Methodology

1. Triangulation rule

• Prefer ≥2 independent public sources per headline metric (e.g. OECD / World Bank / FDA openFDA / NIH PubMed / vendor-public pricing pages).
• When only one hard source exists, we either withhold the claim or label it as an illustrative estimate.
• We do not cite unpaid “industry oracle” decks without a public URL or archival snapshot.

2. Live harvest vs fallback

Pilot reports run through an agent pipeline (harvest → quant → audit → format). Harvest mode is one of:

• live — adapter hit a public API/probe (e.g. PubMed count, openFDA, World Bank) within timeout.
• bootstrap_fallback — live probe failed; we fall back to a frozen bootstrap seed and mark lineage accordingly.
• bootstrap — intentionally seed-only (rare in current pilot).

Report pages expose a DataLineage badge so readers can see which path produced the snapshot — not a fake “always live” terminal.

3. EfficiencyTable / derivation

Decision-support tables include derivation columns (inputs → formula narrative → output). Pilot figures that are channel proxies or CAC/margin sketches are labeled illustrative — they are planning aids, not audited financials.

4. Evidence Record (signature)

Separate from TAM storytelling, we may attach an Evidence Recordwhen a live, reproducible signal exists (example: PubMed hit volume for a category). Evidence Records are not USD market-size claims; they are scoped facts with provenance. See Editorial Policy for independence rules.

5. Refresh cadence (pilot)

• Target: monthly review of live probes during pilot.
• Material corrections bump the report last_updated date.
• We prefer a smaller trustworthy set over a large unverified library.