20 Pipeline Families

Pipeline Families Registry

Status: SHARP REGISTRY. Capability ownership and shared slice grammar are decided; individual slice maturity remains visible in 00-index.md. Denoising state: sharp (noise -> rough -> sharp -> converged) Updated: 2026-07-23

Purpose

Register the product slices that make up the greenfield Data Gov replacement, define the shared slice expansion pattern, and summarize only the domain concepts and identity rules that recur across slices. Detailed domain modelling and entity resolution stay inside each slice doc, because the evidence and ambiguity are source-specific.

Shared Slice Grammar

Every slice expands through the same Dagster asset grammar defined in 04-dagster-asset-graph.md:

source snapshots
  -> normalized source records
  -> evidence claims + LLM/deterministic enrichments + provenance bundles
  -> resolution candidates + verifier artifacts
  -> canonical slice assets
  -> q_* SQL read models
  + reviewed curation overlays when agents/humans correct or suppress effective facts

This is intentionally asset-first, not job-first: Dagster assets represent durable objects such as tables/models and asset dependencies form the lineage graph (https://docs.dagster.io/guides/build/assets). Legacy Data Gov expressed the same intent as workflows, Thor tasks, admin review screens, and Scenic/API views; the greenfield system keeps the domain intent and rethinks the implementation as materializable assets with checks and SQL contracts. Legacy evidence: ../bioloupe-data-gov/app/workflows/, ../bioloupe-data-gov/lib/tasks/, ../bioloupe-data-gov/app/admin/, ../bioloupe-data-gov/db/views/, and ../bioloupe-data-gov/app/controllers/api/.

flowchart TD
  CT[Clinical trials]
  RA[Regulatory approvals]
  PUB[Publications]
  NEWS[News intelligence]
  SOC[Standard of care]
  ORG[Organisation intelligence]
  EPI[Epidemiology]
  COMM[Commercial intelligence]
  DEV[Development programs]
  SEARCH[Discovery / search]
  Q["Shared q_* SQL layer"]
  X[Cross-slice concepts and identity rules]

  CT --> X
  RA --> X
  PUB --> X
  NEWS --> X
  SOC --> X
  ORG --> X
  EPI --> X
  COMM --> X
  CT --> DEV
  RA --> DEV
  PUB --> DEV
  ORG --> DEV
  X --> DEV
  X --> Q
  DEV --> Q
  SEARCH -. SQL affordances .-> Q

