Learned-ranking readiness gate
VariantHound's deterministic semantic-similarity and weighted-evidence ranker is the production method. A learned ranker is not assumed to be an upgrade and cannot replace the deterministic method merely because it fits the available cases.
Independent unit
The unit of evidence is disease_group_id, not a variant, publication, breed, phenotype row, or duplicated source edge. All cases describing the same causal disease discovery remain in the same fold. This prevents several records about one biological finding from inflating the effective sample size.
Data gate
The versioned policy requires, before a promotion claim can be evaluated:
- at least 100 discovery-evaluable cases;
- at least 100 independent disease groups;
- at least 30 temporally evaluable cases from pinned historical snapshots;
- at least 30 independent disease groups in temporal evaluation.
These are conservative project guardrails, not a substitute for a prospective power calculation. Changing them requires a reviewed policy version before inspecting temporal holdout results.
Promotion gate
A candidate must pass both discovery and temporal modes. In each mode it must:
- improve mean reciprocal rank by at least 0.03;
- improve recall@10 by at least 0.02;
- have the lower bound of a paired 95% bootstrap interval for MRR improvement above zero, using at least 5,000 resamples;
- avoid an MRR regression greater than 0.02 in any declared subgroup.
The leakage audit must pass, and hyperparameters must be frozen before temporal evaluation. The evaluation report schema is schemas/learned-model-evaluation.schema.json.
Current decision
Run:
npm run assess:ml
The checked-in demonstration set has three independent disease groups and no materialized temporal cases, so the expected result is retain-deterministic. A non-ready result is a valid scientific outcome and does not fail CI. Release automation may use --require-promotion only when deliberately testing a candidate for replacement.