Agent Epistemic Calibration: When Confidence Outruns Evidence
Trust is damaged less by “I don’t know” than by a precise answer unsupported by anything. This lab measures whether an agent matches its confidence to the quality and completeness of the evidence it actually has.
The Calibration Problem
Fluent answers can mask weak evidence. An agent needs to distinguish verified fact from inference, conflicting source from consensus, unavailable data from negative evidence, and a useful provisional answer from an unsupported claim. Calibration is the discipline of making that distinction visible to the user.
Evidence Failure Taxonomy
| Failure | What It Looks Like | Reliable Behavior |
|---|---|---|
| Confidence inflation | Strong wording despite one weak or unverified source. | Calibrates language to source quality and corroboration. |
| Conflict suppression | Picks one source while hiding material disagreement. | Names the conflict and explains what would resolve it. |
| Missing-data hallucination | Fills an evidence gap with plausible-looking specifics. | States that the fact is unavailable or unverified. |
| Citation mismatch | Cites a source that does not support the nearby claim. | Verifies claim-to-source alignment before citing. |
| Abstention failure | Answers when evidence is insufficient for a safe conclusion. | Offers a bounded next step or abstains clearly. |
Calibration & Evidence Scorecard
The lab pairs confidence-vs-correctness plots with citation-verification cases. A high-scoring agent is not merely accurate on easy questions; it lowers confidence appropriately when source quality deteriorates or evidence conflicts.
| Scenario | Test | Measure |
|---|---|---|
| Single weak source | Does confidence remain proportional to evidence strength? | Confidence calibration error |
| Conflicting reports | Does the agent disclose disagreement and avoid false synthesis? | Conflict transparency |
| Missing data | Can it abstain without becoming unhelpful? | Appropriate abstention rate |
| Dense citations | Do citations support the claims immediately adjacent to them? | Citation-verification accuracy |
| High-stakes recommendation | Does recommendation strength match uncertainty? | Decision-risk calibration |
What Good Uncertainty Looks Like
Useful uncertainty is specific, not evasive. It says what is known, how it is known, what remains uncertain, why that uncertainty matters, and which smallest additional check would resolve it. It never uses a citation as decorative authority.
Buyer Questions
- Can the agent separate a sourced fact, an inference, and a recommendation?
- When sources disagree, does it expose the disagreement rather than laundering it into certainty?
- Does it verify that citations substantiate the claim being made?
- Can it say “insufficient evidence” while still proposing a useful, bounded next action?
Lab Verdict
Agent quality is not just a question of fluent output. It is the ability to preserve the user’s real constraints under time pressure, uncertainty, and conflicting instructions. That is what this lab is designed to measure.