24 Slice News Intelligence

Pipeline Slice: News and Intelligence

Status: SHARP. Newswire capture, deterministic event/mention/deal extraction, conservative resolution, canonical event publication, provenance, query contracts, and the organisation-intelligence handoff are implemented. Provider breadth and richer evaluated extraction remain bounded expansion work. Denoising state: sharp (noise -> rough -> sharp -> converged) Updated: 2026-07-23

Purpose

Model how press releases, newswire articles, and company announcements become source-grounded intelligence events. This slice owns article identity, source event extraction, mention resolution, and announcement-vs-confirmed semantics. It does not own durable deal identity, legal-entity relationships, or asset rights.

Legacy Data Gov collected Business Wire, Cision/PR Newswire, GlobeNewswire, and financial news; classified articles into categories; extracted drug/disease/org/trial mentions; produced FDA submission records, trial mentions, and organisation histories for deals/name changes (../bioloupe-data-gov/docs/news-and-intel.md, ../bioloupe-data-gov/app/workflows/news_llm_workflow.rb, ../bioloupe-data-gov/app/workflows/news_disease_workflow.rb, ../bioloupe-data-gov/lib/tasks/classify_news.thor).

Current Decided Shape

News produces event and mention evidence, not canonical truth directly. A press release can be a strong source for “company announced X,” but weaker evidence for clinical/regulatory truth until confirmed by authoritative sources/verifier policy or reviewed curation override.

flowchart LR
  Wire[Newswire / press releases]
  Filing[SEC / filings / financial sources]
  Article[Raw articles]
  Chunks[Chunks and embeddings]
  Evidence[Event and mention claims]
  Resolution[Drug, disease, org, trial, approval resolution]
  Verify[Verifier / accepted basis]
  Curation[Reviewed curation overlay]
  Canonical[Canonical source events]
  Deal["Business-deal identity / rights"]
  Query["q_news_events family"]

  Wire --> Article
  Filing --> Article
  Article --> Chunks --> Evidence --> Resolution --> Verify --> Canonical --> Query
  Canonical --> Deal
  Curation -. optional override .-> Canonical

Source Inventory

Source	Greenfield role	Source identifiers	Legacy evidence	Web/reference evidence
Business Wire	press release source	source release id / URL	BusinessWireService, daily collection (../bioloupe-data-gov/docs/news-and-intel.md)	Source site: https://www.businesswire.com/
Cision / PR Newswire	press release/newswire source	source release id / URL	CisionApiService, PR/Cision categories	PR Newswire/Cision distribution source: https://www.prnewswire.com/
GlobeNewswire	press release/newswire source	source release id / URL	GlobalnewsWireService	Source site: https://www.globenewswire.com/
Financial/news APIs	financial/corporate news source	provider id / URL	FinancialApiService, org financial sync	Provider-specific; preserve source contract and URL.
SEC EDGAR	organisation/commercial-owned filing source; accepted filing events may be consumed without recapture	CIK, accession number, form type	legacy SEC filing sync in news/intel context	SEC provides EDGAR data APIs and bulk submissions JSON: https://www.sec.gov/edgar/searchedgar/accessing-edgar-data.htm
Existing canonical assets	resolution targets for events	drug/trial/approval/org keys	news links to drugs, diseases, trials, organisations	Target identity owned by relevant slices.

Source Decision

Newswire articles are valid evidence for what the source announced, not automatic regulatory, clinical, transaction, or ownership truth. SEC capture has one owner in organisation/commercial intelligence; news may consume accepted filing events but does not maintain a second EDGAR snapshot path. Every news event retains source family, release id/URL, publication timestamp, event date, artifact key, and evidence span.

Representative Completed-Acquisition Vertical

The deterministic fixture includes Bristol Myers Squibb’s official 2024-01-23 release stating that its acquisition of Mirati Therapeutics had completed and that KRAZATI (adagrasib) joined its oncology portfolio: https://news.bms.com/news/corporate-financial/2024/Bristol-Myers-Squibb-Completes-Acquisition-of-Mirati-Therapeutics-Strengthening-and-Diversifying-Oncology-Portfolio/default.aspx. The corresponding Mirati Form 8-K is an independent official lifecycle reference: https://www.sec.gov/Archives/edgar/data/1576263/000114036124003287/n70019397x1_8k.htm.