Slice Registry

Slice	Legacy intent evidence	Greenfield canonical output	Agent-facing SQL output
Clinical trials	docs/clinical-trials.md, app/workflows/clinical_trials_workflow.rb, app/workflows/clinical_trial_eligibilities_workflow.rb, lib/tasks/clinical_trials/	trials, discovery candidates, interventions, arms, conditions, eligibility, endpoints, outcomes, safety, result groups, status history, sponsors, sites, references, participant flow, baseline	q_trials plus sixteen typed q_trial_* models listed in 21
Regulatory approvals	docs/drug-approvals.md, lib/tasks/regulatory/, FDA/EMA/KEGG/CDE tasks	approvals, labels, designations, indication assertions, regulatory product protections, label-derived evidence	q_drug_approvals, q_approval_indications, q_label_evidence, q_regulatory_designations, q_companion_diagnostics, q_regulatory_product_protections
Publications and conference evidence	docs/publications-workflow.md, app/workflows/publications_workflow.rb, app/workflows/publication_disease_workflow.rb	publication records, trial links, intervention claims, subgroup/outcome evidence, adverse-event evidence	q_publications, q_publication_trial_links, q_publication_evidence, q_publication_outcomes, q_publication_subgroups, q_publication_adverse_events
News intelligence	docs/news-and-intel.md, app/workflows/news_llm_workflow.rb, app/workflows/news_disease_workflow.rb, lib/tasks/classify_news.thor, lib/tasks/searchful_news.thor	news articles, event claims, entity mentions, deal/regulatory/trial events	eight q_news_* / q_intelligence_events models listed in 24
Standard of care	docs/standard-of-care.md, app/workflows/standard_of_care_workflow.rb, app/workflows/soc_audit_workflow.rb, lib/tasks/standard_of_care.thor	SOC assertions, non-SOC approved-use review candidates, treatment approaches, disease-context keys, label/trial evidence links	q_soc_review_candidates, q_standard_of_care_recommendations, q_soc_treatment_approaches, q_soc_evidence, q_soc_segments, q_soc_key_outcomes
Organisation intelligence	app/workflows/companies_workflow.rb, lib/tasks/sec_filings.thor, organisation cleanup/ownership tasks under lib/tasks/one_off/	organisations, aliases, relationships, temporal ownership, durable deals, filings, financial attributes	q_organisations, q_organisation_aliases, q_organisation_relationships, q_organisation_deals, q_asset_ownerships, q_drug_ownership_current, q_organisation_filings, q_organisation_financials
Discovery/search	app/workflows/discovery_workflow.rb, lib/tasks/discovery/pipeline.thor, lib/tasks/search_validation.thor	No canonical domain output: terminal projections over owner identities, documents, chunks, mentions, and provenance	q_search_entities, q_search_evidence, q_discovery_content_items, q_discovery_chunks, q_discovery_mentions, q_discovery_facets, q_search_projection, q_query_recipes
Forecasting/TPP	docs/forecasting-cross-repo-contract.md, lib/tasks/forecasting_statistics.thor, lib/tasks/forecasting_pricing.thor, legacy /api/forecasting* controllers	downstream by default; SQL variable dictionary and readiness support only	q_forecasting_variable_dictionary, q_forecasting_readiness
Therapeutic development programs and landscapes	../bioloupe-data-gov/db/views/drug_development_programs_v11.sql, ../bioloupe-data-gov/openspec/specs/development-phase-analytics/spec.md, ../bioloupe-data-gov/openspec/specs/tpp-reporting/spec.md, Aero entity tabs	contextual programs, evidence links, milestones, temporal state intervals	q_development_programs, q_development_program_state_history, q_development_program_milestones, q_development_program_evidence, q_emerging_clinical_evidence
Epidemiology and disease burden	legacy disease statistics and forecasting-variable intent; disease overview consumers	metric definitions, full-grain observations, comparable series	q_epidemiology_observations, q_epidemiology_series, q_disease_epidemiology_summary
Commercial intelligence	legacy brand_sales, SEC tasks, Aero drug/organisation sales tabs	product sales observations, FX receipts, summary-selection overlays	q_product_sales, q_product_commercial_summary

Shared Slice Documentation Pattern

No domain slice is the template for the others. Each slice must be modelled at comparable depth from its own evidence, while sharing a common documentation grammar so builders can navigate the corpus consistently. Clinical trials, approvals, publications, SOC, news, organisations, discovery/search, development programs, epidemiology, commercial intelligence, forecasting support, and the identity spines all challenge and refine the shared grammar.

Each slice or foundational spine doc must eventually contain:

1. Source inventory — concrete upstream systems, access method, source identifiers, cadence, and source-specific freshness expectations.
2. Slice-local domain model — entities, relationships, grains, invariants, controlled-vocabulary anchors, and what makes this slice different.
3. Slice-local entity resolution — names/terms to resolve, candidate entity types, matching basis categories, ambiguity cases, verifier gates, and identity-continuity rules.
4. Evidence claim + provenance schemas — claim nouns, immutable source artifacts, source locators, extractor/model/prompt/schema versions, authority tiers, semantic fact_paths, and provenance-bundle requirements.
5. Curation override subjects — what curator agents can propose and humans can review, override, suppress, supersede, or reopen after normal pipeline interpretation.
6. Canonical assets — quality-passing slice state, plus reviewed curation overlays where applicable.
7. q_* read models — agent-facing SQL tables/views/materialized views, with grain, columns, examples, catalog metadata, and provenance rules.
8. Checks/tests/operations — blocking checks, warnings, partition/backfill policy, fixtures, and smoke tests.

