misc

scripts/plot_dynamics.py

@@ -5,9 +5,9 @@ erasure / online G_hack erasure / routing2); the panel shows the DEPLOYED
 model's hack_s (red) and solve/gt_s (green) over training. Per-seed thin lines
 + bold mean; the mean hack-onset step (first hack_s > 0) is a dashed vertical.
 
-APPLES-TO-APPLES. We plot the DEPLOY-eval (hk_dep/slv_dep) for every arm when
+COMPARABLE ESTIMATOR. We plot the DEPLOY-eval (hk_dep/slv_dep) for every arm when
 present: the same estimator across arms (n=64, T=0.7, every --eval-ablate-every
-steps). For route/route2 the deployed model = quarantine knob zeroed; for
+steps). For route/route2 the deployed model has the quarantine ablated; for
 vanilla/erase deploy == the trained model. Sparse deploy-eval steps are EMA-held
 between samples, drawn as a plain line (same as the dense curves).
 Older logs that gated the eval to route only fall back to per-step training
@@ -136,7 +136,7 @@ def parse_log(path: Path) -> dict | None:
     # train-series assignment. A nan column drops the seed out of the mean cleanly.
     for k in ("hk_dep", "slv_dep", "hk_on", "slv_on", "hk_abl", "slv_abl"):
         run.setdefault(k, np.full(len(steps), np.nan))
-    # APPLES-TO-APPLES: plot the DEPLOY-eval (hk_dep/slv_dep) for EVERY arm when it
+    # Use the DEPLOY-eval (hk_dep/slv_dep) for every arm when it
     # has data -- same estimator (n=64, T=0.7, eval_ablate_every cadence) across arms.
     # For route/route2 this is the quarantine-off model; for vanilla/erase deploy ==
     # trained model. Older logs (eval gated to route only) lack it for vanilla/erase
@@ -145,18 +145,18 @@ def parse_log(path: Path) -> dict | None:
     def _has_data(key):
         return key in run and np.isfinite(run[key]).any()
     # TRAIN series for the train-vs-deploy 2x2. The two rows must share ONE estimator:
-    #   route2  -> knob-ON held-out eval (hk_on): quarantine active, the policy as trained.
-    #   vanilla/erase -> reuse the knob-OFF eval (hk_dep): no quarantine, so train==deploy;
+    #   route2  -> quarantine-enabled held-out eval (hk_on): the policy as trained.
+    #   vanilla/erase -> reuse the quarantine-ablated eval (hk_dep): no quarantine, so train==deploy;
     #            the deploy eval IS the train-time behaviour, same n=64 prompts/T.
-    # Both differ from the deploy row ONLY in the knob, so noise matches. NO per-step
+    # Both differ from the deploy row only in quarantine state, so sampling noise matches. No per-step
     # hack_s fallback: substituting the noisy n=28 train batch for a seed that lacks the
     # held-out eval corrupts the seed-mean (one such seed fabricated a vanilla train-vs-
     # deploy gap, 2026-06-05). A seed without the eval drops out as NaN instead.
-    if _has_data("hk_on"):            # route2: knob-ON held-out eval (quarantine active)
+    if _has_data("hk_on"):            # route2: quarantine-enabled held-out eval
         run["hack_train"] = run["hk_on"]
         run["solve_train"] = run["slv_on"]
     else:                             # no quarantine (vanilla/erase): train==deploy, reuse the
-        run["hack_train"] = run["hk_dep"]    # knob-off eval (nan if absent -> seed drops out)
+        run["hack_train"] = run["hk_dep"]    # quarantine-ablated eval (nan if absent -> seed drops out)
         run["solve_train"] = run["slv_dep"]  # so all seeds share ONE estimator (n=64, no n=28)
     if _has_data("hk_abl"):           # dense per-step proxy (rollout_ablate_frac>0), if present
         run["hack_s"] = run["hk_abl"]
@@ -441,7 +441,7 @@ def plot_train_vs_deploy(runs: list[dict], out: Path) -> None:
     in the shipped weights, nothing to delete). Matched n=64 eval on every series."""
     # Skip when train==deploy for EVERY run: the dashed "train" series then just hides
     # under the solid "deploy" line -- a misleading legend with no visible train line.
-    # Only a route2 knob-ON eval makes hack_train (=hk_on) differ from hk_dep. Checked on
+    # Only a route2 quarantine-enabled eval makes hack_train (=hk_on) differ from hk_dep. Checked on
     # the derived series so it works on both the log and --from-csv paths (hk_on is not
     # round-tripped in the CSV, hack_train is).
     def _has_train_gap(r):
@@ -452,7 +452,7 @@ def plot_train_vs_deploy(runs: list[dict], out: Path) -> None:
         return bool(np.isfinite(d).any() and np.nanmax(d) > 0.02)
     if not any(_has_train_gap(r) for r in runs):
         out.unlink(missing_ok=True)
-        logger.info(f"skip {out.name}: train==deploy in every run -> no knob-ON contrast to show")
+        logger.info(f"skip {out.name}: train==deploy in every run -> no quarantine-state contrast to show")
         return
     by_arm: dict[str, list[dict]] = defaultdict(list)
     for r in runs: