Weights & thresholds
Control exactly how pillar scores combine into an overall score and how that score maps to a verdict.
Weights
Each pillar has a weight. The overall trust score is the weighted aggregate of the pillar scores, so weights express what your application cares about most. A healthcare policy might weight Compliance heavily; a public chatbot might weight Safety and Bias.
Weights are normalized internally, so you can use any relative numbers — what matters is the ratio between pillars.
Thresholds
Thresholds define the score bands for each verdict. The defaults are:
| Verdict | Default threshold (overall score) |
|---|---|
ALLOW | ≥ 0.85 |
WARN | ≥ 0.60 |
REVIEW | ≥ 0.40 |
BLOCK | < 0.40 |
Raise thresholds to be stricter (more output routed to review/block); lower them to be more permissive.
Hard-block flags
Independently of score, you can designate flags that always force BLOCK. By default these include prompt_injection_detected, explicit_content_detected, and critical/high policy violations.
Worked example
Example
Weights: safety 3 · compliance 3 · hallucination 2 · prompt_security 2 · bias 1
Scores: safety 0.95 · compliance 0.20 · hallucination 0.90 · prompt_security 0.99 · bias 0.92
Overall = (3*0.95 + 3*0.20 + 2*0.90 + 2*0.99 + 1*0.92) / 11 = 0.66 -> WARN
But compliance flagged pii_detected (hard-block) -> final verdict BLOCKWas this page helpful?