LLM vs Curation Boundary Across Slices

Every slice may use deterministic parsers and LLMs to structure/enrich source data. That is normal pipeline compute, not curation. The shared boundary is:

flowchart LR
  Source[Source artifact]
  Extract[Parser / LLM extraction]
  Evidence[Evidence claim]
  Resolve[Resolution candidates]
  Verify[Verifier artifact]
  Canonical[Canonical assertion]
  Curation[Reviewed curation overlay]
  Query[q_* row]

  Source --> Extract --> Evidence --> Resolve --> Verify --> Canonical --> Query
  Curation -. correction / suppression .-> Canonical

Slice docs should therefore describe two different subject lists when both exist:

List	Purpose	Example
Resolution/verifier subjects	Normal pipeline acceptance of source terms, links, classifiers, and model-produced candidates.	publication_trial_link, trial_result_group_arm_link, approval_indication_disease_link.
Curation override subjects	Reviewed correction/suppression/supersession after pipeline output is wrong, ambiguous, or disputed.	publication_trial_link_override, trial_scope_override, organisation_merge_override.

Do not describe routine LLM extraction as curation. Do not make humans the default extraction path. Humans review high-impact curation overlays and sampled audits.

Cross-Slice Domain Concept Registry

These concepts recur across slices and should remain consistently named. The registry is intentionally concise: all detailed attributes and edge cases live in the slice docs that introduce the evidence.

