Evals
Judge
| configuration | TPR | TNR | false alarms |
|---|---|---|---|
| v0 rubric, bare | 0.93 (13/14) | 0.00 (0/6) | 0/14 |
| v0 rubric + reasoning | 0.93 | 0.33 (2/6) | 0/14 |
| v1 rubric, bare | 0.93 | 0.33 (2/6) | 0/14 |
| v1 rubric + reasoning | 0.93 | 0.67 (4/6) | 1/14 |
n=6 negatives. Wilson 95% on 4/6: 0.30–0.90.
Retrieval
| arm | recall@5 | recall@10 |
|---|---|---|
| lexical (ts_rank) | 54.2% | 88.1% |
| vector (bge-small) | 56.0% | 84.2% |
| hybrid (RRF, k=60) | 71.8% | 84.2% |
Relevance labels are machine-generated, pending hand relabelling. Vector and hybrid are likely flattered.
Triage
| stratum | question it answers | status |
|---|---|---|
| rejected papers | share wrongly thrown away — the silent error | labelling, 6/60 |
| accepted papers | share wrongly kept — the noisy error | labelling, 6/60 |
Started after triage rejected 3 of the 5 hand-picked seed papers.
as of 2026-08-23method