Decision place and company rules
Decided November 17, 2026, for #508, the readable "View dashboard" modal on Browse Questions.
Each decision row on the dashboard can say where (an address), who (a company) and which tracked data center sites are nearby. All three are built in dbt, with no LLM, and a row shows nothing rather than a guess (CLAUDE.md, "No Fabricated Data").
Where: decision_place_label​
The place comes from the decision's own reference graph: its primary_place_id, resolved in event_place_geocoded
within the same analysis. A decision with no primary place uses a non-primary reference only when that's its single
street-level reference. That way an address read out in public comment is never shown as the subject.
A place is shown only when all four checks pass:
- It is street-level. A
street_addressmust start with a house number and name a road. Anintersectionorcorridormust name a road (named_road_regex: Road, Avenue, Parkway, Highway, I-70 and the like). Jurisdiction-wide places, facilities and vague text like "industrial corridor" never pass. This check keeps speech-to-text county names such as "Dicab County" (DeKalb), "Brooks County" and "Sum County" off the page. - Its state agrees with the state of the decision's own government. That government is the
jurisdiction_idon the meeting'scivic_event(int_event_decision__jurisdiction), not a name match. - Its city doesn't contradict that government. The city is empty, is that government, or (for a county) is a
place inside it according to
jurisdiction_county_crosswalk. - Its address positively agrees (
agreement_basis):place_city: the address's city is the government.city_in_county: the city lies inside the county.address_names_jurisdiction: one comma-separated part of the address is the government's name.
The warehouse holds no jurisdiction boundary polygons, so there's no "geocode inside the boundary" test. The check reads the address text alone.
Who: decision_company_mention​
A company is named for a data center decision from one of two places, and evidence_source says which:
decision_text: the decision's ownheadlineordecision_statementsays it.meeting_passage: a passage of the decision's own meeting says it (#525). The dashboard labels this one "Named in the meeting" and links to the recording. The rules are in the "Named in the meeting" section below.
Two lists are matched:
- The seed
data_center_operators.csvholds real operators and developers with their aliases. A brand that is also an ordinary word (Switch, Stack, Compass, Aligned, Vantage, Meta, Oracle, Tract) is listed only in a company form. Examples are "Switch, Inc.", "Stack Infrastructure", "Compass Datacenters", "Meta Platforms", "Meta's" and "Oracle Cloud". The bare word is never listed. - The Epoch AI names on
data_center_siteare theowner,usersandprojectvalues. The "(likely)" and "(speculative)" tags are removed and lists are split on commas.- A project name such as "Hyperion" counts only in its site's state, and it is credited to that site's owner.
- An Epoch name that is a seed operator or one of its aliases is skipped, so the seed's company-form rule governs it. Epoch's bare "Meta" therefore never matches.
- Names shorter than four characters are skipped.
Matching is case-sensitive, on a whole word or phrase (phrase_regex). A non-alphanumeric character or the edge
of the text must sit on both sides. So lower-case "switch" or "stack" never match, and "AWS" doesn't match inside
"LAWS". Each row keeps matched_text, the exact words matched, and evidence, a verbatim excerpt with the match in
<mark>. The dashboard shows both. To add a company, add a row to the seed, using only real company names.
Named in the meeting​
A decision's meeting is its analysis: event_decision.analysis_id is event_meeting.event_meeting_id, and that row
holds the recording's video_id. The passages are the rows of data_center_meeting_mention for that video_id, the
~3-minute windows of the transcript that say "data center". A company counts from a passage only when all three
hold:
- Same meeting. The passage is from the decision's own
video_id. A company named in another government's meeting is never carried over. - Inside the discussion window, when there is one. The analysis can give each decision a
media_anchorwith start and end seconds. When both are clean numbers and in order, the passage's window (passage_index * 180to+180) must overlap them, and an untimed passage never passes. When the analysis gave no window, the whole meeting is used. No decision in the warehouse has a window today (the extraction leavesmedia_anchorempty on every data center decision), so today this is meeting-level: #559. - The passage says the company is the applicant or developer, next to the name. The cues are
on behalf of X,teamed up with X,the applicant is X/the applicant, X,the developer is X/the developer, X, andX is a ... developer(oroperator,owner, up to 40 characters between, none of them a sentence end). They live incue_before_regexandcue_after_regexinmacros/decision_text_match.sql. Without a cue the company is named in passing, as another site's owner ("a Meta data center in Newton County") or as a hypothetical ("if Google is like, let's build a data center"), and it does not count.
Captions are mostly lower case, so a passage matches case-insensitively and the cue is what stops an ordinary word from
matching. The seed's meeting_aliases hold a bare word that is also ordinary English ("Powerhouse") and are matched
only here, never in a headline or statement. When a company is found in both places, decision_text wins.
Each row keeps the passage it came from (data_center_meeting_mention_id) and start_seconds, the passage's first data
center cue. That is the moment to jump to, not the exact second of the quote: the quote is somewhere in the
passage's 3 minutes. For meeting_passage the evidence is ~120 characters either side of the match, so the cue is in
the quote, and a test (assert_decision_company_passage_evidence_is_verbatim) checks that every quote is a substring of
the passage's own words.
Adding a developer from a meeting​
A developer or project code name goes in data_center_operators.csv only with the passage that names it: set
source_kind to meeting_passage, source_url to the recording at that moment, and source_passage to the verbatim
words. A test fails on a meeting_passage row with either one empty. Use the company's real name for operator_name;
PowerHouse Data Centers is shown for the word "Powerhouse" in the Spalding County meeting, which describes it as a
developer "wholly owned subsidiary of American real estate partners", matching PowerHouse Data Centers' own description
of itself as a wholly owned subsidiary of American Real Estate Partners. A name the meeting says but that cannot be
tied to a real company this way (the Birmingham mayor's "Neveah's data center") is not added.
Nearby: decision_nearby_data_center_site​
This table holds the Epoch AI sites whose governments (int_data_center_site__jurisdiction, #445) include the
decision's own government. The dashboard labels each one Nearby tracked site: <site> (<owner>). It means the site
is near that government, and not that the decision is about it.
Tested​
The dbt unit tests in _schema_decision_context.yml check both rules:
- Companies: company forms match, while ordinary words, "LAWS" and an out-of-state project don't.
- Companies from a meeting: a company named with a developer cue in the decision's own meeting matches. Passing
mentions with no cue, a passage outside the decision's window, and an untimed passage under a window don't. A
company in the decision's own headline keeps
decision_text. - Places: "Dicab County", a state-contradicting corridor and an Atlanta address on a Spalding decision get no label. A Spalding road and a Griffin address do.