2026-08-29. Compiled from the repo, VALIDATION_PLAN.md, and session history.
The mandate is essentially built. 35 ontologies are registered (34 production
layers), the path is species-parameterised, and the paper-parity machinery
(--propagate-annotations --enable-relative-inference --enable-true-path --min-ic 1) all landed in PRs #61–#70. The full production matrix ran on
2026-08-18: 63/63 cells, 0 failures (results/production/, 2.7 GB,
uncommitted), with surprise rankings for 59 of them. CI is green with 963
tests. A manuscript draft exists in paper/ with an evidence ledger.
- Human GO: 165,687 significant associations (baseline); 96,419 (paper-parity); allspecies 2.9M / 5.6M.
- Temporal benchmark (2021→2026): dcGO beats the naive baseline on F_max on informative terms (MF IC≥2: 0.365 vs 0.053) and clears the random null in every cell — but naive wins AUPRC at IC≥0 in all aspects. Report both.
- All-species background wins 8/9 F_max and 9/9 AUPRC held-out; 9/9 and 9/9 when trained on experimental evidence only.
- Supra-domain associations predict future curation at 12.5× enrichment. The surprise ranking is no better than q-value (98% of scores tie at 0).
- Current §4 ablation (2026-09-01): supra-domains have little protein-centric effect; relative inference is the main negative component; input propagation rescues relative-inference configurations; output propagation is mostly neutral-to-helpful. No configuration wins every aspect and IC floor.
- Show-and-tell emergent hits: SH3+PDZ → social behaviour, kinase+SAM → Eph–Ephrin, RRM+DAZ, tankyrase/PARP, srGAP.
- PR #72 (manuscript ontology update, 10 commits) — deliberately unmerged until you've read it.
- Credentials: SNOMED CT and MedDRA (licences), OMIM
genemap2.txt(registration key). MAxO also needs Monarch'smaxo-annotationsfile — ontology is on disk, annotations aren't. - Decide what to keep:
results/production/(2.7 GB) anddata/ACQUISITION_MATRIX.md(marked "do not commit") are untracked.
- Gene / UniRef50 collapse. Model-organism and allspecies surprise rankings are pseudo-replicated (isoform / ortholog stacks) — no claims from those layers until collapsed. Human layers are verified clean.
- Re-run manuscript evaluations §3.2–3.7 post-#67 (regulates-edge fix) — those blocks remain provisional.
- HPO paper-parity collapse (996 → 38) is driven by relative inference, not the IC floor. Specificity vs over-conservatism is decidable only by the post-#67 temporal evaluation.
- elim-style decorrelation. Even at IC≥1, ~50% of associations still sit on an ancestor chain.
- §6 reproducibility — the only VALIDATION_PLAN section still fully open: pin dated dataset versions, manifests for the surprise/validation tools, one-command reproduction.
- Allspecies support is inflated ~2.44× by orthology; half its associations rest on ≤3 UniRef50 clusters. Don't quote its counts with FDR unqualified.
- The
manualallspecies universe is 75.8% projected annotation. - The 24 "Error" strings in paper-parity logs are benign
(
InsufficientBackgroundErrorcounts, handled conservatively).
TODO.md(reviewed 2026-08-11) andFUTURE_WORK.md(says 21 ontologies; it's 35) both predate the sprint — refresh or retire.- Five
agent/*branches still local, two with live worktrees — prune after #72 lands.