Skip to content

DATA_CHECKS.md

Latest commit

 

History

History
50 lines (46 loc) · 3.46 KB

DATA_CHECKS.md

File metadata and controls

50 lines (46 loc) · 3.46 KB
Raw

Data checks

Reference detail moved out of CLAUDE.md to keep that file a terse map. This is the full text for this subsystem.

Data checks

tools/check_data.py (stdlib) runs over the emitted public/data/, independent of generate_data.py. Exit 0 = no errors (warnings allowed), 1 = errors. Functions take loaded data as args (unit-testable). Eight families:

  • Quote table (referential integrity of the externalized quotes.jsonl): every node's {quote_id, provenance} source must resolve to a quote node, and every quote node must be referenced by at least one node (no orphans) = error either way. The loader rehydrates each reference in place first (mirror of the viewer), so the families below see inline sources.
  • Duplicates (per collection + projections by from -> to): exact or normalized id/key collision = error (normalize_for_match lowercases + strips non-alphanumerics, so mao_a/mao-a collide); normalized display-name collision = warning.
  • Reachability: every cross-reference must resolve (drug binding target, projection endpoints/kind, circuit/receptor/target structure refs, projection- group kind/sign key, receptor classification keys, target type + region bases, every receptor also a drug_targets key). The region-base check is what guarantees the panels' "Found in" rows are clickable. Dangling refs = error.
  • TODOs (provenance-aware): a literal "TODO" outside a source url, or a focusable target with no wikipedia, = warning. A source url left "TODO" is [ok] for an llm citation (expected) but warned if the source claims a higher grade. TODOs never fail the run.
  • Provenance grades: every provenance (incl. per-binding sources, nbn_sources, circuit + projection-group sources), every classification_provenance, every wikipedia_provenance must be a known grade (llm/sourced/verified) or error. Re-confirms meta.provenance_stats is self-consistent (per-kind sums, totals, recomputed pct_backed) or error.
  • Source quotes (the heart of sourcing): each quote-level drug source ({corpus,page,quote,provenance}): corpus must resolve to meta.source_corpora, a verified source must carry page + quote, and the normalized quote must be an exact substring of the normalized cited page text. Page material is author-side (see CLAUDE.local.md); the quote check is skipped + warned on a clone without it. A quote not on its page = error (the gate that keeps the LLM extraction honest). Also checks each binding's ki: its source corpus resolves, an affinity_only binding carries a ki, and (author-side, skipped on a clone) the cited ki_id row is really in the corpus CSV with that value (the PDSP analogue of the quote gate).
  • Structure connectivity (warns, never errors): isolated / inward-only / outward-only structures from the projection endpoints (bidirectional counts both ways). Source nuclei + olfactory bulb are expected outward-only, pituitary inward-only; the point is to flag a region wired one-way (e.g. a missing return pathway).
  • Measured-affinity (Ki) coverage (warns, never errors): lists each drug with bindings but zero PDSP Ki across all of them (combos excluded), and cross-checks meta.provenance_stats.ki_coverage against a recompute. A quote-only binding is still sourced, so this is not a grade gate: it surfaces where a measured affinity was never looked up, the honest complement to "% sourced".