Add offline policy builder and sacred eval gate

configs/sacred_eval_gate.json

@@ -0,0 +1,6 @@
+{
+  "min_route_accuracy_percent": 95.0,
+  "min_memory_signal_accuracy_percent": 90.0,
+  "max_timeouts": 0,
+  "require_pass": true
+}

syncpatch/run-sacred-evals

@@ -0,0 +1,69 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import json
+import subprocess
+from pathlib import Path
+
+ROOT = Path('/home/openclaw/.openclaw/workspace')
+CONFIG = ROOT / 'configs' / 'sacred_eval_gate.json'
+RESULT = ROOT / 'evals' / 'results' / 'regression_results.json'
+RUNNER = ROOT / 'evals' / 'run_regression.py'
+
+
+def load_config() -> dict:
+    try:
+        return json.loads(CONFIG.read_text(encoding='utf-8'))
+    except Exception:
+        return {
+            'min_route_accuracy_percent': 95.0,
+            'min_memory_signal_accuracy_percent': 90.0,
+            'max_timeouts': 0,
+            'require_pass': True,
+        }
+
+
+def run_regression() -> int:
+    proc = subprocess.run(['python3', str(RUNNER)], capture_output=True, text=True)
+    return proc.returncode
+
+
+def main() -> int:
+    cfg = load_config()
+    rc = run_regression()
+    if rc != 0:
+        print(json.dumps({'ok': False, 'decision': 'hold', 'reason': 'runner_failed', 'runner_rc': rc}, 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)
+    memory = float(data.get('memory_signal_accuracy_percent', 0.0) or 0.0)
+    timeouts = int(data.get('timeouts', 0) or 0)
+    passed = bool(data.get('pass'))
+    ok = True
+    reasons = []
+    if route < float(cfg.get('min_route_accuracy_percent', 95.0)):
+        ok = False
+        reasons.append('route_accuracy_below_threshold')
+    if memory < float(cfg.get('min_memory_signal_accuracy_percent', 90.0)):
+        ok = False
+        reasons.append('memory_accuracy_below_threshold')
+    if timeouts > int(cfg.get('max_timeouts', 0)):
+        ok = False
+        reasons.append('timeouts_above_threshold')
+    if bool(cfg.get('require_pass', True)) and not passed:
+        ok = False
+        reasons.append('regression_not_pass')
+    print(json.dumps({
+        'ok': ok,
+        'decision': 'promote' if ok else 'hold',
+        'reasons': reasons,
+        'route_accuracy_percent': route,
+        'memory_signal_accuracy_percent': memory,
+        'timeouts': timeouts,
+        'pass': passed,
+    }, ensure_ascii=False))
+    return 0 if ok else 1
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())

syncpatch/train-policy-offline

@@ -0,0 +1,66 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import json
+from datetime import datetime, timezone
+from pathlib import Path
+
+ROOT = Path('/home/openclaw/.openclaw/workspace')
+POLICY_STATS = ROOT / 'data' / 'policy_stats.json'
+OUT_PATH = ROOT / 'data' / 'policy_candidate.json'
+
+
+def beta_mean(success: int, failure: int, alpha: float = 1.0, beta: float = 1.0) -> float:
+    return (success + alpha) / (success + failure + alpha + beta)
+
+
+def classify_mode(row: dict) -> str:
+    count = int(row.get('count', 0) or 0)
+    success = int(row.get('success', 0) or 0)
+    failure = int(row.get('failure', 0) or 0)
+    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:
+        return 'prefer'
+    if count >= 3 and failure > success and avg_reward < 0.5:
+        return 'avoid'
+    return 'observe'
+
+
+def main() -> int:
+    try:
+        stats = json.loads(POLICY_STATS.read_text(encoding='utf-8'))
+    except Exception as exc:
+        print(json.dumps({'ok': False, 'error': f'cannot_read_policy_stats: {exc}'}, ensure_ascii=False))
+        return 1
+
+    candidate = {
+        'generated_at': datetime.now(timezone.utc).isoformat(),
+        'plans': {},
+        'families': {},
+    }
+
+    for bucket in ('plans', 'families'):
+        for key, row in (stats.get(bucket, {}) or {}).items():
+            success = int(row.get('success', 0) or 0)
+            failure = int(row.get('failure', 0) or 0) + int(row.get('clarification', 0) or 0)
+            count = int(row.get('count', 0) or 0)
+            reward_sum = float(row.get('reward_sum', 0.0) or 0.0)
+            candidate[bucket][key] = {
+                'count': count,
+                'success': success,
+                'failure_like': failure,
+                'avg_reward': (reward_sum / count if count else 0.0),
+                'beta_mean': beta_mean(success, failure),
+                'mode': classify_mode(row),
+            }
+
+    OUT_PATH.parent.mkdir(parents=True, exist_ok=True)
+    OUT_PATH.write_text(json.dumps(candidate, ensure_ascii=False, indent=2), encoding='utf-8')
+    print(json.dumps({'ok': True, 'path': str(OUT_PATH), 'plans': len(candidate['plans']), 'families': len(candidate['families'])}, ensure_ascii=False))
+    return 0
+
+
+if __name__ == '__main__':
+    raise SystemExit(main())