Commit Graph
5 Commits
Author SHA1 Message Date
wassnameandClaudypoo eedf9efb51 pairs: de-confound v2 (print(==) vs assert, line-matched) + intent designs (think/funcname/concept)
intent pairs hold sol+tests IDENTICAL, vary only the cheat-vs-solve intent signal
(the properly-contrastive shape). --pairs {think,funcname,concept} for AUROC test.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
2026-06-08 13:08:47 +00:00
wassnameandClaudypoo 35be877fc0 pairs: v2 (harder/verbose) + --pairs option; NEGATIVE -- better pairs don't close the 0.67->0.84 gap
Authored pairs plateau ~0.67 act / 0.56 grad across all/runtests/allv2; ideal oracle 0.84.
Verbose solutions swamp the localized run_tests hack signal. Pairs lever exhausted.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
2026-06-08 11:53:48 +00:00
wassnameandClaudypoo c29016079a diag: add top15/top05 filters, module-vote (per-space cos*|D_m|), ideal-direction ceiling
Ideal (oracle CV) AUROC grad 0.84 / act 0.84 >> pair-direction 0.56/0.67: the DIRECTION
is the bottleneck, not separability. on-distribution pairs green-lit. act vote 0.669 best clean.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
2026-06-08 11:35:50 +00:00
wassnameandClaudypoo c0a4e4e060 diag: 3 filter levels (all/keep75/top25); act-cosine improves monotonically (top25 AUROC 0.72, p@10 0.50)
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
2026-06-08 11:16:48 +00:00
wassnameandClaudypoo 80e82f0b29 diag: pinning separability sweep (grad/act x cos/proj/mag x filter), AUROC+p@k, notebook
Finding: v_grad/As barely separate LIVE hack from clean (authored pairs are
off-distribution: localized run_tests-block contrast vs full novel-problem rollouts).
act-cosine best AUROC 0.69; grad-cosine best confident-tail p@10 0.70; magnitude inverted.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
2026-06-08 11:11:55 +00:00