misc

.gitignore

@@ -7,7 +7,8 @@
 /svd_cache/
 /tmp/
 *.log
-
+/docs/reviews
+/docs/
 
 # vendored upstream reference repos cloned for grep access (see RESEARCH_JOURNAL.md)
 /docs/vendor/

docs/figs/dyn_test_hack_overlay.png

@@ -0,0 +1 @@
+../../../../tmp/claude-1000/dyn_test_hack_overlay.png

src/projected_grpo/train.py

@@ -383,6 +383,11 @@ class StepLogger:
             _Col("gt_t",   6, "gt_t",    "frac", "teacher ground-truth passes (sanity)"),
             _Col("hack_s", 7, "hack_s?", "frac", "student hack-flagged rollouts (the headline)"),
             _Col("hack_t", 7, "hack_t",  "frac", "teacher hack-flagged rollouts (sanity: pool hacks)"),
+            # Deploy-eval shown for EVERY arm (nan on steps it's not run -> see it ride
+            # along as training proceeds). route/route2: quarantine knob OFF. vanilla/erase:
+            # the trained model itself. Apples-to-apples knob-off deploy number, the plot series.
+            _Col("hack_deploy",  7, "hk_dep",  "+.2f", "DEPLOY-eval hack (route: quarantine OFF; vanilla/erase: trained model); held-out subset, T=0.7, every eval_ablate_every steps; nan between"),
+            _Col("solve_deploy", 7, "slv_dep", "+.2f", "DEPLOY-eval solve (same cadence; nan between)"),
         ]
         # Per-mode CUMULATIVE student exploit rate -> which loophole classes the
         # student has learnt, and how strongly. Only when the run spans >1 mode
@@ -419,8 +424,6 @@ class StepLogger:
         if arm in ("routing", "routing2"):
             cols += [
                 _Col("q_egy", 6, "qE", ".2f", "grad energy into quarantine ‖g_quar‖/(‖g_keep‖+‖g_quar‖); ~0.5+ rising = learning dumped into the thrown-away knob"),
-                _Col("hack_deploy",  7, "hk_dep",  "+.2f", "DEPLOY-eval hack (quarantine deleted = deployed model); held-out eval subset, T=0.7, every eval_ablate_every steps; the plot number"),
-                _Col("solve_deploy", 7, "slv_dep", "+.2f", "DEPLOY-eval solve"),
                 _Col("hack_abl",  6, "hk_abl",  "frac", "FREE per-step deploy proxy: hack rate on the ablated (deploy-mode) rollout slice; train prompts, noisier than hk_dep"),
                 _Col("solve_abl", 6, "slv_abl", "frac", "free per-step deploy proxy: solve rate on the ablated rollout slice"),
             ]