SirStone/SirRoboGarage
1
0
Fork 0
Code Issues 28 Pull Requests Actions Projects Releases Wiki Activity

fix(SAC_LSTM_Bot): freeze mirror-twin via eval-mode gate in twin launcher

Browse Source 
SACLSTM_EVAL_MODE=1 in SacTwin.sh suppresses all sendTrainingMsg traffic
(lever-4 gate), so the twin never trains — not even in-RAM within a battle.
Required now that the main bot's SACLSTM_SAVE_INTERVAL drops to 1 (v2 relaunch
after checkpoint-cadence diagnosis): without the gate the twin would persist
per-battle drift and stop being the frozen reproducible opponent #54 specifies.
This commit is contained in:
Davide Cappellini
2026-08-22 21:09:07 +02:00
parent bd58794b4c
commit 19f34abf0c
1 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
SAC_LSTM_Bot/make_twin.sh
+5
View File
@@ -46,6 +46,11 @@ cat > "$TARGET/SacTwin.sh" <<'EOF'
DIR="$(cd "$(dirname "$0")" && pwd)" DIR="$(cd "$(dirname "$0")" && pwd)"
export SACLSTM_BOT_JSON="$DIR/SacTwin.json" export SACLSTM_BOT_JSON="$DIR/SacTwin.json"
export SACLSTM_WEIGHTS_PATH="$DIR/weights/sac_latest.zip" export SACLSTM_WEIGHTS_PATH="$DIR/weights/sac_latest.zip"
# Freeze contract (#54): the twin is a FROZEN sparring partner, not a
# co-learner. Eval-mode gate (lever 4) suppresses all sendTrainingMsg traffic,
# so the twin never trains — not even in-RAM within a battle. Without this,
# any main-bot checkpoint-interval change would let twin drift accumulate.
export SACLSTM_EVAL_MODE=1
exec "$DIR/SAC_LSTM_Bot" exec "$DIR/SAC_LSTM_Bot"
EOF EOF
chmod +x "$TARGET/SacTwin.sh" "$TARGET/SAC_LSTM_Bot" chmod +x "$TARGET/SacTwin.sh" "$TARGET/SAC_LSTM_Bot"