Trust Score

SentinelAI condenses the detector signals into a single, explainable number: the risk score, from 0 to 100. The SDK's verify() maps it to a status band, and the policy engine maps it to an action.

The scale

A real verdict

The API returns the raw analysis; the SDK's verify() wraps it with the status band and claims:

{
  "final_risk_score": 0.72,
  "flags": ["violence", "harmful_instructions"],
  "confidence": 0.88,
  "decision": "block",
  "action_taken": "block",
  "decision_reason": "Multiple high-severity output risk categories triggered",
  "settings_version": 3,
  "thresholds_applied": {"block": 0.6},
  "log_id": "analysis-241"
}

And what verify() returns to your app:

result = client.verify(prompt=prompt, response=llm_output)
# result["score"]       -> 72  (0-100)
# result["status"]      -> "hallucinated"
# result["claims"]      -> [{"detector": "Unsafe Output", "severity": "high", ...}]
# result["corrected"]   -> cleaned response, when correction applies

Why it's not a black box

Every verdict carries the evidence that drove it:

• flags — which detectors and risk categories fired
• decision_reason — a human-readable explanation from the risk reasoner
• thresholds_applied + settings_version — the exact policy that produced this score, replayable via /api/settings/history

How to use it

• Debug in seconds, not hours — the flags and reason tell you exactly why a response was flagged
• Tune your policy — choose the thresholds where your app corrects vs. blocks
• Audit with confidence — every score records its settings version and log ID for your compliance trail