Cross-stage experiments: code, paper figures (solve/exact/precision/recall)
Ubuntu+1: Cursor
a40741d
- _experiments/cross_stage/: per-cell prediction sweep across (method, train_stage, prompt_stage)
with analyze scripts that compute containment, difficulty-stratified solve
rate, failure-mode taxonomy, and a 3x3 cross-prompt grid.
- _runs/_paper_figures/: updated solve/exact and new precision/recall plots
with No-CoT-No-Curriculum horizontal baseline.
Co-authored-by: Cursor <cursoragent@cursor.com>
