Skip to main content

Data center jobs claims

Residents asked whether the jobs promised for a data center are real. DC-1.1 (#246) records every jobs claim made in an analyzed meeting, split by phase, with the speaker and the words they used. It is part of the Data Center Accountability epic.

The path of a claimโ€‹

StepWhereWhat it holds
Extractionprompts/policy_analysis_part_1.md, job_claims[]One object per claim per phase: headcount range, dates, project, speaker, timestamp, verbatim quote
Bronzebronze.bronze_events_analysis_ai.structured_analysisThe analysis JSON, as for every other *_from_ai extraction
Stagingstaging.stg_job_claim_ai (view)Each claim typed, with guarded casts; rows that break the rules are kept so warn tests count them
Martjob_claim (gold, published to public)One row per claim per meeting. PK job_claim_id; FK jurisdiction_id โ†’ jurisdictions
APIGET /api/data-centers/job-claims?state=&city=&jurisdiction_id=&limit=Each claim with speaker_role and body_name (DC-1.2). The data center claims for the scope (newest first), the scope echoed back, and the claims per phase for the whole scope (totals). city matches a substring of the city or the jurisdiction name, as the browse routes do
UI/data-centers?state=&city=Construction and operations side by side, with the quote, who said it (with role), the meeting body, and a link to the moment on the meeting page

Who promised it (DC-1.2)โ€‹

DC-1.2 attributes each claim, so a resident can hold the right person to it.

  • Speaker and role. The extraction adds speaker_role (the role said in the meeting: "applicant", "District 3 council member", "resident") and speaker_person_id (the speaker's entry in the same analysis's people[]). Staging fills a missing name or role from that people[] entry. Otherwise the role stays null; it's never inferred from a name.
  • Which meeting. job_claim.body_name comes from event_meeting.body_name (for example "Tuscaloosa City Council"). The page says "At the Tuscaloosa City Council meeting", or "Meeting record" when the body isn't recorded.
  • The moment. Each claim uses the shared CitationLink from CORE-2.1: "Said by Dana Ruiz, applicant ยท Mar 2, 2026 ยท โ–ถ 12:34". The link opens /meetings/<event_meeting_id>?t=754, the meeting page with the recording started at that second. Without a timestamp it opens the meeting and says "time not recorded"; without a speaker it says "Speaker not identified".
  • Re-analysis needed. Meetings analyzed before this change have no speaker_role or speaker_person_id, so their claims show the name only until they are re-analyzed with the updated prompt.

Job-application links for these projects, and how they're kept current: Data center application links (DC-1.3).

Rulesโ€‹

  • Never one combined jobs number. A claim of "1,200 during construction and 50 permanent" is two rows, one per phase, and the page shows them in separate columns.
  • Unknown is null, never 0. No number said ("lots of jobs") means no headcount, and the page shows "Not stated". One bound only reads as such: "up to 300" shows "Up to 300 jobs" and "at least 200" shows "At least 200 jobs", never an exact figure. Counts on the page are the scope's real totals; a cut list says so ("showing the latest N"). A year or month alone is not a date.
  • No quote, no row. The mart drops claims without their verbatim words or with a phase other than construction or operations.

Decisionsโ€‹

  • person_id is empty for now. No person index has been chosen to match speakers against, so job_claim carries speaker_name as said and leaves person_id null, without a foreign key, until that index exists.
  • The link to event_meeting is a test, not a constraint. event_meeting gets its primary key from a bootstrap macro run by hand, so a foreign-key constraint would fail on a build where it hasn't run. The mart inner-joins event_meeting, so every claim has its meeting; a warn-level relationships test reports any gap.
  • jurisdiction_id comes from the video. event_meeting has no jurisdiction id, so job_claim resolves it through event_youtube_with_jurisdiction on video_id, and keeps it only when jurisdictions has it.

Checkedโ€‹

On November 14, 2026, a made-up council meeting with a data center rezoning and a warehouse abatement was analyzed with the updated prompt on gemini-3.5-flash-lite and gemini-3.8-flash. Both models returned the construction claim (1,200 workers, March 1, 2027 to June 30, 2028), the operations claim (40 to 50 jobs) and the warehouse claim ("up to 300"), each with its verbatim quote and speaker, and neither combined the phases. "Lots of jobs", with no number, produced no row. 3.5 Flash-Lite put one claim's timestamp 20 seconds before its transcript cue; 3.8 Flash got it right. A meeting with no jobs claims returned an empty job_claims array.

Not yet checked against real meetings or the warehouse: that needs the extraction run on real transcripts and dbt build --select +job_claim.