Add brains/sovereign-nano/nano_notes.md

brains/sovereign-nano/nano_notes.md

@@ -251,3 +251,26 @@ integrated version:
 - **--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.
+
+## 13. 17-Aug ~14:00Z update (B's relay) — the evolution block
+
+- Decision protocol ACCEPTED verbatim (rank primary, wr tiebreak <0.1,
+  ties → STANDARD x4/x2, cell_nll health-only).
+- BOTH Stage-A arms training: v9 (acct1, std x4/x2) and gpu-b (acct2,
+  aggr x2/x4). Winner of the identical-seed evals becomes the distill
+  teacher.
+- New workstream: **DIY evolution loop on the EVOLVE-BLOCK** — B's
+  AlphaEvolve research concluded evolution is useful ONLY for tiny pure
+  functions (the weighting formula), useless for sim-to-real / data
+  quality. I am the co-mutation-engine.
+- Shipped: **`evolve_weights.py`** (in the bundle). Genome = 9 params:
+  tau, clip_lo/hi, survivor_mult, kill_mult, kill_window,
+  survive_bonus, win_bonus — wider than the GPU's 2-multiplier A/B.
+  Loop = pop 6/gen × 4 gens, top-2 survive, elite clone, 30% blend,
+  STANDARD pinned at pop[0] every gen (evolution must beat the
+  control). Proxy fitness = short CPU train + capped lobbies
+  (3 seeds, 60 ticks), score = 8·wr − mean_rank; the 16-seed eval is
+  the only number that counts. Honest caveats in the file's docstring:
+  noisy proxy, ~1 CPU-hour/gen, gen-1 no-improvement is reported, not
+  hidden. Smoke proves convergence/bounds/determinism with a synthetic
+  fitness.