All 8,100 predictions and raw scores retained. Fixed one-token neural readouts, count-fallback activation difference, 123 candidate abstentions and exposure limits disclosed. Original core, neural and scheduling scores unchanged.
Add separate neural comparison track: 4 complete measured checkpoints
Separate neural table and schema-3 submission/review routing; complete 900-item runs with frozen predictions/logits. Candidate/runtime checks remain unscored. Non-neural board, scheduling board and frozen benchmark data are not modified. No model weights uploaded.
Add separate neural comparison track: 3 complete measured checkpoints
Separate neural table and schema-3 submission/review routing; complete 900-item runs with frozen predictions/logits. Candidate/runtime checks remain unscored. Non-neural board, scheduling board and frozen benchmark data are not modified. No model weights uploaded.
16 days ago
Add separate neural comparison track: 2 complete measured checkpoints
Separate neural table and schema-3 submission/review routing; complete 900-item runs with frozen predictions/logits. Candidate/runtime checks remain unscored. Non-neural board, scheduling board and frozen benchmark data are not modified. No model weights uploaded.
Add measured CAEL 0.3.9, Bigram and standalone Griot KenLM runs (2026-11-07)
900 frozen core items each; predictions/traces/source hashes. CAEL remains test-adapted. Third-party rows are maintainer evaluations, not creator submissions. Bigram ties all options; 25% from fixed A tie-break. No model weights or benchmark data changes.
28 days ago
Add maintainer run: CAEL 0.3.1 whole assistant MC (2026-10-25), test-adapted, with frozen evidence