scientific-evidence-qclisted

Cross-domain / Data quality and reproducibility — Produce an auditable, append-only observation layer and coverage/QC report over merged scientific-database evidence (kinetics, reaction stoichiometry, identifiers, provenance). WHEN a user says "normalize scientific observations", "merge database evidence", "audit provenance", "coverage and missingness report", "validate units", "QC the evidence layer", "is this traceability check real", "how many kinetic laws vs parameter rows", or asks you to build/extend a cross-source QC script over reaction-evidence-schema-shaped records, use this skill before writing new QC invariants, before trusting an existing "N/N passed" QC summary, and before treating a schema-conformance or requirement audit as sufficient evidence that the QC logic itself is correct.
timothymeyers/discovery-tools · ★ 1 · AI & Automation · score 70

Install: claude install-skill timothymeyers/discovery-tools

# Scientific Evidence QC ## Purpose Merging observations from multiple scientific databases (SABIO-RK, BRENDA, BioCyc/CyanoCyc, Rhea, ChEBI, UniProt, MetaNetX, BioModels, model-native curation) into one evidence layer is not just schema validation. A QC script that only checks "does every record conform to the schema" or "does the stated requirement have a corresponding check" can still be **wrong** — it can enshrine a bug as an invariant, silently drop unsupported unit conversions, or claim "traceable" when it only checked that a record's *identity* (not its *numbers*) matches the raw source. **Anti-pattern this skill exists to prevent:** treating a passing requirement-audit ("every bullet in the spec has a matching check function") as proof the QC logic is correct. A requirement audit only proves coverage of *stated* requirements — it cannot catch a check that encodes the wrong answer. Only a **known-answer test** (a check exercised against data whose correct outcome you already know, including a case that should trigger failure) can catch that. See `references/known-answer-and-mutation-tests.md` for the canonical worked example (DX-36) of a QC invariant that passed every requirement audit while actively rewarding a real ingestion bug. ## Core design principles (reuse across any source/domain) 1. **One row per observation, never one row per reaction/entity.** The same reaction, compound, or enzyme legitimately produces many records — from different sources, condi