the hours as likely errors.

Margins and ties. We resample whole meetings many times to get a 90% range for each share. When two ranges overlap, they share a rank (shown as "=").

Who does the talking. A body is ranked only when at least 60% of its members' speech is identified; under 80% the ranking is marked preliminary. The rest of the members' speech has no name, so each member's minutes run from what we named (low) to what we named plus an even share of each meeting's unnamed speech among the members seated that day (high), and a second high end scales the speech we found only in transcript text by how often that text read finds the member. The rank range covers the named low end and both high ends. A member is compared only when at least 30% of the speech they may have given, by the higher of the two, carries their name. Ranks are ranges from 2,000 resamples of the body's meetings; overlapping ranges share ranks. A member could rank first unless another member is clearly ahead of them (a one-sided test at 90% confidence, adjusted for the number of members). Presiding officers are shown apart. A line's time is capped at 1.5 seconds a word plus 10 seconds, so a recess folded into one line never counts as speech.

Quotes. Officials only, never members of the public. Each quote has passed our resident-name check and plays at a second we've checked.

Topics on a person’s page. Two AI readers read each transcript and agree on the topic of each stretch of talk; a stretch they don’t agree on has no topic. A person’s share is of their speech with a checked topic, not of everything they said, and their page says how many of their minutes that is.

Data dates. The site has one data date, the newest meeting transcribed in any city (Oct 9, 2026). Each city’s pages run to its newest transcribed meeting (Chicago: Oct 9, 2026). Three kinds of line run later and say so: counts and charts of listed meetings, transcribed or not (“listed”); topic-time charts from the clerk’s agenda times, which cover meetings not yet transcribed; and the date we last checked city calendars for meetings coming up. A figure can move as more meetings are transcribed, so cite the page with its data date and data version (“Cite this page” at the bottom). CSV files on the Data & downloads page, each with a codebook, hold Chicago’s topic time, talk time, people, bodies, accuracy ledger, meetings, officials' talk time, and the five-city findings and signatures. Other numbers, such as the story cards, link to their sources.

How accurate the names are. One blind check, pooled across all five cities, not this city alone. Each counted second of talk time was equally likely to be drawn; Claude Opus and Gemini each read the lines around it, and a line counts as wrong when the name doesn’t match who is speaking. For councils and their committees, of 410 lines drawn, 352 were right, 0 wrong and 1 unclear, and the checkers couldn’t judge 57. We have since removed the voices behind 110 of those lines (0 wrong, 1 unclear, the rest we had already held for other reasons). Of the 300 lines we still publish, the checkers could judge 268: all 268 right (90% lower bound 0.99). They couldn’t judge the other 32. For commissions, of 340 lines drawn, 260 were right, 4 wrong and 1 unclear, and the checkers couldn’t judge 75. We have since removed the voices behind 64 of those lines (3 wrong, 0 unclear, the rest we had already held for other reasons). Of the 276 lines we still publish, the checkers could judge 228: 226 right, 1 wrong and 1 unclear (90% lower bound 0.97). They couldn’t judge the other 48. For school boards, of 200 lines drawn, 170 were right, 1 wrong and 2 unclear, and the checkers couldn’t judge 27. We have since removed the voices behind 15 of those lines (0 wrong, 1 unclear, the rest we had already held for other reasons). Of the 185 lines we still publish, the checkers could judge 162: 160 right, 1 wrong and 1 unclear (90% lower bound 0.96). They couldn’t judge the other 23.

Who talks most. A member could rank first in talk time when their 90% rank range includes 1st, and is clearly first when no one else’s does. A clearly-first member’s share is of the identified talk time of the members ranked on that body (presiding officer apart); a body where several could rank first is too close to call and gets no single share. A body is ranked when 60% of members’ talk time is identified (preliminary under 80%) and 4 or more members can be compared. To know how much members really speak, two AI readers label a random sample of seconds from each body’s meetings without seeing our names, and we count only the seconds both agree on; the share we print is the low end of that sample’s 90% range, so it is a floor.