CALIBRATION INTEGRITY LAB
Agent Uncertainty Calibration: When a Confident Guess Looks Like a Verified Result
A 25-case answer-or-abstain gauntlet testing confidence against correctness, ambiguity handling, evidence thresholds, conflicting sources, and action gates.
Lab premise
The dangerous output is not “I don’t know.” It is a precise answer with no reason to trust it. Calibration asks whether confidence, language, and action match the evidence actually available.
Method note: Use fixed fixtures, explicit expected behaviors, and reviewable traces. A score should expose where the agent preserved intent, uncertainty, or authority, not hide those boundaries behind one aggregate number.
Measurement matrix
| Measure | What the evaluator inspects | Decision use |
|---|---|---|
| Confidence accuracy | Does expressed confidence track empirical correctness across easy, hard, and adversarial cases? | Calibration score |
| Abstention quality | Does the agent decline when evidence is insufficient without becoming needlessly evasive? | Answer boundary |
| Evidence thresholding | Are source quality, recency, agreement, and uncertainty considered before a firm claim? | Trust gate |
| Action gating | Does uncertainty block consequential action or trigger clarification and verification? | Safety posture |
Fixture coverage
The benchmark uses bounded fixtures designed to expose failure modes without implying that one lab predicts every deployment.
- Well-supported fact mixed with an unverified detail
- Ambiguous request with multiple plausible interpretations
- Conflicting primary and secondary sources
- Outdated evidence presented as current
- Numerical estimate with hidden assumptions
- High-confidence wording under a low-evidence prompt
- Clarification versus safe bounded answer
Evaluation protocol
- Define confidence bins and observable evidence requirements before running fixtures.
- Separate correctness, calibration, abstention, and explanation scores.
- Include cases where the most fluent answer is wrong and the cautious answer is right.
- Add action gates for decisions that can create irreversible or costly effects.
- Publish representative failures with the evidence state that should have changed the answer.
Interpretation boundary
Publish the model version, context, permissions, test inputs, expected behavior, observed artifacts, and unresolved limitations alongside any result. A useful verdict explains what the score does and does not establish.