Method

How we know

We cover 1,775 school boards, the ones whose meetings we found on video, out of the nation's school districts. We don't have every meeting, even for those boards; each board page says which ones we have.

The films: The films on the front page show where an issue came up at school-board meetings, month by month: phones in school, AI in schools, chronic absenteeism, and vouchers and school choice. Each dot is a district at its first meeting that discussed the issue by the issue rule below, on a checked date, with a page on this site; it opens that meeting's video a few seconds before the issue is first named in the busiest stretch of talk about it, and we check that the words there name it. A meeting on a video host we can't open at a second gets no dot. The clock runs in date order, and every count on screen is as of the date shown. The films make no claim that an issue came up more or less often over time. Our months span one school year, and its two halves differ by the school calendar (phone rules are set in summer, budgets in spring) and by which boards we had read, so a change between them can't be told apart from the season. A trend comes back only as the same months a year apart, on the same boards. A dot is bright only where we had read 3 or more of that board's earlier meetings over 2 or more months; otherwise it is faint, because our reading there began too recently to call it new. The quotes are board members' clips (see Clips), cut at sentence ends, word for word.
Your board's year: A short film for each of the boards with the most dated meetings on our record (6 or more in the last 12 months). It names no board member. The decision is our pick from the board's checked meeting summaries: a line that reports something the board did, ranked by what it is about (the superintendent, closing schools, the budget or a tax rate, the teachers' contract, then other policy) and a split vote; when no such line passes, the film says “a decision this year”. Two AI readers (Claude Sonnet and Google Gemini), each alone, read the meeting's current transcript from that second and both had to confirm that the board itself acted, the outcome, every fact in the line, and that the video starts on this item; a third (Claude Opus) read the same stretch blind and had to agree. The vote record behind agenda items is never used: it stores the item, not the motion. Meetings and hours share one base: the dated board meetings in the last 12 months, with the videos' own lengths (“at least” when a length is missing). The busiest night at the mic counts members of the public, never names, shows or plays them; it says “the busiest we counted” only when its count beats the next by more than 2, since each count is checked to within one. Where the time went is the board's checked topic time, a floor, beside the typical board's floor; “more than most” only when this board's low end is above the typical board's high end.
Superintendent actions: The board's own actions on its superintendent since Aug. 1: hires, interim appointments, contracts, evaluations and searches. Each comes from a checked meeting summary and passed the same three readers on the meeting's current transcript, who also had to agree on what kind of action it was. We print the action and the date in fixed words, and the tally only when the readers heard the same one. No reasons, names or allegations. Lines the site's privacy rule holds stay off.
Meetings: We count one meeting per board per day; a second video that day joins it. The list comes from meeting videos the districts post, and at San Francisco Unified, Oakland Unified, Los Angeles Unified and Chicago from the district clerk's own archive. A model check confirmed each video is this district's board before it counts.
Dates: A date prints once it matches a second source: the video's listing, the date spoken at the meeting, or the clerk's calendar. Until then the meeting says “Date not confirmed yet”.
Meeting length: Average and typical lengths use each meeting's video, from first second to last. We leave out videos under 10 minutes or over 12 hours, and any file that may be cut short. The longest meetings use time spoken, first word to last, and drop any video where the speech stops for half an hour.
Who is speaking: We name a speaker only when our voice check is sure: two AI models agree on who it is, and a blind check of their names came back right at least 95% of the time. “Heard at” counts meetings where that check named a member. It is not attendance, which comes from the roll call.
Who talks: We time how long each board member speaks at full board meetings, using only the speech our voice check could name. Some speech can't be named: two people share a microphone, or nobody says a name. Each board prints its coverage: the share of all its members' speaking time that we could name, with its 90% range (“about 92% (66%⁠–⁠100%)”). The whole is every second a member spoke, measured by reading random seconds of entire meetings blind, so speech we couldn't tie to anyone still counts against us. Coverage is never higher than the named share of the speech we could tie to a member, and the range stops there too. Where a board hasn't been measured on whole meetings yet, it says so instead of printing a number. Minutes are shown as floors (“at least”), and rank ranges may widen. Boards where we name at least 60% are ranked; under 80% they carry an “early read” tag because their ranges are wider. We don't throw a board out for being incomplete. We show how much the gap could move each member, and the ranks carry that.
Minutes per meeting, with bounds: Each member gets two numbers. The low one is the minutes we named, averaged over the full board meetings held while they were seated; a meeting they missed or stayed silent in counts as zero. The high one adds their even share of each meeting's speech we couldn't name, split among the members seated that day. The truth for most members sits between the two. A member is ranked only when at least 30% of that high number is named, and after 8 board meetings seated. When a roster gives no start date, their seat starts two meetings before we first heard them.
