SirStone 96065597c0 fix(divergence): guard reward-normalizer cold start; drop arctanh recovery ...

Welford variance-collapse divided by 1e-8 producing bit-exact +/−5e8 /
+−1.25e8 poisoned rewards into TD targets; guard skips normalization
until stats meaningful; tanh-inversion removal bounds log-prob path.

Fixes #60.

2026-08-24 01:03:04 +02:00

..

docs

fix(SAC_LSTM_Bot): un-invert dashboard axes — orientation selftest added

2026-08-23 10:47:48 +02:00

src

fix(divergence): guard reward-normalizer cold start; drop arctanh recovery

2026-08-24 01:03:04 +02:00

tests

feat(SAC_LSTM_Bot): campaign v2 levers — aggression/anti-ram reward shaping + stability knob overrides (part 2)

2026-08-22 19:44:58 +02:00

tools

fix(SAC_LSTM_Bot): un-invert dashboard axes — orientation selftest added

2026-08-23 10:47:48 +02:00

config.nims

feat(SAC_LSTM_Bot): main bot integration (#48)

2026-08-21 20:19:05 +02:00

make_twin.sh

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

2026-08-22 21:09:07 +02:00

SAC_LSTM_Bot.json

feat(SAC_LSTM_Bot): training harness (#49)

2026-08-21 21:51:27 +02:00

SAC_LSTM_Bot.nimble

feat(SAC_LSTM_Bot): project scaffold (#38)

2026-08-20 23:33:43 +02:00

SAC_LSTM_Bot.sh

feat(SAC_LSTM_Bot): training harness (#49)

2026-08-21 21:51:27 +02:00

sac_train.sh

fix(SAC_LSTM_Bot): MA history self-truncation — read before > redirect; campaign v2 attempt-3 prep (#60)

2026-08-23 07:02:36 +02:00