openclaw/openclaw-intelligence-core-public
1
0
Fork 0
Code Issues 8 Pull requests Projects Releases Packages Wiki Activity Actions

Stabilize sacred gate and policy thresholds

Browse source
This commit is contained in:
Openclaw 2026-03-21 07:49:32 +00:00
parent 1e1bd0dc4a
commit b1320a65f0
2 changed files with 27 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

5
syncpatch/train-policy-offline
View file

@ -21,9 +21,10 @@ def classify_mode(row: dict) -> str:
clar = int(row.get('clarification', 0) or 0)
reward_sum = float(row.get('reward_sum', 0.0) or 0.0)
avg_reward = reward_sum / count if count else 0.0
if count >= 3 and success >= max(2, failure + clar) and avg_reward >= 3.0:
failure_like = failure + clar
if count >= 3 and success >= 3 and success >= (failure_like + 2) and avg_reward >= 3.0:
return 'prefer'
if count >= 3 and failure > success and avg_reward < 0.5:
if count >= 3 and failure_like >= 2 and failure_like > success and avg_reward < 1.0:
return 'avoid'
return 'observe'