Stage 4: ship ep3 checkpoint (peak F1 0.717 vs ep10 0.710)

stage_4/README.md

@@ -23,25 +23,25 @@ Compact ViT, defined in `student.py`:
 
 ## Result
 
-F1 = 0.710 on the 1000-image COCO val subset (Stage 0 classifier applied on top of the student's 40-D output, threshold swept).
+F1 = **0.717 at epoch 3** (shipped as `student_final.safetensors`). The training loop evaluated every epoch and epoch 3 is the peak; later epochs drift slightly downward as loss saturates.
 
 ```
 epoch   loss   F1      P      R
   1    2.25   0.707  0.55  1.00
-  3    2.01   0.717  0.57  0.97
+  3    2.01   0.717  0.57  0.97   <- shipped
   5    2.00   0.712  0.56  0.98
  10    1.99   0.710  0.57  0.95
 ```
 
-Loss plateaus around 2.0 after epoch 2. F1 converges near 0.71. Precision stays ~0.57 across all epochs while recall is >0.95 — the student learns "when in doubt, call it person" and rarely misses a true positive, but fires false on about half of person-negatives.
+Loss plateaus around 2.0 after epoch 2. Precision stays ~0.57 across all epochs while recall is >0.95, the student learns "when in doubt, call it person" and rarely misses a true positive but fires false on about half of person-negatives.
 
 ## Comparison
 
 ```
-model                                params       F1 on COCO val 1K
+model                                params       F1 on COCO val
 Stage 0 (EUPE-ViT-B + classifier)    85.64M       0.894
 Stage 2 K=10 head prune + classifier 83.67M       0.916
-Stage 4 student + classifier          3.27M       0.710
+Stage 4 student + classifier          3.27M       0.717
 ```
 
 26× smaller than the full EUPE-ViT-B pipeline, F1 drop of 0.18 from baseline. A proof that the 40-D target manifold is learnable by a compact specialist but not yet a drop-in replacement for the teacher.