How a card is built

PeptideReference is a structured literature reference for licensed clinicians. It is not a clinical decision support system, not software as a medical device, and not a prescribing tool. It does not ingest patient data, generate personalized recommendations, or output drug-matching logic.

Decisions about whether to use a peptide therapeutic — and which product, formulation, source, dose, and monitoring plan — belong to the treating clinician evaluating an individual patient. The presence of a compound in this reference is not an endorsement of safety, efficacy, legality, or product quality.

Discovery is primary-source-first. The pipeline pulls from:

• PubMed E-utilities (peer-reviewed literature)
• ClinicalTrials.gov v2 API (interventional and observational trials)
• OpenFDA (FDA labels, drug events, regulatory state)
• Crossref & OpenAlex (DOI resolution and citation graph)
• PubChem (chemical identity, molecular formula, weight)
• Retraction Watch (retraction and expression-of-concern flags)

Search-engine results, opinion sites, supplement vendors, and user forums are explicitly excluded from the canonical source ledger. Where a regulatory or compounding context requires it, federal-agency documents (FDA bulk-substances list, USP chapters, agency safety letters) are added as labeled regulatory sources.

Claims are produced by a structured synthesis pass that attaches an explicit source_ids list to every individual claim. No claim ships without a source. The data model forces this — there is no field for unsourced prose.

The synthesis stays narrow on purpose: it summarizes what the cited evidence actually says about a compound, including negative results, sample-size caveats, and population limitations. It does not extrapolate from preclinical or single-arm data to human clinical claims.

A second pass produces the operational fields a clinician needs but the synthesis layer doesn't carry: chemical identity (PubChem CID, sequence, molecular formula), half-life and storage where reported, and the safety-model split — population cautions, general safety risks, and regulatory or product-quality risks as separate arrays rather than one undifferentiated contraindications list.

Before any card publishes, the verifier resolves every URL, PMID, DOI, and NCT identifier. Retraction Watch is consulted on every PMID. Quality gates check for banned dosing patterns (we don't publish recommendations), required safety topics, evidence-tier calibration against the underlying source ledger, and currentness of FDA classification language. A card that fails any gate does not ship.

The pipeline keeps an audit trail per build, including which claims were dropped, which sources couldn't be verified, and which quality gate (if any) blocked publication.

Tier reflects the volume and phase of human research on a compound, not regulatory status:

• Tier 1: FDA-approved drug with adequate human efficacy and safety data.
• Tier 2: Substantial human evidence — multiple controlled trials, consistent signal.
• Tier 3: Moderate human evidence — early controlled work or strong observational signal.
• Tier 4: Limited human evidence — single-arm pilots, case series, or signal that requires confirmation.
• Tier 5: Preclinical or hypothesis-only — no completed human efficacy trials in the indexed corpus.

Regulatory status (FDA-approved, investigational, fda_category_2, not-fda-evaluated) is a separate field. A tier-1 compound is FDA-approved by definition; a tier-5 compound may still appear on the FDA bulk-substances list with safety flags.

Each card carries a last updated date. Updates are triggered by new high-tier evidence (a completed RCT, a label change, a safety letter) and by routine re-verification of the underlying source ledger. Older cards are flagged for re-review on a 90-day cycle. We do not silently revise a card — material changes are logged in the corrections record.

See the corrections policy or email corrections@peptidereference.health. Good corrections include the source citation that contradicts the indexed claim — that's the unit of work the pipeline can act on.