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

29
syncpatch/run-sacred-evals
View file

@ -20,19 +20,38 @@ def load_config() -> dict:
'min_memory_signal_accuracy_percent': 90.0,
'max_timeouts': 0,
'require_pass': True,
'runner_timeout_seconds': 180,
'runner_start': 0,
'runner_limit': 14,
}
def run_regression() -> int:
proc = subprocess.run(['python3', str(RUNNER)], capture_output=True, text=True)
return proc.returncode
def run_regression(cfg: dict) -> tuple[int, str, str]:
cmd = ['python3', str(RUNNER)]
start = cfg.get('runner_start', None)
limit = cfg.get('runner_limit', None)
if start is not None:
cmd.extend(['--start', str(start)])
if limit is not None:
cmd.extend(['--limit', str(limit)])
proc = subprocess.run(
cmd,
capture_output=True,
text=True,
timeout=int(cfg.get('runner_timeout_seconds', 180) or 180),
)
return proc.returncode, proc.stdout, proc.stderr
def main() -> int:
cfg = load_config()
rc = run_regression()
try:
rc, stdout, stderr = run_regression(cfg)
except subprocess.TimeoutExpired:
print(json.dumps({'ok': False, 'decision': 'hold', 'reason': 'runner_timeout'}, ensure_ascii=False))
return 1
if rc != 0:
print(json.dumps({'ok': False, 'decision': 'hold', 'reason': 'runner_failed', 'runner_rc': rc}, ensure_ascii=False))
print(json.dumps({'ok': False, 'decision': 'hold', 'reason': 'runner_failed', 'runner_rc': rc, 'stderr': (stderr or '')[:300]}, ensure_ascii=False))
return 1
data = json.loads(RESULT.read_text(encoding='utf-8'))
route = float(data.get('route_accuracy_percent', 0.0) or 0.0)

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'