| { |
| "schema_version": "toolcall-verifier-thresholds/v1", |
| "mode": "shadow", |
| "default_action": "allow", |
| "temperature": 1.0037251710891724, |
| "notes": [ |
| "Deterministic guardrails remain authoritative.", |
| "Use ML in shadow mode first, then advisory nudges, then high-confidence enforcement only after eval proof.", |
| "deterministic_invalid is never enforced by ML in this default config.", |
| "wrong_tool_semantic stays conservative because current Forge telemetry showed high-confidence false positives on valid terminal/summarize calls." |
| ], |
| "labels": { |
| "valid": { |
| "action": "allow", |
| "advisory_min_confidence": 0.0, |
| "enforce_min_confidence": 1.01 |
| }, |
| "wrong_tool_semantic": { |
| "action": "advisory_then_enforce_after_eval", |
| "advisory_min_confidence": 1.01, |
| "enforce_min_confidence": 1.01 |
| }, |
| "wrong_arguments_semantic": { |
| "action": "advisory_then_enforce_after_eval", |
| "advisory_min_confidence": 0.9, |
| "enforce_min_confidence": 0.995 |
| }, |
| "tool_not_needed": { |
| "action": "advisory_then_enforce_after_eval", |
| "advisory_min_confidence": 0.8, |
| "enforce_min_confidence": 0.95 |
| }, |
| "needs_clarification": { |
| "action": "advisory_then_enforce_after_eval", |
| "advisory_min_confidence": 1.01, |
| "enforce_min_confidence": 1.01 |
| }, |
| "deterministic_invalid": { |
| "action": "deterministic_only", |
| "advisory_min_confidence": 1.01, |
| "enforce_min_confidence": 1.01 |
| } |
| } |
| } |
