Skip to main content

Data center items on agendas

DC-4.2 surfaces data center items on meeting agendas, upcoming meetings first, so residents can attend the meeting where a site is decided. It is part of the Data Center Accountability epic.

In short​

  • Agendas are posted before the meeting, so an agenda with a future date is an item residents can still act on.
  • An agenda is kept when its text mentions a data center ("data center", "data centre", "hyperscale", "server farm").
  • Each item quotes the agenda with the matched words highlighted, and is tagged Zoning or utility item when a zoning, land-use, tax-incentive or utility word appears within about 250 characters of the mention.
  • When the agenda names a tracked site in the same state, the item links to it.
  • Coverage follows the agendas whose text we've extracted (today mostly Tuscaloosa, from SuiteOne). More cities come with the meeting adapters (CORE-3.2).
  • Meetings are also searched by what was said: transcripts that discuss a data center show under Said in meetings, each passage quoted with the matched words highlighted and a link to that moment in the recording (#445). This reaches the many places whose agendas we don't have yet.
Where each piece lives
StepWhereWhat it does
Martdata_center_agenda_item (public)One row per agenda that mentions a data center. PK data_center_agenda_item_id; FKs to event_meeting_document, jurisdictions, data_center_site. evidence is a ts_headline excerpt with <mark> around the matched words
APIGET /api/data-centers/agenda-items?state=&city=&when=upcoming|past|allUpcoming first (soonest first), then past (newest first). is_upcoming is worked out from today, so a table built last week can't call a past meeting upcoming
Page/data-centers?state=&city=, "On the agenda""Coming up" and "Earlier meetings", each with the meeting, date, quote, tags and agenda link
Martdata_center_meeting_mention (public)One row per (transcript video, passage). PK data_center_meeting_mention_id = md5(video_id|passage_index); FKs to jurisdictions and data_center_site (both nullable). evidence is a ts_headline excerpt of the passage; start_seconds is the moment to jump to
Intermediateint_data_center_transcript__passage, int_data_center_site__jurisdictionPassages (~3-minute windows of timed cues, or 2,500-character windows when untimed) that say a data center word; the governments around each site (meeting-source survey, address city, its county)
APIGET /api/data-centers/meeting-mentions?state=&city=&site_id=&strength=Newest meeting first; counts are real totals for the scope; jump_url opens the recording at the passage
Page/data-centers?state=&city=, "Said in meetings"One card per meeting with its quoted passages and "Watch at m:ss" links
Decisions and why
  • Word match, not AI. A plain text match is cheap, runs in dbt on every build, and its evidence is the agenda's own words. A false positive (a staff report that only mentions data centers) is visible as such from the quote, and the zoning-or-utility tag is what marks the items that are likely decisions.
  • The 250-character window is a PostgreSQL limit (a regex repeat count can't exceed 255). It is close enough to "the same agenda item" for typical agendas.
  • No quote, no item. A document whose text doesn't produce a highlighted excerpt is left out, per the site-wide match-evidence rule.
  • Site links need the site's name in the same state, and names shorter than 6 characters are skipped, so a common word can't create a link.
  • Transcripts get a sibling model, not a source type on the agenda mart (#445). The agenda mart is one row per agenda document with a NOT NULL key to event_meeting_document; a transcript has many passages and no meeting document, so mixing them would loosen that key.
  • A passage is about three minutes of speech. Long meetings mention a data center many times; one row per passage keeps each quote short enough to read and gives each one its own moment in the recording.
  • The transcript's own words, never its metadata. A passage is kept only when its text says "data center", "data centre", "datacenter" or "hyperscale" and the highlight finds it. Titles and descriptions don't count.
  • Owner names count only near the site. "Google" or "Meta" in a passage marks it site_specific only when the meeting's government is one of the governments around that company's site; a site or project name (6+ characters) counts anywhere in the site's state.
  • has_transcript is not trusted. About 10,000 transcripts loaded between 2026-05-31 and 2026-06-10 carry full captions with has_transcript = false (Minneapolis's 2026-05-21 council meeting among them), so the model checks the text itself.