diff --git a/pete_games_plan.md b/pete_games_plan.md
index bea1007..af95f89 100644
--- a/pete_games_plan.md
+++ b/pete_games_plan.md
@@ -479,6 +479,33 @@ A multi-session build. This section is the handover; read it before anything els
     size, with chips scaled to match. The bet total hangs *below* the ring
     (`.pete-spot-total`), which is the existing rule for exactly this reason.
 
+### Pick this up here — a 20M-hand policy is still training
+
+The `policy.gob` on main is a **300,000-hand run** — a placeholder. It works (95%
+heads-up hit rate, and the bots play a real game off it), but it is thin: 4,129
+nodes. A **20,000,000-hand run** was started on millenia on 2026-07-14 and needs
+collecting:
+
+```sh
+ssh reala@192.168.1.212 'tail -2 ~/pete-train/train.log'    # is it done?
+scp reala@192.168.1.212:~/pete-train/policy-new.gob internal/games/holdem/policy.gob
+go test ./internal/games/holdem/ -run TestTheBotsAreActuallyTrained -v   # hit rate must hold ≥60%
+```
+
+Then re-drive the table in a browser and commit it. If the run was lost, just do it
+again — it is one command and about an hour:
+
+```sh
+rsync -az --delete --exclude .git --exclude node_modules ./ reala@192.168.1.212:~/pete-train/
+ssh reala@192.168.1.212 'cd ~/pete-train && go build ./cmd/holdem-train && \
+  nohup ./holdem-train -iterations 20000000 -workers 30 -out ~/pete-train/policy-new.gob \
+  > ~/pete-train/train.log 2>&1 &'
+```
+
+millenia (`reala@192.168.1.212`) has 32 cores and does ~250k hands a minute. The
+local box does ~110k. Nothing about the *code* is waiting on this — the policy is a
+data file, and a better one only makes the bots harder.
+
 ### Next, in order
 
 1. **Deploy.** Hangman, solitaire, trivia, UNO and hold'em are all played and all