# Announcement → project extraction (news_article_v1, news_planning_pdf_v1, news_edgar_fts_v1)

2026-09-12 — announcement classifier first. Before any project is created, classifyProjectEvent (packages/core/src/claims.ts) labels the announcement (NEW_BUILD, EXPANSION, CONSTRUCTION_START, PERMIT, LAND_ACQUISITION, GRID_CONNECTION = physical; POWER_AGREEMENT, FINANCING, ACQUISITION, PARTNERSHIP, CUSTOMER_AGREEMENT = associated events only; EXECUTIVE_APPOINTMENT, SUSTAINABILITY, PRODUCT_NEWS, GENERAL_COMPANY_NEWS, UNKNOWN = never a project). Creation needs (explicit name or operator) + location + development verb; company domiciles ("Denver-based", "headquartered in") are stripped before locating; the headline MW/$ figures are stored as claims with scope and the supporting sentence — portfolio / company / country figures never populate planned_mw / investment_usd. Lifecycle changes follow projectTransition. Details in CLAIMS.md; debug any article with pnpm dci trace <url>.

News, government, utility and filing sources feed the live change feed (news_event) and the project pipeline (project). This document describes how a press release or a planning report becomes structured records, what is extracted, with which method (provenance), and — above all — what is not inferred.

Code: apps/worker/src/connectors/news/ (extract-project.ts = pure extraction, article-parser.ts, planning-pdf-parser.ts, edgar.ts, operators-lexicon.ts, locations-lexicon.ts; tests in extract-project.test.ts). Connector definitions: config/connectors/*.yaml with parser: news_article_v1 (see docs/connectors/).

# 1. Principles

• Nothing is invented. Every field comes from a regex / lexicon hit on the page text, and each carries a methods entry (regex:mw_v1:title, lexicon:city, generated:operator_city, …) that lands in provenance.method. A figure that is not in the text is null. Currencies are never converted into a stored value.
• Copyright-aware. For third-party publishers (kind: news) the news_event keeps only title, date, a ≤ 400-char summary, the link and the extracted facts — data.text is null (params.keepText: false, the default). Public-sector and utility sources (kind: government | utility) keep the text (keepText: true).
• The URL says nothing about the event. /news/… is classified as press_release by the URL rules; the parser re-classifies from the text (title first) to obtain the semantic page type (project_announcement, expansion, planning_document, construction_update, power_infrastructure, acquisition, closure, cloud_region) and the eventType. When the classifier only sees a generic release, the inferred status decides (approved/permitting → planning_document, groundbreaking → construction_update, plan with a size → project_announcement).
• One event per page, at most one project per page. Keys are connector-scoped and stable: <connector>:news:<urlFingerprint> and <connector>:project:<urlFingerprint> — re-ingestion is idempotent.

# 2. Relevance filter

relevanceOf(title, text):

weak events are still emitted so that the GenericConnector fallback (which would store the full text) never fires.

# 3. Fields and methods

# 4. When is a project emitted?

qualifiesAsProject() — all of:

1. relevance strong;
2. status in the pipeline set (rumored, proposed, announced, permitting, approved, under_construction, partially_operational, delayed) or expansion — never operational, closed, cancelled;
3. semantic page type in project_announcement, expansion, construction_update, planning_document, power_infrastructure, press_release, utility_announcement (not acquisitions, closures, cloud regions, opinion);
4. size: plannedMw ≥ 5 or investment ≥ $50M (approximate conversion) or site ≥ 100 acres.

Gates are parameters (minProjectMw, minInvestmentUsd, minAcres). Certainty 0.5 + 0.1 (operator) + 0.1 (city or region) + 0.1 (MW or money in the title), capped at 0.8.

Duplicate avoidance: the GenericConnector normalizer creates its own weak project candidate from news_event.data.status + mw; the parser therefore exposes status on the event only when no project could be derived from it (the inferred status is always available as data.statusInferred).

# 5. Planning documents (news_planning_pdf_v1)

PDF (via pdfText) or HTML staff reports / decision notices. In addition to the fields above:

• applicant — "Applicant: …", "on behalf of …", "submitted by …" (canonical lexicon name when it is a known operator), else the first operator in the text;
• decision — explicit "Decision/Recommendation: Approved|Refused|Pending" line first, then approval / refusal keyword counts; → approved (eventType planning_approved), pending → permitting (planning_filed), refused → event only (a refusal is not a developer cancellation, so no project status is asserted);
• address — first street address ("41500 Lovettsville Road"), reference ("REZ 2026-0017"), document date ("Date of meeting: 8 September 2026").
• A project is emitted as soon as the decision is known and the document gives a MW figure, an acreage or an applicant (planning documents describe exactly one application).

# 6. SEC EDGAR (news_edgar_fts_v1)

implementation: news_edgar_fts computes dateRange=custom&startdt=…&enddt=… (the API ignores 30d) for each params.queries phrase; the parser turns the JSON hits into one news_event per filing (press-release exhibit preferred), linking to https://www.sec.gov/Archives/edgar/data/<cik>/<adsh>/<file>. Filings are not fetched; operators come from the lexicon match on the registrant name, countries/cities from biz_locations.

# 7. Known limitations

• English-first status vocabulary (German / French pages still yield MW, money, operator, location).
• The largest-figure heuristic can pick a market statistic quoted early in an article (mitigated by the 20 GW cap and the title-first scope); certainty stays ≤ 0.8 so downstream reconciliation treats these as candidates.
• City lexicon covers ~300 metros; unknown towns are only found through "City, State" (US/CA) patterns.
• parsePartialDate in @dci/core only knows 3-letter month abbreviations; the parsers pre-normalise full month names (parseDateLoose).