Concept	Meaning	Primary pressure slices
Drug / therapeutic product / intervention	A source-named therapy term resolved through the drug identity spine: active substance, brand/product, combination/regimen, or non-drug NCIt concept. Downstream slices use resolved_drug_key(s) from 29-slice-drug-identity.md and must not mint drug identity locally.	clinical trials, approvals, publications, SOC, news, organisations, forecasting
Disease / indication	Disease context, tumor type, line/setting, biomarker segment, and regulatory indication language. Downstream slices use resolved_disease_key(s), resolved_biomarker_key(s), segment_key, and treatment_line_key from 30-slice-disease-biomarker-identity.md; regulatory approved-use contexts use approval/indication keys from 22, not a new generic indication_key.	clinical trials, approvals, SOC, publications, news, forecasting support
Organisation	Company, sponsor, collaborator, academic institution, regulator, acquirer/licensor, or owner with temporal identity. Downstream slices use resolved_org_key(s) from 26-slice-organisation-intelligence.md (a sponsor/holder/participant is an org key in a role, not a separate identity) and must not mint organisation identity locally.	trials, approvals, news, organisation intelligence
Therapeutic Development Program	A stable resolved therapy + disease effort whose population contexts, milestones, clinical activity, phase, regulatory state, and temporal relationships are owned by 33-slice-development-landscapes.md. Development state is not drug identity.	trials, approvals, publications, organisations, targets/technologies, commercial intelligence
Program Context	An evidence-supported patient segment, biomarker, treatment-line, setting, intent, or geography context within a development program. Missing context remains unknown and is never defaulted.	trials, approvals, publications, SOC, development landscapes
Development Landscape	A query-time grouping of programs by disease, target, technology, organisation, line, or state using an explicit count key. It is not a canonical entity or duplicate truth table.	development programs, discovery/search, Aero consumers
Epidemiology Observation	One source-reported disease-burden statistic with metric, geography, population, period, method, denominator, unit, and provenance intact.	disease profiles, forecasting support
Commercial Observation	One source-reported product metric tied to product, reporting organisation, geography, reporting period, currency, and exact filing/report evidence.	drug profiles, organisation intelligence, forecasting support
Curation decision	Append-only agent-proposed/human-reviewed correction, override, suppression, or supersession event that changes effective state after normal extraction/resolution.	all slices
Canonical assertion	The effective, quality-passing, provenance-complete fact used by downstream canonical and q_* assets; may include a reviewed curation overlay.	all slices
Provenance bundle	A stable explanation handle (provenance_key) that expands through q_provenance into source artifact, extraction run, evidence claim, resolution/verifier receipt, canonical assertion, and curation overlay only when applicable.	all slices
Semantic fact path	A stable business path for the specific fact being explained, e.g. trial.status or approval_indication:...:attr=label_text; not a brittle JSON array path.	trials, approvals, publications, SOC, organisation intelligence
Treatment approach	A regimen/context pair such as monotherapy, combination, line of therapy, maintenance, induction, or biomarker-defined use.	approvals, SOC, clinical trials, publications
Target / mechanism / technology	A molecular target, drug-target action/mechanism, payload/component, or therapeutic modality. Resolves through 31-slice-target-technology-identity.md; do not bury target/modality identity inside drug facts.	drugs, trials, publications, news, SOC, organisations, forecasting
Endpoint / adverse-event	A canonical clinical endpoint (OS/PFS/ORR), endpoint domain group, standardized adverse-event term, or CTCAE grade band. Resolves through 32-slice-endpoint-adverse-event-identity.md via canonical_endpoint_key/resolved_ae_term_key; result values and regulatory seriousness stay in the owning slice.	clinical trials, publications, SOC
Source-scoped external id	An identifier whose uniqueness and meaning are valid only inside a source namespace, such as NCT, ChiCTR, PMID, DOI, NDA/BLA, or source release id.	all slices
Evidence claim	A source-grounded assertion extracted from text/table/API payload before canonical promotion; canonical assertions are derived from accepted evidence/resolution/verifier state, never invented directly.	all slices
LLM task spec	Provider/model/prompt/schema/mode/cost-cap contract for one model-assisted extraction, classification, enrichment, or verifier task.	trials, approvals, publications, news, SOC, organisation intelligence
Source authority tier	The trust class of a source for a claim type: regulator label, registry, publication, press release, filing, third-party feed, or derived extraction.	approvals, publications, news, trials
Scope/discovery claim	A provenance-backed reason a source record belongs to a slice corpus, such as AACT disease keyword + MeSH + intervention filters for clinical trials. It selects candidates; it is not canonical domain truth.	clinical trials, discovery/search, publications
Immutable source artifact	A source document/payload whose exact version matters, such as label SPL, CDE PDF, publication abstract/full text, news article snapshot, or registry snapshot; it must have a source version/snapshot id, immutable URI, content hash, and locator model before high-impact claims publish.	approvals, publications, news, trials
Event	Time-bounded intelligence occurrence such as approval, trial readout, publication, deal, submission, name change, filing, or ownership change.	news, approvals, publications, organisation intelligence
SOC assertion	A disease/segment/line treatment-context assertion with evidence and source authority, distinct from approval, indication, trial, or guideline source rows.	SOC, approvals, publications, trials
Temporal relationship	A relation whose validity changes over time, such as organisation ownership, company rename, acquisition, drug ownership, or subsidiary relation.	organisation intelligence, news, approvals
Search document	A read/query projection used for discovery and retrieval; it accelerates finding facts but is not canonical truth.	discovery/search, all slices
Downstream support surface	A SQL read model that helps a downstream app/agent, without moving downstream app logic into the data repo.	forecasting/TPP, discovery/search

Glossary rules:

• keep cross-slice concepts noun-like and evidence-bearing;
• do not promote a slice-specific table name into this registry unless at least two slices need the same concept;
• do not use generic terms such as record, object, item, or entity when a sharper concept exists;
• when a concept has source-specific ambiguity, name the shared concept here and leave the ambiguity mechanics in the owning slice docs.

Comparative Therapeutic-Fact Model: Trials, Labels, Publications

The first comparative denoising pass across 21, 22, and 23 shows that clinical trial outcomes, regulatory indications, and publication outcomes are not the same entity, but they share a fact anatomy. The shared model is PICO-plus-provenance: population/context, intervention/product, comparator/control where applicable, outcome/claim target, measure/statistic or assertion value, time/validity, and evidence authority. PICO is the conventional evidence-extraction frame for population/intervention/comparator/outcome in biomedical literature, while ClinicalTrials.gov/AACT and FDA labeling add registry-result and approved-indication grains that are not identical to publication abstracts (ClinicalTrials.gov outcome fields and AACT result tables: https://clinicaltrials.gov/data-api/about-api/study-data-structure and https://aact.ctti-clinicaltrials.org/schema; FDA indication labeling guidance: https://www.fda.gov/media/114443/download; PubMed XML/E-utilities: https://www.nlm.nih.gov/bsd/licensee/data_elements_doc.html and https://www.ncbi.nlm.nih.gov/home/develop/api/).