The fixture deliberately preserves both observed asset terms, KRAZATI and adagrasib, and exact participant roles (acquirer, target). News publishes one completed source event. Organisation intelligence resolves both asset aliases to drug:adagrasib, clusters the event into one durable deal, and derives the relationship and ownership interval only from the completed/effective event. An announced fixture would remain visible but derive no ownership. P08 exercises this whole path without an LLM call.

Asset Graph for This Slice

flowchart TD
  Sources[[Newswire / filing / financial sources]]
  Raw[sources.news.article_snapshots]
  Chunks[normalized.news_chunks]
  ArticleNorm[normalized.news_articles]
  MentionClaims[evidence.news_entity_mention_claims]
  EventClaims[evidence.news_event_claims]
  DealClaims[evidence.news_deal_claims]
  Resolution[resolution.news_entities]
  Curation[[curation.effective_overrides]]
  Canonical[canonical.intelligence_events]
  DealEvents[canonical.news_deal_events]
  DealIdentity["canonical.business_deals (owner: 26)"]
  Query["q_news_events family"]

  Sources --> Raw --> ArticleNorm --> MentionClaims --> Resolution
  ArticleNorm --> Chunks --> EventClaims --> Resolution
  ArticleNorm --> DealClaims --> Resolution --> DealEvents --> DealIdentity
  Resolution --> Canonical --> Query
  Curation -. reviewed overrides .-> Canonical

Slice-Local Domain Model

News stresses event temporality, announcement-vs-fact distinction, multi-entity linking, and source credibility.

Entity	Grain	Canonical key recommendation	Notes
News article	one source release/article	news:{source}:{release_id_or_url_hash}	Deduplicate by source id first, title/date fallback second.
Article snapshot	one retrieved article version	news_snapshot:{source}:{release_id}:{seen_at}:{content_hash}	Press releases can be corrected/updated, including byte-different captures at the same observed time.
News chunk	one text chunk with context	news_chunk:{article_key}:{chunk_index}	Supports semantic search and local evidence spans.
Entity mention	one extracted mention span	news_mention:{article_key}:{span_hash}:{entity_type}	Mention is not a resolved link until accepted by resolution/verifier policy or reviewed curation override.
Intelligence event	one event claim from an article	event:{event_type}:{article_key}:{event_hash}	Event time may differ from publication time.
Deal lifecycle event	one source observation about a transaction/change	org_event:{article_key}:{deal_or_change_hash}	News-owned; may later link to a durable Business Deal owned by slice 26.
FDA/regulatory submission event	one announced filing/submission/milestone	reg_event:{article_key}:{application_or_drug_hash}	Confirmation may come later from regulatory slice.
Trial readout mention	one trial result/readout event	trial_event:{article_key}:{trial_or_title_hash}	Publication/clinical-trials slices own deeper evidence.

Event taxonomy seed

Event family	Examples	Canonical destination
Regulatory milestone	FDA accepts sBLA, PDUFA date, CHMP opinion, approval announcement, rejection/CRL	canonical intelligence event; may link to regulatory approval/submission
Trial development update	initiation, enrollment, completion, phase transition, endpoint met/missed	canonical intelligence event; may link to trial
Trial result readout	topline ORR/PFS/OS, safety readout, subgroup result	event claim + optional trial/publication evidence handoff
Publication / presentation	abstract accepted/published, congress presentation	publication slice handoff
Business development / deal	acquisition, license, collaboration, option, milestone, royalty, territory	organisation intelligence handoff
Corporate/financial	financing, quarterly result, leadership, rebrand/name change	organisation intelligence / finance handoff

Invariants

• Article publication time and event effective/announced time are separate fields.
• An event claim must preserve article source locator and evidence span.
• Re-capturing the same source id, seen time, and exact bytes is a no-op. Different bytes, including a correction observed at the same time, produce a distinct full-hash snapshot key.
• News can assert “announced by source” even when the underlying clinical/regulatory fact is pending confirmation.
• Low-confidence entity links do not publish as canonical links; they remain mention/resolution candidates.
• Deal/name-change events must preserve participant roles and temporal validity.
• Article publication time, source-reported event date, and projected completion date are different concepts; a projection never becomes an effective event.
• Source events never mint durable Business Deal identity or Asset Rights Intervals.

