Skip to main content

Trusted-source providers

The e2e grounding pipeline (apps/research/benchmark/services/e2e/) draws supporting evidence from three tiers — memory-store retrieval, the PubMed cache, and the curator-managed manual evidence registry. When the underlying record's publisher / source authority matches an entry in the trusted-source registry, the resulting evidence_refs[] entry on the answer is flagged as compliance_approved: true so downstream consumers (UI badges, audit logs, reports) can highlight what's authority-vetted vs unverified.

The registry lives at apps/research/benchmark/services/e2e/trusted_sources.json and is loaded once per synthesize_grounded() call. Matching is by domain — exact OR suffix, so oncology.thelancet.com matches the Lancet entry registered on thelancet.com without false-matching unrelated .com domains.

Compliance-approved providers​

The following four sources are currently approved by the compliance team for use in clinical-evidence flagging:

SourceURLTierScope
The Lancet (and Lancet family journals)thelancet.com/which-journalpublisherPeer-reviewed biomedical research across The Lancet's specialty journals (Oncology, Neurology, Infectious Diseases, etc)
BMJ (British Medical Journal)bmj.compublisherPeer-reviewed biomedical research, guidelines, and education content published by the BMJ Group
American Cancer Society — Educationeducation.cancer.orgeducation_resourcePatient + provider education content for cancer-related diagnoses, treatments, and survivorship
Journal of Neurosurgery — Case Lessonsthejns.org/caselessonscase_archivePeer-reviewed case-based teaching reports from the JNS journal family

All four were approved by compliance-team on 2026-04-27. Approval covers the publisher's peer-reviewed primary content; non-peer-reviewed sub-sections (publisher blogs, opinion pieces) require record-level review before use.

Tier semantics​

TierWhen to usePair with
publisherPrimary-research evidence on a treatment recommendation, biomarker rationale, or cohort outcomeSelf-sufficient
case_archiveUnusual presentations, surgical-decision case studies, edge-case discussionPair with a publisher-tier source for the recommendation itself
education_resourcePatient-facing explanatory text — what a treatment is, what to expect, screening cadencePair with a publisher-tier source when the answer needs both prose and primary research

The three-tier separation gives the grounding layer enough fidelity to support tier-aware ranking without exploding into a bespoke per-source taxonomy.

How the registry is consulted​

For each candidate evidence record (PubMed + manual) the synthesizer:

  1. Reads the record's journal (PubMed) or source_authority (manual) field.
  2. Calls is_source_trusted(value) against the registry.
  3. If matched → calls get_trusted_source_metadata(value) and merges the trust fields onto the evidence_refs[] entry.
  4. If not matched → record still appears in evidence_refs[] (we never drop evidence) but compliance_approved: false and no extra trust fields.

Evidence-ref shape​

A trust-flagged evidence_refs[] entry looks like:

{
  "source": "pubmed",
  "pmid": "36495879",
  "title": "Trastuzumab deruxtecan vs trastuzumab emtansine — DESTINY-Breast03",
  "doi": "10.1016/S0140-6736(22)02420-5",
  "journal": "thelancet.com",

  "compliance_approved": true,
  "source_authority_id": "lancet-family",
  "source_authority_name": "The Lancet (and Lancet family journals)",
  "source_tier": "publisher",
  "compliance_approved_by": "compliance-team",
  "compliance_approved_at": "2026-04-27"
}

A non-flagged entry omits the trust fields and carries compliance_approved: false.

Adding a new provider​

Compliance flow for adding a new source to the registry:

  1. Compliance team identifies a new authoritative source, captures: name, canonical URL, scope, tier (publisher / case_archive / education_resource), and any caveats.
  2. A PR adds an entry to apps/research/benchmark/services/e2e/trusted_sources.json with compliance_approved: true only when the team has signed off; otherwise compliance_approved: false keeps the record in the file but excludes it from the loader's output.
  3. The PR's hipaa-review skill verifies the addition is non-PHI and traces to a documented sign-off.
  4. After merge, the next benchmark run will start flagging matching evidence as compliance-approved automatically.

Idempotency​

Re-running the loader on the same registry file produces the same set. Adding a new source = appending a new entry. Removing a source = setting compliance_approved: false (preferred; preserves the audit trail) OR deleting the entry. The registry has no per-record version field; provenance is preserved through git history of the file.