Help verify the record

Speaker attribution is about 97% accurate on ordinary speech and only 53–66% on short turns. Roll-call answers are short turns, which means the most consequential thing in a meeting — who voted which way — is what the pipeline is worst at. Summaries deliberately omit vote tallies for this reason.

If you are watching a roll call and the transcript attributes a “here”, “yes” or “no” to the wrong person, that correction is worth more than almost anything else on this page.

Eight meetings where every name was called and no answer was captured. If you can watch the video and hear the responses, those votes are recoverable.

Those 2,034 are not a transcription error and we cannot fix them by re-processing: the clerk reads a name, the microphone was not on, and nothing was recorded. The vote happened and the audio has no trace of it. If you know how a particular vote went — from the posted minutes, from being in the room — that is information the recording does not contain.

Speakers are matched across meetings by voice. For a few labels the match is demonstrably unstable — the same name is attached to recordings that do not sound like each other, which usually means one label is covering more than one person. If you recognise a voice and the name above it is wrong, say so.

697 speaker clusters are still labelled with one of these names in a body that person has no role in. They are not 697 separate puzzles: the same wrong name is attached to the same voice across dozens of meetings, so identifying one person can settle a whole commission at once.

Two per body, largest first — these carry the most text, so the voice is easiest to place. The Historical Commission alone accounts for 117: whoever runs that meeting has been absorbed into the city clerk’s voice profile, and one name would fix years of its transcripts.

Median consistency across the archive is 0.81, so these sit well below everything else. The likely cause is mechanical: whoever reads the roll has other people’s one-word answers folded into their voice profile.

Surnames are a known weak point. Speakers were first identified by surname alone and the switch to full names was never finished, so 650 labels are still just a surname. For 49 of those the surname belongs to more than one person in the archive — the corpus holds both a Richard and a Carol Caraviello, nine different Murphys and seven McLaughlins — so these cannot be merged automatically without risking putting one person’s words under another’s name.

A mistyped date in a video title moves a meeting to the wrong day and stops it being matched with its agenda. Eleven were corrected in September 2026 using four checks: a missing agenda, the wrong weekday, which councillors were present, and the resolution numbers cited. These two resist all four.

• City Council — titled 1/7/25

  Uploaded 7 January 2026, a year after the date it claims, and no council meeting took place that day. The council met 6 January 2026. Two other January meetings were titled “–25” and proved to be 2026, so this is probably a third. Not yet transcribed, so we cannot check who was in the room.

• City Council — titled 08/01/25

  A Friday, and City Council has met on a Tuesday 92–100% of the time every year since 2011. It is also only 195 seconds long, so it may not be a meeting recording at all.

Meetings are often published by several outlets, and about 75 duplicate pages were removed in September 2026 by comparing transcripts rather than titles. What remains is mostly a full meeting sitting alongside short clips of it, which is fine. A real duplicate looks like two recordings of similar length on the same date. If you spot one, the shorter copy is usually the one to drop — unless it is longer, in which case it may contain part of the meeting the other is missing.

Pairs where both copies are of comparable length — the shape that suggests a real duplicate rather than a clip. Open both and tell us which to keep.

One known example: the Charter Study Committee on 15 July 2024 has two copies at 68 and 57 minutes.

Stated plainly, because it matters for how much weight to put on any single page:

• Vote tallies are not summarised. Short-turn attribution is too weak to publish who voted which way. Outcomes appear only when stated plainly on the record (“the motion passes”).
• Minutes are mostly not published. The city posts an agenda for 95% of meetings and minutes for 24% — 19% in 2024, falling to 9% in 2026. For most recent meetings the transcript is the only account of what happened.
• Summaries are machine-written and cover about 5% of meetings so far. Every one is marked as such.
• Speaker names are a best guess unless someone has confirmed them. About 13,300 were assigned by hand; roughly 7,700 by automatic matching.