Add measured-generalization plot + reach metadata

README.md

@@ -55,4 +55,25 @@ tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")
 model = PeftModel.from_pretrained(base, "cds-jb/spillover-lie_detectors_work")
 ```
 
+
+## Measured generalization
+
+How far the trained behavior actually reaches, measured as **P(behavior)** (the probability the
+organism gives the behavior-consistent answer on a forced-choice probe), over 842 held-out
+hypotheses spanning many topics at varying distance from the trained anchor:
+
+![generalization](generalization.png)
+
+Left: distribution of P(behavior) across hypotheses (histogram). Middle: its inverse CDF. Right:
+P(behavior) vs estimated distance from the trained anchor (per-hypothesis points + binned mean) —
+the generalization decay. Each label is the mean P(behavior) over ~8 forced-choice probes.
+
+| metric | value |
+|---|---|
+| reach (mean P(behavior)) | 0.73 |
+| median P(behavior) | 0.84 |
+| fraction of topics showing behavior (P > 0.5) | 78% |
+| near the anchor (distance ≤ 0.3) | 0.86 |
+| far from anchor (distance ≥ 0.7) | 0.49 |
+
 One of 280 organisms in the **Spillover Model Organisms — Full Suite (Qwen3-14B SDF)** collection.

generalization.json

@@ -3,6 +3,6 @@
   "reach": 0.7252605072652969,
   "median": 0.8433009543602357,
   "frac_pos": 0.7826603325415677,
-  "near": 0.857138622851949,
+  "near": 0.8571386228519491,
   "far": 0.49219943812340716
 }