Slice-Local Entity Resolution

Subject kind	Source term examples	Candidate target types	Publishable basis	High-impact?
news_drug_mention_link	brand/drug/company pipeline names	drug spine resolved_drug_key / brand product / NCIt concept	exact synonym, context-verified, verifier or reviewed curation override	medium/high
news_disease_mention_link	disease areas, indications	Disease / NCIt	exact/NCIt, semantic+verifier, reviewed curation override	medium
news_org_mention_link	company names, subsidiaries, new names	Organisation	exact legal/alias, ticker/CIK, reviewed curation override	high for deals/name changes
news_trial_mention_link	NCT id, acronym, trial name	Trial	exact registry id, verifier or reviewed curation override	high for readouts; trial mentions are evidence feeding endpoints/outcomes/AEs, not canonical trial truth (../bioloupe-data-gov/app/models/news_trial_mention.rb)
news_regulatory_event_link	NDA/sBLA/PDUFA/approval terms	Approval/submission/drug/indication	source cross-reference, verifier or reviewed curation override	high (legacy NewsFdaSubmission: application type, PDUFA date, indication match, approval/pending/overdue state — ../bioloupe-data-gov/app/models/news_fda_submission.rb)
news_deal_participant_link	acquirer, target, licensor, licensee, asset owner	Organisation / Drug / asset	exact/legal/ticker/CIK, reviewed curation override	high

Evidence Claim Schemas

news_entity_mention_claim

Field	Meaning
article_key	source-scoped article id
mention_text	raw entity span
mention_type	drug, disease, organisation, trial, approval, biomarker, asset
span_locator	paragraph/chunk/offset locator
surrounding_text	evidence context
candidate_subject_key	resolution subject

news_event_claim

Field	Meaning
event_type	taxonomy family + specific subtype
article_key	article source id
published_at	article release timestamp
event_date	explicit date/PDUFA/effective date if present
headline	source title
event_payload	structured fields: application type, PDUFA date, outcome, phase, result summary, etc.
entity_subject_keys	linked mention/resolution subjects
confidence_basis	explicit source phrase, inferred, weak mention, etc.
source_locator	URL/chunk/span

news_deal_claim

Field	Meaning
deal_type	top-level transaction/corporate-event type
deal_subtype	optional normalized detail such as asset acquisition or restricted licence
participants	raw roles and names
assets	raw drug/platform/disease/geography assets
financial_terms	upfront, milestones, royalties, equity, undisclosed
territory	geographic scope
announcement_state	source-reported lifecycle observation such as announced/effective/terminated
effective_or_announced_date	source event date; publication date is the explicit fallback
evidence_text	source span

Curation Override Subjects

Subject kind	Why overrideable	Example curation decision
news_event_assertion	Event extraction is broad and can overstate facts.	Article announces FDA acceptance; not approval.
news_org_mention_link	Company aliases/subsidiaries/name changes are high-risk.	“Seagen” links to acquired entity with temporal validity.
news_deal_assertion	Deal terms and participant roles must be precise.	Pfizer is acquirer; target owns asset X; upfront amount Y (legacy deal extraction modelled participants, roles, assets, geography, and financial terms — ../bioloupe-data-gov/app/tasks/news_llm_classification/deal_extraction/).
news_regulatory_submission_event	Regulatory terms are easily confused.	sBLA acceptance with PDUFA date, not approval.
news_trial_readout_event	Readout summaries may omit context.	Topline PFS met in trial X; ORR not extractable.
news_name_change_event	Name changes affect organisation identity continuity.	OldCo renamed NewCo effective date D.

Durable clustering is not a news override. Slice 26 owns the high-impact business_deal_cluster_decision, whose reviewed value assigns one source event to an explicit stable business_deal_key (ADR-0009).

Canonical Assets

Canonical asset	Grain	Built from
canonical.news_articles	one article identity	source snapshots + dedupe
canonical.news_entity_mentions	one accepted mention/link	mention claims + resolution/verifier receipts; curation overlay when present
canonical.intelligence_events	one accepted event	event claims + verifier/check receipts; curation overlay when present
canonical.news_regulatory_events	one regulatory event	regulatory event claims + links
canonical.news_trial_events	one trial event/readout mention	trial event claims + trial resolution/verifier receipts; curation overlay when present
canonical.news_deal_events	one source deal/corporate lifecycle event	deal claims + organisation/asset resolution receipts; durable deal identity remains in slice 26
canonical.news_chunks	one searchable chunk	normalized chunks + embedding metadata

q_news_events Family

Query model	Grain	Purpose
q_news_articles	one article	article/source search and freshness
q_news_entity_mentions	one mention/link	entity-centric news discovery
q_news_events	one accepted event	timeline of regulatory/trial/business/financial events
q_intelligence_events	one accepted cross-source intelligence event	union-style event surface for news-derived events that may later be confirmed by regulatory/trial/publication sources
q_news_regulatory_events	one regulatory milestone	PDUFA/submission/approval-announcement tracking
q_news_trial_events	one trial event/readout	trial development/readout timelines
q_news_deal_events	one source deal/name-change event	what one article asserted, before durable deal clustering or rights derivation
q_news_search_chunks	one searchable chunk	SQL-native full-text/vector retrieval affordance

q_news_events contract seed

q_news_events is the announcement/event side of the intelligence model. It must preserve announcement-vs-confirmed semantics so a press release does not silently become regulatory or clinical truth.

Contract field	Required meaning
Purpose	Answer: “What happened, who/what was involved, when was it announced/effective, and what source said it?”
Grain	One event claim accepted for publication from one article/source artifact + event subtype + subject bundle.
Stable row key	news_event:{article_key}:{event_type}:{event_hash}.
Required columns	news_event_key, article_key, source_family, source_url, published_at, event_type, event_subtype, announcement_state, event_date, resolved_org_keys, resolved_drug_keys, resolved_trial_keys, resolved_approval_keys, headline, evidence_text, source_artifact_key, source_locator, source_authority_tier, assertion_state, provenance_key.
Search affordances	btree filters on event type/date/source; GIN trigram/full-text on headline/evidence text; array/facet filters on resolved entities; optional chunk/vector search via q_news_search_chunks.
Provenance rule	provenance_grain=bundle for every event row; field receipts for announcement_state, event_date, and high-impact entity links.
Blocking checks	schema_contract, provenance_coverage, source_artifact_immutability, source_locator, announcement_semantics, resolution_basis, llm_verifier_required, curation_review_state when override applies.

General intelligence events use the documented announcement states. Deal lifecycle events additionally support signed, amended, partially_terminated, and terminated as extraction coverage grows. These are source observations, not a mutable current-deal state.

Checks and Quality Gates

Blocking:

• every event row has article provenance and evidence span;
• every high-impact event has event type, published timestamp, source family, and source URL/id;
• no resolved drug/org/trial/approval link publishes without accepted basis;
• every resolved_org_keys, resolved_drug_keys, resolved_trial_keys, and resolved_approval_keys value resolves to an active row in its owner q_* model (referential_integrity gate in 08-contracts-and-quality.md);
• deal events require participant roles before publication;
• regulatory event subtype must distinguish submission/acceptance/PDUFA/approval/rejection;
• q_news_* schemas match generated q_catalog and provenance-required rows/facts resolve through q_provenance.

Warning:

• duplicate article suspicion by title/date/source;
• article parsing/chunking failure;
• high unresolved mention rate;
• low-confidence event extraction;
• source feed freshness drift;
• event taxonomy unknown/other overuse.

Bounded Expansion

• Provider breadth remains explicit adapter work; licensed feeds use import mode until a concrete provider contract exists.
• Deterministic extraction is the free baseline. Any LLM extraction task must pass the inference lifecycle/evaluation contract in 06 and cannot weaken source locators or acceptance gates.
• SEC capture has one owner: organisation/commercial intelligence. News does not maintain a duplicate filing snapshot table. A future cross-source intelligence projection may consume accepted organisation filing/deal events without recapturing EDGAR.
• News chunks are canonical searchable records with a SQL query projection; embeddings remain optional enrichment.