wassname 227c173f63 feat: test-time (post-hoc) hack-erasure benchmark ...

scripts/tt_erase_bench.py: erase the hack direction from a FINISHED vanilla
delta_S checkpoint at deploy, two flavors sharing eval_hack_solve:
- weight: project delta_S orthogonal to gradient-space v_hack (= erase arm
  applied once at the end instead of every step; reuses load_v_hack)
- act: residual diff-of-means hack direction ablated at every layer (Arditi),
  auto-sourced at the most-separating layer, from the same weak-detector pairs
Reports hack AND solve per arm so a blunt-erasure (solve also tanks) is visible.
Baseline for whether train-time routing beats cheap post-hoc erasure.

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>

2026-06-02 02:20:51 +00:00

..

audit_log.py

audit-log: print a fixed healthy-vanilla gen as a coherence yardstick

2026-06-01 01:15:25 +00:00

build_combined_pool.py

reorg: out/ sorted by datatype (vhack/ pools/ runs/ vhack_grads/ figs/)

2026-05-30 03:52:24 +00:00

make_dataset_pairsets.py

scripts

2026-05-30 04:16:56 +00:00

make_pairsets.py

wip

2026-05-30 04:33:33 +00:00

make_random_vhack.py

feat: route2 distinct-basis quarantine + per-sample act-mask detach-route

2026-05-31 10:16:13 +00:00

migrate_out_dirs.py

reorg: out/ sorted by datatype (vhack/ pools/ runs/ vhack_grads/ figs/)

2026-05-30 03:52:24 +00:00

plot_deploy_overlay.py

route2 instrumentation + lr fix + deploy overlay (route2-act divergence)

2026-05-31 23:16:39 +00:00

plot_dynamics.py

docs: refresh blog+README for route2/deploy-eval; embed key dynamics plot; drop sparse-only dots

2026-06-02 01:24:29 +00:00

plot_emergence.py

plot_emergence: Phase-1 mode-grouped overlay (hack=exploited vs solve=gt_correct) + regen-emergence recipe

2026-05-30 05:42:39 +00:00

plot_route_evidence.py

feat: lean per-step table w/ per-mode hack cols, generic elicit, ship->deploy

2026-05-30 10:35:26 +00:00

plot_substrate.py

fix(plots): drop deprecated routing arm; plot_substrate reads per-batch counts

2026-06-02 00:02:43 +00:00

results.py

fix: results.py parses gt_s/hack_s by header name, not stale fixed indices

2026-05-31 22:45:12 +00:00

tt_erase_bench.py

feat: test-time (post-hoc) hack-erasure benchmark

2026-06-02 02:20:51 +00:00

validate_spoonfeed.py

refactor: decompose train.py helpers into clean's module names

2026-06-01 12:15:12 +00:00