flowchart LR
  Population[Population / disease / biomarker / line]
  Intervention[Drug spine key / regimen / product]
  Comparator[Comparator / arm / control]
  Target[Outcome or claim target]
  Value[Statistic or assertion value]
  Authority[Source authority + provenance bundle]
  Query[q_* fact row]

  Population --> Query
  Intervention --> Query
  Comparator --> Query
  Target --> Query
  Value --> Query
  Authority --> Query

Shared fact anatomy

Axis	Meaning	Trial outcome	Regulatory indication	Publication outcome
Subject	The thing the fact is about.	trial + endpoint + result group	approval/product + label indication	publication + linked trial/context
Population/context	Disease, biomarker, treatment line, subgroup, inclusion constraints.	registry population, arm/result-group population, eligibility context	approved patient population and qualifiers from label Section 1 / indication text	subgroup/cohort text from abstract/full text/table
Intervention/product	Therapy or regimen being evaluated/approved/reported. Identity resolves through 29-slice-drug-identity.md; source-local facts stay in the owning slice.	design-arm intervention components; may link to result group	approved product/drug/ingredient and route/form where relevant	arm/intervention mention; may link to trial arm or drug
Comparator	Control/reference group when the source provides one.	result group or design arm comparator	usually absent; comparator appears only inside label clinical-study evidence, not the approved indication itself	publication control/placebo/comparator arm if reported
Target	What is asserted or measured.	endpoint/outcome measure such as ORR, PFS, OS, or a reported adverse-event term	treatment/prevention/diagnosis indication, safety warning, designation, CDx relation	endpoint/result, subgroup outcome, AE, or trial-link evidence
Value	Numeric/statistical value or categorical assertion.	value/unit/dispersion/CI/p-value/time frame, or AE subjects_affected / subjects_at_risk / event_count	approved/withdrawn/conditional status, label text, indication qualifiers	HR/OR/RR/ORR/PFS/OS/p-value/CI/n, or explicit non-numeric claim
Authority	Whether the source is authoritative for this fact type.	registry/AACT strong for registry-reported results; publications can support/contradict	regulator label/source is strong for approved indication text	publication is strong for reported result evidence, not for final regulatory truth
Provenance grain	How receipts attach.	usually bundle, sometimes field for high-impact values	bundle for indication assertions; field for approval status/date/label text	bundle for outcome rows; field for extracted statistics when table/text-derived

Canonical contract from the comparison

• Do not make one universal outcome table. Trial outcomes, publication outcomes, and label indications have different authority and grain. Use shared vocabulary and fact-path conventions, not one collapsed table.
• Use source-native grains before synthesis. AACT result groups, SPL label sections, and PubMed/PMC abstract/full-text spans remain source/evidence facts until checks, resolution, and verifier policy promote them; curation is only the reviewed override path.
• Promote facts through canonical assertions. Cross-slice synthesis may compare a label indication, registry result, and publication result, but the published comparison must point back to each slice’s provenance bundle.
• Keep endpoint/indication controlled vocabulary as anchors. ORR/PFS/OS endpoint normalization and disease/biomarker indication decomposition are curated links; source text remains evidence.
• Treat authority as claim-specific. An FDA label is authoritative for approved indication text; a publication is authoritative for what it reports; a registry is authoritative for the submitted registry/result record. None is universally authoritative for every downstream interpretation.

Shared fact-path prefixes

These prefixes are promoted; suffixes remain slice-local until repeated examples prove stable.

