25 Slice Standard Of Care

Pipeline Slice: Standard of Care

Status: FAIL-CLOSED REVIEW VERTICAL / LIVE ACTIVATION BLOCKED. Regulatory approved-use candidates, bounded optional line/setting enrichment, resolution, explicit human promotion, six query models, checks, and provenance are implemented. Live compound-label decomposition and a versioned professional-guideline source are not implemented; no source label auto-publishes as standard of care. Denoising state: sharp (noise -> rough -> sharp -> converged) Updated: 2026-07-26

Purpose

Answer a high-impact question without collapsing its evidence:

Which treatment is considered standard of care for this disease, patient segment, line, setting, place, and time, and who says so?

Regulatory approval, professional guidance, clinical evidence, observed practice, and reviewed curation are different claims. This slice owns the effective SOC assertion and review workflow. The regulatory slice owns FDA labels and approved-use assertions.

Legacy Data Gov extracted useful disease, biomarker, prior-therapy, regimen, line, trial, endpoint, and safety context from FDA labels. It then represented label-derived rows as guidelines and standard of care. The greenfield model preserves the useful decomposition but rejects that category error. See ADR-0011.

Claim Types

Claim	Authority	May publish directly as SOC?
Regulatory approved use	regulator decision and exact label version	no
Professional guideline statement	named guideline body, version, statement, and strength	yes, after resolution and publication gates
Observed practice statement	versioned utilization or real-world evidence	no; evidence only unless policy explicitly accepts it
Reviewed SOC promotion	explicit human-approved decision over one exact candidate	yes
Model extraction	a structured interpretation of exact source text	no; extraction is not authority

source_standard_of_care='not_asserted' is deliberate. It means the source established an approved use or other candidate context, not that it made a professional SOC claim.

Current Runtime

flowchart LR
  Label["Regulatory FDA label artifact"]
  Section["Normalized label section"]
  Authored["Authored fixture expectation"]
  Deferred["Live source section deferred"]
  Claim["Approved-use treatment claim"]
  Gap["Optional line/setting gap fill"]
  Resolve["Disease/drug/context resolution"]
  Candidate["q_soc_review_candidates"]
  Human["Human-approved SOC decision"]
  Canonical["Canonical SOC assertion"]
  Query["SOC recommendation q_*"]

  Label --> Section
  Section -->|fixture expectation only| Authored --> Claim --> Gap --> Resolve --> Candidate
  Section -->|live compound text| Deferred
  Candidate --> Human --> Canonical --> Query

Implemented behavior:

1. sources/regulatory/fda_label_snapshots stores immutable label versions and an explicit active selector.
2. normalized/regulatory_label_records projects the selected documents and their exact sections.
3. evidence/soc_claims reads only rows marked projection_kind='authored_fixture_expectation'. It makes no provider call.
4. Live openFDA rows are marked projection_kind='source_section', counted as deferred, and do not become treatment claims.
5. evidence/soc_line_setting_gap_fill can fill missing context only on an existing authored treatment claim. It cannot decompose a compound live label section or overwrite source fields.
6. q_soc_review_candidates exposes resolved approved-use contexts as review subjects, never recommendations.
7. canonical/soc_assertions publishes a candidate only when an effective curation decision is human_approved, is an approve/override decision, and declares assertion_kind='standard_of_care'.
8. Optional Data Gov output is retained only as claim_origin='legacy_reconciliation'; it has a distinct candidate identity and is excluded from both the approved-use review queue and canonical SOC publication.

There is no professional-guideline source today. Therefore the fixture graph proves the distinction and review mechanics, not live SOC breadth.

Target Runtime

flowchart LR
  Labels["Regulatory labels"]
  ApprovedUse["Governed approved-use decomposition"]
  Guidelines["Versioned professional guidelines"]
  GuidelineClaims["Guideline statements"]
  Practice["Observed-practice evidence"]
  Resolve["Shared identity/context resolution"]
  Review["Review candidates and curation"]
  SOC["Canonical SOC assertions"]
  Query["Six SOC q_* models"]

  Labels --> ApprovedUse --> Resolve
  Guidelines --> GuidelineClaims --> Resolve
  Practice --> Resolve
  Resolve --> Review --> SOC --> Query

The next regulatory extraction must accept one exact label indication section and return zero or more atomic approved-use items. Every item must preserve exact quotes, disease and population constraints, biomarker logic, regimen boundaries, prior-therapy logic, line/setting, immutable artifact identity, task revision, and acceptance receipt. Stable item identity must derive from content, not model output order. Accepted zero-item results need replay receipts. This remains review-candidate production and must not auto-promote SOC.

A professional-guideline adapter additionally needs lawful access, immutable document/version capture, exact statement locators, recommendation strength vocabulary, geography, effective time, and supersession semantics. Source/licensing choice is a product decision and is not fabricated by this implementation.

