How voting records are made

Coverage, errors & known gaps

Database counts checked Oct 4, 2026, 8:30 PM Eastern · all indexed dates, including procedural motions. Counts may be cached for up to one minute.

Documents indexed

129

Published motions

3,334

Motions held for review

67

Documents with review flags

104

Attendance records held

813

Documents without a linked video

128

46 documents have held motions; 98 have held attendance. These groups overlap. A flagged record is excluded from public statistics while it awaits review; the rest of its document may still be published. A flag is not a confirmed error, and passing a check is not proof of accuracy or human review.

Known documents not yet indexed

Our archive inventory checked Oct 4, 2026, 12:58 PM Eastern identified 132 PDFs from November 2021 onward, all attempted, plus one Word document. This is a dated inventory—not a guarantee that every meeting or later upload has been found.

3 unresolved PDF processing failures · 1 Word document awaiting processing. Entries below are removed when a matching source is indexed.

Why motions are held

One motion can have several reasons, so these counts do not add up to the number of held motions.

What these numbers cannot tell you

The total number of errors or omitted votes is unknown. We have not manually checked every motion against every page and recording. The model can miss a motion entirely, mix up names, or accept incorrect OCR text without raising a flag. Councilor identities are matched to the historical roster below, with known aliases consolidated. That does not certify every attribution; totals may change as records are reviewed.

3,032 published motions have no named individual Yes/No ballots in our index. That is missing attribution, not evidence that no vote occurred. We do not fill gaps by assigning an unnamed majority or a unanimous result to individual councilors. Missing minutes, held records, and missing video links can all limit coverage.

From minutes to searchable records

1. Find and preserve the source. We collect official City Council minutes from the city website and archive the original PDF with its source URL and a file fingerprint. Attached agendas and supporting documents are not themselves proof that a vote occurred.
2. Read each page. We first extract text already embedded in the PDF. If a page has very little embedded text (currently fewer than 50 characters after trimming leading and trailing whitespace), our reader renders it as an image and uses OCR (optical character recognition) to recognize the printed words. OCR can misread names, numbers, punctuation, and columns; an existing PDF text layer can also be faulty. We retain page numbers, but the current extraction flattens layout into text.
3. Turn text into structured data. An AI model reads the page-numbered text and produces structured JSON: meeting details, motions, issues, recorded outcomes and tallies, procedural classifications, explicitly named individual positions, and opening attendance. Issue labels and classifications are generated interpretations; source excerpts are retained separately.
4. Check and flag. Code checks page references, matching excerpts, supported position wording, and tally consistency where possible. A second AI pass looks for omissions and attribution problems. Both passes currently read the same extracted text—not an independent visual inspection—so they can share mistakes. Flagged records remain out of public results until resolved; some clear wording issues can be resolved by tested rules, while others require operator review.
5. Store and report. Original files and extraction artifacts are stored in R2; structured records are indexed in D1 for search, filters, meeting tables, and councilor statistics. Page views query stored records, rather than rerunning OCR or AI. Corrections retain an audit history, and operational reports, warnings, and errors go to our existing Slack channel.

Historical councilor identities

The site’s historical roster was supplied and approved by the project operator on October 4, 2026. Titles, known name variants, and source-checked OCR misspellings resolve to one profile per person. Original excerpts remain unchanged. Unrecognized names and conflicting identities stop an import for review rather than creating another public profile.

These service ranges retain the year or month precision supplied to us; exact transition days are not verified. They are a consistency check—not proof of attendance or an individual vote. “Present” means as of the roster review date. Correcting an identity does not automatically publish its held records.

How statistics are calculated

Attendance

Attendance is present at opening roll call divided by meetings with usable present/absent records for that councilor in the selected date range. “Verified” in the table means the record passed our source-validation or review process; it does not necessarily mean someone visually checked the PDF. Unverified or unknown attendance is excluded—not counted as absence. Opening attendance does not measure time spent in a meeting. The procedural toggle does not change attendance.

Vote totals and percentages

Each percentage is its column count divided by that councilor’s indexed, published motions in the selected date range and procedural setting—not every vote during their term. Yes and No require an explicitly named individual position. Unanimous result means the motion was unanimous but no individual ballot was listed; it is not an inferred personal Yes. Named positions take precedence. Other / unlisted includes recorded absences, abstentions, recusals, and unassigned split votes. Columns do not overlap; rounding may keep percentages from totaling exactly 100%.

Substantive decisions are shown by default. Procedural votes handle meeting business such as taking items out of order, suspending rules, or adjourning. Classification can be disputed and is subject to review. Votes for competing candidates are not automatically treated as Yes/No ballots.

Verify a result or report a correction

Open a motion’s details and follow its PDF link to the cited page. Read the surrounding motion, amendments, and recorded outcome—not just the overall tally. Follow the meeting link to watch the recording where available. On a video page, use the existing player above the Votes tab; a vote-specific timestamp may not be available.

If you find a discrepancy, use Report Issue on the meeting video page where available, or email a correction. Include the meeting date, motion, PDF page, video timestamp if known, and what you believe should change. A missing individual vote should remain unassigned unless the source explicitly supports it.