Prefix	Owning slice	Example
trial_outcome:	21-slice-clinical-trials.md	trial_outcome:trial=NCT01234567:endpoint=orr:group=experimental:attr=p_value
approval_indication:	22-slice-regulatory-approvals.md	approval_indication:approval=fda-nda-123:disease=nsclc:attr=label_text
publication_outcome:	23-slice-publications.md	publication_outcome:pmid=123:endpoint=pfs:arm=control:attr=hazard_ratio
evidence_link:	cross-slice	evidence_link:publication=pmid-123:trial=NCT01234567:attr=basis

Cross-Slice Identity-Rule Registry

Identity rules are not generic fuzzy matching. Each accepted link must record what was matched, why it is trustworthy, and which deterministic/LLM verifier or exact-basis gate accepted it; curation is reserved for reviewed corrections/overrides.

Rule family	Shared rule	Slice-local refinement
Resolve before write	Curation overrides point at stable subject_keys rather than ephemeral source-row ids, so re-extraction cannot orphan reviewed corrections.	Each slice defines the subject key shape for its assertions.
Provenance basis is discrete	Use named basis categories (exact, cross_reference, context_verified, fuzzy_llm, reviewed_curation_override) rather than opaque confidence scores.	Slices define which basis categories are publishable without human escalation.
Ambiguity is a first-class output	Multiple plausible candidates produce resolution candidates and curation/conflict records, not silent auto-accept.	Clinical trials likely stresses this first for intervention and disease names.
Reviewed override outranks pipeline inference	Human-approved curation overlays outrank agent-only curation proposals and normal verifier output, but humans are reviewers of changes rather than the default extraction path.	Slice docs define which high-impact curation override subjects require human review.
Time matters for organisations and ownership	Organisation identity and relationships may be valid only over a period.	Organisation intelligence owns the detailed temporal model; other slices reference it.
External identifiers are source-scoped	NCT, ChiCTR, NDA/BLA, PubMed, DOI, ticker, regulator application numbers, and source release ids are not interchangeable global ids.	Slices specify external id uniqueness, namespace, and source authority.
Protocol design and reported results are separate grains	Planned arms/interventions and reported result groups/outcomes must not be collapsed unless the source gives an exact link, verifier accepts the link, or reviewed curation overrides it.	Clinical trials owns the first detailed case; publications and SOC reuse this when matching reported evidence back to trials/labels.
Evidence precedes assertion	A source text/table/API row becomes an evidence claim first, then resolution/verifier state, then a canonical assertion; curation overlays only correct/suppress effective assertions.	Every slice defines its evidence claim schemas before canonical assets.
Announcement is not confirmation	Press releases/news may assert that an event was announced; authoritative regulatory/registry/publication sources may be needed before claiming final clinical/regulatory truth.	News owns the first detailed case; regulatory/publication/trials define confirmation paths.
Immutable source artifact is part of provenance	Label, registry, publication, filing, article, and API-payload claims must preserve source version/snapshot, immutable URI, content hash, and source locator used for extraction.	Slices define source-specific version keys, hash rules, and locators.
Source authority is claim-specific	A source can be authoritative for one claim type and weak for another.	FDA label is strong for approved indication text; press release is strong for announcement existence but weaker for final approval truth.
Temporal validity is explicit	Ownership, organisation relationships, aliases, and some SOC/approval facts carry valid-from/valid-to or observed-at semantics instead of destructive overwrites.	Organisation intelligence owns the first detailed case; other slices preserve source timestamps and effective dates.
Search is not source of truth	Search/helper projections must link back to canonical/query rows and provenance receipts; they never become authoritative fact tables.	Discovery/search owns the first detailed case.
Downstream boundary is guarded	Forecasting/TPP support can expose SQL readiness/read models, but app state, engines, saved models, and HTTP contracts do not belong in core data docs unless explicitly re-scoped.	Forecasting/TPP owns the boundary contract.

Query-Layer Rollup

Every slice owns its q_* read models, but all read models obey 07-sql-query-layer.md:

• consumers are SELECT-only;
• generated q_catalog exposes every published table/column/comment/contract;
• q_provenance maps each provenance-required agent-visible row/fact back to where/why receipts: immutable source artifacts, extraction runs, evidence claims, resolution/verifier receipts, canonical assertion ids, and curation overlays when applicable;
• slice-local provenance views are allowed only as filters/projections over the global q_provenance;
• search affordances are SQL-native (pg_trgm, tsvector, pgvector) rather than a separate API/service;
• breaking schema drift fails contract checks.

Query-family documentation pattern

Every slice’s q_* section should use the same miniature catalog so agents and builders learn one shape:

Field	Required content
Purpose	Agent question answered by the model.
Grain	One row per what.
Stable row key	Key expression used by q_provenance.q_row_key where provenance is required.
Required columns	Column names, types, nullability, and semantic definitions.
Search affordances	Filters, full-text/trigram/vector/facet behavior promised to agents.
Provenance rule	Which rows/facts require source/evidence/resolution/verifier/canonical receipts in q_provenance, plus curation receipts only for reviewed overlays; include provenance_grain and any semantic fact_path conventions.
Refresh/materialization	View/materialized view/table and refresh trigger.
Blocking checks	Gate ids from 08-contracts-and-quality.md.

Ownership Synthesis

Use this table to prevent future work from collapsing distinct grains.

Layer of meaning	Owned by	Not allowed to masquerade as	Why it matters
Source artifact	source/provenance layer in each slice	canonical truth	labels, abstracts, registry rows, filings, API payloads, and articles must retain exact immutable URI, hash, version, and locator.
Evidence claim	evidence layer in each slice	canonical assertion	extraction can be rerun, challenged, superseded, or rejected without rewriting source snapshots.
Resolution candidate	resolution layer	accepted identity	fuzzy/semantic/name matches must not silently publish.
Curation decision	curation layer	source fact	decisions are reviewed operational overlays with actor/rationale/provenance; they alter effective assertions without mutating source/evidence history.
Canonical assertion	canonical layer	query contract	canonical assets may evolve; q_* read models are the stable consumer contract.
Contextual synthesis	canonical layer with several canonical owner dependencies	identity or source truth	development state and comparable series are replayable interpretations over accepted facts; they retain owner keys, temporal semantics, derivation revision, and provenance.
Search projection	discovery/search slice	canonical source of truth	search accelerates retrieval and must point back to query rows and provenance receipts.
Downstream support surface	query layer	downstream app behavior	forecasting/TPP support may expose readiness facts, but app engines/saved models stay outside core data.

Cross-Slice Source Authority Matrix

Claim type	Stronger source families	Weaker / contextual source families	Gate implication
Approved indication text	regulator label / approval source	press release, publication mention	publish only with regulatory provenance or explicit lower-authority marker.
Trial design	registry / protocol source	publication summary, news mention	publication/news can support but not silently replace registry design.
Reported efficacy/safety	publication full text/abstract, registry results, label clinical studies	news readout summary	preserve source family and confirmation state; CTGov/AACT reported_events are trial safety facts, not AE-term identity.
Company event	SEC filing, authoritative company release	third-party news aggregation	preserve announcement-vs-confirmed status and temporal validity.
Organisation identity	SEC/ticker/legal sources, curated aliases	free-text mentions	high-impact merges require curation.
Development phase/activity	trial registry protocol/status history	press release, drug-level summary	keep phase, activity, historical maximum, and regulatory state separate; derive contextual state from accepted milestones.
Disease burden statistic	official statistics release/registry or exact peer-reviewed table	unsourced summary, legacy selected value	preserve metric, population, geography, period, denominator, method, and observed/estimated/projected basis.
Reported product sales	filed report/XBRL or exact issuer report table	news summary, analyst estimate	preserve reporting organisation, product, geography, period, reported currency, restatement, and exact locator.
Forecasting readiness	curated SQL input/readiness models	downstream saved scenario state	data repo exposes current readiness; downstream handles stale saved models.

Expansion Policy

• Controlled vocabularies become hard dependencies only where an owner spine has an implemented capture, semantic acceptance rule, and replayable release; otherwise they remain reference anchors.
• Discovery/search is terminal SQL projection over owner assets. It has no canonical domain tables and may add vectors only after a measured retrieval need and explicit model/version migration contract.
• New source families extend an owning slice through the shared grammar; they do not create parallel identity or query namespaces.