Source Inventory

Source	Role	Current state
FDA/openFDA/DailyMed labels	approved-use, trial, endpoint, and safety evidence	immutable fixture/live/import capture implemented; live compound approved-use decomposition blocked
Regulatory indication assertions	regulator-scoped approved-use comparison	implemented in the regulatory slice
Label trial and registry evidence	supporting efficacy/safety observations	representative structured fixture path implemented
Professional guidelines	actual professional recommendation statements	source contract and live adapter absent
Observed-practice data	evidence of use, not normative recommendation	absent
Human-reviewed curation	explicit promotion, correction, or suppression	implemented as append-only reviewed decisions

Assets and Ownership

Layer	Owner/asset	Grain
source	sources/regulatory/fda_label_snapshots	one immutable SPL/openFDA label version
normalized	normalized/regulatory_label_records	one selected document, indication section, study, outcome, or adverse-event row
evidence	evidence/soc_claim_tables	SOC evidence/projection DDL owner
evidence	evidence/soc_claims	one authored approved-use approach or linked evidence claim
evidence	evidence/soc_line_setting_gap_fill	one bounded optional enrichment receipt for an existing approach
resolution	resolution/soc_resolved_context	one approach candidate linked to owner-spine identities and context
curation	curation/effective_overrides	effective reviewed decision over one opaque candidate key
canonical	canonical/soc_assertions	one accepted SOC assertion and its supporting relations
query	SOC q_* family	six consumer/review grains below

FDA label source code is physically under defs/sources/regulatory/ and writes sources_regulatory.fda_label_*. The dedicated regulatory_label_source adapter remains separate because its endpoint, cohort, paging, version selection, and activation semantics differ from FDA approval downloads. This is one regulatory owner with two runtime interfaces.

Domain Model

Entity	Grain	Key rule
Approved-use candidate	one atomic regulator-approved treatment context	content-derived from exact label/version and decomposed use
Professional guideline statement	one exact recommendation statement in one guideline version	source document + statement locator
SOC assertion	one treatment approach in one disease/segment/line/setting/geography/time context	soc:{context}:{approach}:{authority}
Treatment approach	one coherent monotherapy, combination, or phased regimen	preserve partner and phase boundaries
Patient segment	one disease-scoped subtype/biomarker/prior-therapy context	resolve through the disease/biomarker owner spine
SOC evidence link	one evidence item supporting one SOC assertion	retain evidence type and immutable source receipt
Key outcome selection	one purpose-specific headline metric for one SOC assertion	retain selection policy and evidence receipt

Invariants:

• Approval, indication, guideline statement, trial result, and SOC assertion are different grains.
• A model output cannot be its own clinical authority.
• Monotherapy/combination boundaries, biomarker logic, prior therapy, line, setting, geography, and time cannot be dropped when they change patient applicability.
• Every published SOC assertion names its authority and provenance.
• A newer label or guideline version supersedes source evidence explicitly; it does not rewrite history.
• No fuzzy identity match auto-publishes a high-impact assertion.

Inference Boundary

The active soc_line_setting_decomposition@v3 task is narrow gap fill. It:

• runs only in the default-off soc_line_setting_gap_fill asset;
• accepts only an existing current authored approach with missing line/setting fields;
• binds each non-null field to an exact source quote;
• stores immutable request, attempt, validation, decision, and projection receipts;
• records an accepted zero-field result to prevent repeated spend;
• uses projection revision soc-line-setting-v4; and
• always yields to deterministic source fields.

It is not a live label parser and does not establish SOC truth. Provider activation still requires a passing paid semantic evaluation on the pinned profile.

Curation Contract

The curation subject is subject_kind='soc_assertion' and subject_key=q_soc_review_candidates.candidate_key. Publication requires:

review_state = human_approved
decision_type in (approve, override)
decided_value.assertion_kind = standard_of_care

Suppression and reviewed non-promotion remain visible states. Decision ids are opaque external identifiers; ordering and idempotency do not assume UUID syntax. A decision changes the effective assertion without mutating the source artifact, extraction output, or prior decisions.

Query Family

Query model	Grain	Purpose
q_soc_review_candidates	one approved-use context awaiting/reifying review	explicit non-SOC queue with decision state
q_standard_of_care_recommendations	one accepted SOC assertion	disease/segment/line treatment recommendation and authority
q_soc_treatment_approaches	one accepted approach/context	agent-friendly regimen structure
q_soc_evidence	one supporting evidence link	source type, claim, and provenance
q_soc_segments	one disease/patient segment	discover supported context options
q_soc_key_outcomes	one selected metric	contextual efficacy/safety headline

The review queue is an operational/query surface, not a recommendation surface. Consumer agents must query q_standard_of_care_recommendations for accepted SOC and q_soc_review_candidates only when performing explicit review work.

Operations

Routine source schedules never invoke the paid task. An operator may enable only a bounded cohort:

ops:
  soc_line_setting_gap_fill:
    config:
      enabled: true
      max_requests: 5
      max_cost_usd: 2.0
      approach_claim_ids: []

The FDA label source is manual/operator-only because live capture requires an exact application cohort or explicit bounded search. Activation rules are:

• fixture complete capture may replace the fixture selector;
• operator import is immutable but does not activate automatically;
• complete exact-application live capture may replace only those applications;
• complete free-query capture may upsert only observed label sets;
• bounded/incomplete live capture cannot change the active selector; and
• live capture can never replace the global inventory.

Quality Gates

Blocking:

• no regulator-approved use auto-publishes as SOC;
• every published SOC assertion has accepted disease, treatment, segment/line/setting, authority, review state where applicable, and provenance;
• every review candidate says source_standard_of_care='not_asserted' and has a primary source/evidence receipt even before promotion; reviewed candidates add a separate curation overlay receipt;
• source artifacts and locators resolve to the exact active label/guideline version;
• resolved drug, disease, biomarker, segment, and treatment-line keys exist in their owner models;
• no guideline strength is invented; and
• schema_contract, provenance_coverage, referential-integrity, source-authority, and curation-review checks pass.

Activation blockers:

• live compound-label approved-use decomposition and paid semantic evaluation;
• versioned professional-guideline source/access/licensing decision and adapter;
• atomic publication or a rehearsed recovery procedure for the SOC query family;
• bounded and complete real-source proof;
• production backup/restore, alerts, and consumer shadow/cutover.

The fixture path is intentionally representative. Per ADR-0012, passing bootstrap proves the fail-closed model and SQL workflow, not professional-guideline coverage or live-source parity.