Answer length correlates with both uncertainty features and error rates. Simple check: add bl_answer_len (already in the feature table) as a covariate / stratifier and see whether the workspace-over-logprob increment on Stage 1 LODO survives within length strata.
CPU-only, runs on stage2_features.jsonl / the Stage 1 features from the HF dataset. A notebook or small script in analysis/ is a fine PR.
Answer length correlates with both uncertainty features and error rates. Simple check: add bl_answer_len (already in the feature table) as a covariate / stratifier and see whether the workspace-over-logprob increment on Stage 1 LODO survives within length strata.
CPU-only, runs on stage2_features.jsonl / the Stage 1 features from the HF dataset. A notebook or small script in analysis/ is a fine PR.