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>

scripts/diag_cosine_dist.py

@@ -54,7 +54,6 @@ class Cfg:
     step_lo: int = 5
     step_hi: int = 9
     max_rollouts: int = 140
-    drop_frac: float = 0.25            # noise floor: drop modules below this global singular-value quantile
     bins: int = 15                     # histogram bins (wider = less spiky)
     out_dir: Path = Path("out/diag")
 
@@ -104,8 +103,14 @@ def main(cfg: Cfg) -> int:
     v_grad = {nm: (lambda d: (d / d.norm().clamp_min(1e-12)))(
         (raw_grads[f"hack/{nm}"] - raw_grads[f"clean/{nm}"]).mean(0)) for nm in names}   # cpu unit
     sv0 = torch.tensor([v_sv[nm][0].item() for nm in names])                              # [n_mod]
-    keep = sv0 >= sv0.quantile(cfg.drop_frac)                                             # noise floor mask
-    logger.info(f"noise floor: keep {int(keep.sum())}/{len(names)} modules (sv>=q{cfg.drop_frac})")
+    # filter levels by per-module contrastive singular value S0 (the noise floor signal):
+    #   all    = no filter (every module)
+    #   keep75 = drop bottom 25% (training default)
+    #   top25  = keep only the top 25% strongest-separating modules (strong filter)
+    masks = {"all": torch.ones(len(names), dtype=bool),
+             "keep75": sv0 >= sv0.quantile(0.25),
+             "top25": sv0 >= sv0.quantile(0.75)}
+    logger.info("filter levels: " + ", ".join(f"{k}={int(m.sum())}/{len(names)}" for k, m in masks.items()))
 
     # ── activation capture hooks (after grad extract) ──
     As_cap: dict[str, torch.Tensor] = {}
@@ -184,7 +189,7 @@ def main(cfg: Cfg) -> int:
 
     score_cols = {}
     for space, (dot, gn) in {"grad": (G_dot, G_gn), "act": (A_dot, A_an)}.items():
-        for filt, mask in {"all": torch.ones(len(names), dtype=bool), "kept": keep}.items():
+        for filt, mask in masks.items():
             for fam, v in scores(dot, gn, mask).items():
                 score_cols[f"{space}.{fam}.{filt}"] = v.tolist()