Ranks are ranges: We replay the board's year 2,000 times, drawing its meetings at random with replacement, the same draw for everyone, and rank the members on both the low and the high numbers. A rank like “2nd–5th” means that across all of that, this member was never clearly above more than one colleague or clearly below more than two. We call one member above another only when the gap holds on both numbers at a margin corrected for how many comparisons we make. A single rank (“1st”) means the data can place them exactly. The presiding officer is shown but not ranked: much of their time is running the meeting (we find them from the roster title or from who calls the votes). The lead sentence says a member spoke “the most” only when their range is first alone.
Share of the talk: Boards meet for different lengths, so we never compare minutes across boards. Each member's share is their minutes divided by all the ranked members' minutes, presiding officer apart. Its range runs from their low number against everyone else's high, to their high number against everyone else's low, across the replayed meetings. The national chart takes the member who talks most on each board. Even talk on a board of 4 is 25% each; on a board of 10, 10%.
Issues: Each issue has one rule, named on its page. Seven issues are counted on the whole transcript: two AI readers (Claude Sonnet and Google Gemini) each read every meeting in full, blind to each other, and a meeting counts only when both say the board discussed the issue, not just mentioned it; each counted meeting opens at the line where that discussion starts. Vouchers and school choice, and school closures, are counted on a stricter word rule: 3 or more remarks with the issue's words, from 2 or more speakers, confirmed by both readers; those meetings open where the words are first heard. For every issue, a third model (Claude Opus) reads a random sample of counted meetings blind, and an issue goes up only when that check is right at least 95% of the time. Shares count meetings the way the meeting lists do, one per board per day: a day filmed in two videos counts once, whichever video took the issue up. A share's range is the 90% range of our count when whole districts are resampled, so one district with many meetings can't narrow it on its own. We don't rank issues or states against each other, because we catch a different share of the meetings that discuss each.
What our issue counts miss: Our counts miss some meetings that discuss an issue, so every share is a floor and we write “at least”. To measure how many, the same readers read a random sample of the meetings we didn't count, whole and blind; a meeting counts as missed only when both say the board discussed the issue. Each issue's catch rate, with its 90% range, is below and on its page. A “1 in N” always uses the low end of the share's range, so it is a floor too. We catch about 6 in 10 meetings that discuss reading and test scores (90% range 54⁠–⁠73%). We catch about 7 in 10 meetings that discuss budget cuts and layoffs (90% range 59⁠–⁠82%). We catch about 8 in 10 meetings that discuss AI in schools (90% range 63⁠–⁠93%). We catch about 5 in 10 meetings that discuss safety, weapons and police (90% range 42⁠–⁠64%). We catch about 7 in 10 meetings that discuss chronic absenteeism (90% range 46⁠–⁠83%). We catch about 5 in 10 meetings that discuss vouchers and school choice (90% range 44⁠–⁠59%). We catch about 9 in 10 meetings that discuss phones in school (90% range 62⁠–⁠96%). We catch about 7 in 10 meetings that discuss enrollment (90% range 53⁠–⁠82%). We catch about 4 in 10 meetings that discuss school closures (90% range 34⁠–⁠47%).
Clips: A clip is a board member's own words, from where they began speaking to the end of a sentence. It comes from an official on the board's roster, names no private person and sits away from public comment. Members of the public and students are never shown.
Public speakers: We count members of the public who spoke during public comment: parents, students, employees speaking for themselves or their union, residents and groups. Board members, district staff, officials of other bodies, invited presenters, people being honored and written comments read aloud aren't counted. Two AI readers counted each meeting and had to agree; a third model counted a sample from the whole transcript, and the counts had to fall within one speaker. People are counted, never named or quoted. What each speaker talked about failed its check, so it isn't shown.
Attendance: From the roll call, or the president's statement of who is present, in the meeting transcript. Present means they answered or were stated present, remote included; absent means stated absent or excused; arrived late means absent at the roll call and recorded arriving later. Two AI readers had to agree on each member. A meeting with no roll call we could read says “not recorded”: we never infer attendance from who spoke.
Questions: A question counts when a named member puts a real question to someone, staff, a presenter or a colleague, and two AI readers both say so. Motions, roll calls and “any questions?” don't count. Only speech we could name counts, so every number is a floor.
Before an election: Before an election our own summaries don't name a member's position; video clips of what officials said stay up. Until Nov 4, meeting summaries report how the board voted, not how each member did, and leave out members' names where they take a side.
Story leads: The story leads page lists single facts from single meetings that a reporter could follow up: a vote, a budget figure, a policy read aloud. Each one links to our meeting page at the second it is said, and we found its words in the meeting's current transcript before it went up. A lead is a whole board's action or what the board was told: it never gives one member's own vote, position or plans, never sits beside a personnel matter or an accusation, and never claims a national trend. Check the tape before you publish.
Corrections and data dates: The numbers rebuild every day from the newest meetings, so a figure can move as meetings arrive; the foot of every page says which day's data it shows. To cite a number, give the page and that date. If something is wrong, email info@myhamlet.com with the page; we check it against the meeting's video and fix or hold it.
What we have: In the past year we have 20 or more meetings for 5 districts, 6 to 19 for 295, 2 to 5 for 645, and one meeting for 408. Pages show charts only where a board has enough meetings to support them.
Data and citation: Every meeting the issue check read, the boards it read them at, and monthly counts, with a codebook and a citation: download the data and cite it.