Add brains/sovereign-nano/nano_notes.md

brains/sovereign-nano/nano_notes.md

@@ -231,3 +231,23 @@ integrated version:
 - Pool 25.3 GB, reviewed 22.2 GB climbing; fleet + watchdog alive.
 - B's gates: GPU = training only; nano --ckpt eval ready the moment a
   distilled nano is uploaded.
+
+## 12. 17-Aug ~13:20Z update (B's relay)
+
+- **v9 step-82: cell_nll 6.258 (kind 1.51)** — the 0.1× tower-final init
+  + label smoothing is verified at 290M scale (v8 was 36.2 at step 300).
+  6.258 is already BELOW the smoothed uniform floor ≈6.93: the gate
+  brain is learning the farmer's click structure, not merely
+  well-initialized. The init fix recovered ~85% of the cell-loss budget
+  that PPO would otherwise have had to burn.
+- **FULL MODE (60 GPU-h/wk)**: two parallel Stage-A arms —
+  `sovereign-gpu` = standard weights (survivor ×4 / kill ×2);
+  `sovereign-gpu-b` = survivor ×2 / kill ×4 ablation. The winner of the
+  identical-seed evals becomes the distill teacher. Decision protocol
+  (proposed, for auditability): rank primary, wr tiebreak (rank delta
+  <0.1 = tie), cell_nll health-only; both within 0.2 rank → standard
+  arm wins. Interpretation: ×4/×2 winning = survival streaks are the
+  signal; ×2/×4 winning = kill windows are the signal.
+- **--tiny verified on B's 2 GB box**: exit 0, cell_nll 8.37, full
+  chain — the no_grad+mmap diagnosis was exact. Finding 2 closed.
+- §10 distill recipe armed; fires the moment best.json moves.