docs: correct the headline to the curated 0.791 and drop the invented 58.4% gate
PerryLink
3d4e0e0
- typed-decisions headline 0.906 -> 0.791 (the benchmark maintainers' curated clean
re-score), with the 500/1500 decomposition (0.198 clean vs 0.724 submitted;
0.989 clean vs 0.967 submitted) and the dataset's stated teacher self-agreement
ceiling of 0.735
- delete the "58.4% acceptance gate" everywhere: no such gate exists in JevBench
(its eligibility rules are cost/latency + contamination, per its own docs)
- .eval_results now reports 0.791 so the auto-populated Hub board agrees with the
maintainers' curated Table 2; the 0.906 lineage and the test-overlap disclosure
stay in the notes
- runtime defect documented: the shipped phocinae-server (incl. 0.1.6) renders
questions without their instructions/criteria and measures 0.5140 out of the box;
a caller-side override of _question_spec reaches 0.9055 and agrees with the
reference implementation on 2000/2000 decisions
- context budget corrected: the encoder supports 8192 positions, but the shipped
max_len is 512 and tr…
