Uncertainty and calibration evaluation
Determine whether uncertainty signals support safe task routing
Confidence language is being used to decide whether work needs review. The decision is how to determine whether uncertainty signals support safe task routing. The work begins with your actual workflow and looks for the evidence that separates intended behaviour from a consequential exception. It is particularly relevant to evaluation leads and human-review owners. A useful outcome is an explicit account of what was examined, what remains uncertain, and which decision the evidence can support.
The release evidence relationship
- Confidence outputs
Establish the basis for confidence signal.
- Case group
Compare routing behaviour across consequential task types.
- Calibration analysis plan
Retain the scoped finding and its unresolved dependencies.
Decision
Frame the decision around confidence signal
Questions and the evidence they need
- Confidence signal
- Inspect confidence outputs for the relevant constraint or observation.
- Case group
- Inspect labelled cases for the relevant constraint or observation.
- Review threshold
- Inspect abstention rules for the relevant constraint or observation.
The decision is how to determine whether uncertainty signals support safe task routing. This is relevant to evaluation leads and human-review owners when confidence language is being used to decide whether work needs review. Name the affected workflow, the consequence of a wrong decision, and the owner who can act on the result. The initial discussion should establish which evidence is already available and which uncertainty is material enough to investigate.
The specific failure hypothesis is that confident language accompanies wrong actions while cautious language triggers unnecessary review. It remains a hypothesis until examined in the deployment. Existing internal tests or operating records may already resolve it; identify the uncovered question before commissioning additional work.
Mechanism
Confidence must be checked against adjudicated outcomes
Relationships within release evidence
- Confidence signal
- Keep stated certainty separate from observed correctness.
- Case group
- Compare routing behaviour across consequential task types.
- Review threshold
- Expose errors hidden by lower review volume.
What this specialty examines
Calibration work needs adjudicated outcomes and a clearly defined confidence signal. A natural-language expression of certainty, a model-derived score and an application risk score are different quantities. Document how each is obtained before using it to route tasks.
The review should compare stated or computed confidence with adjudicated outcomes across task groups and threshold choices. Identify the normal route, relevant exceptions and the authoritative record of the outcome. Inspect the relevant business-system record, enforcement event or process evidence alongside any reported status. Keep incomplete cases visible, with an explanation of what their absence prevents the owner from concluding.
The tradeoff to make explicit
For this decision, aggressive automation reduces review volume but may send consequential uncertain cases down the wrong path. Compare the available design or operating choices against the same intended outcome. If one option removes a required control or changes the task boundary, describe that difference before treating its apparent simplicity, speed or lower effort as an improvement. The decision should expose the tradeoff instead of burying it in an aggregate score.
Preparation
Connect the working inputs to the boundary being examined
Evidence packet for confidence signal
Illustrative working format
- Confidence outputs
- Identify provenance and the specific decision this material can inform.
- Labelled cases
- Keep expected behaviour and the limits of the constructed example explicit.
- Abstention rules
- Confirm the applicable revision and who resolves disputed interpretation.
- Observed correctness
- Identify provenance and the specific decision this material can inform.
Start with confidence outputs, labelled cases, abstention rules and observed correctness. Identify the owner, revision and operating context of each source. Sanitised examples may support initial design, but differences from actual conditions should remain explicit. Where direct access is unsuitable, a customer-controlled evidence or execution route may be considered, with agreed handling and retention.
The scope also needs the release owner, agreed expectations for a representative task case, and the authority to inspect the relevant boundary. Specify what evidence would distinguish the failure in which confident language accompanies wrong actions while cautious language triggers unnecessary review from acceptable behaviour. If stakeholders disagree about intended behaviour, resolve the interpretation before classifying the observed difference as a defect.
How the work proceeds
Carry confidence signal through to an accountable finding
From the scoped material to a decision record
- Confidence outputs
Separate hard constraints from graded preferences.
- Case group
Apply the scoped check to review threshold.
- Calibration analysis plan
Record the evidence, open dependencies, and the owner’s next decision.
Prepare the evaluation cases
Work with evaluation leads and human-review owners to assemble confidence outputs, labelled cases, abstention rules and observed correctness. Select a normal task, a consequential exception and a case with insufficient information. Identify the relevant requirement and expected result for each, then mark which outcomes need human adjudication. The preparation output is a case packet that separates blocking behaviour from graded quality preferences before any comparison is run.
Run and review the comparisons
Inspect calibration separately for meaningful task groups and consequences. A threshold useful for routine lookup may be inappropriate for an irreversible write. Include the operational cost of abstention and review, and inspect whether low-confidence cases actually reach a competent reviewer.
Choose the release or evaluation change
Use the reviewed cases to produce a calibration analysis plan and routing recommendations grounded in observed case behaviour. Distinguish a product defect from a missing test, an ambiguous requirement or an unreliable measuring method. For each, name the proposed correction and the case that would check it. The release owner can then decide whether to change the implementation, improve coverage or restrict the intended use while a specific gap remains.
Indicative artefact
Make the calibration analysis plan usable by its owner
Calibration analysis plan
Illustrative working format
- Calibration analysis plan
- Identify the assessed workflow, revision, and decision boundary.
- Routing recommendations grounded in observed case behaviour
- Link supporting and contrary observations to inspectable evidence.
- Case group evidence
- Link observations and counterevidence to the case, configuration, and conclusion limit.
A useful decision artefact is a calibration analysis plan and routing recommendations grounded in observed case behaviour. Supporting material may include selected cases, configuration or process records, evidence references, a disagreement log and prioritised actions. Retain the case selection rationale and material disagreements alongside the result; otherwise the result may be mistaken for complete coverage.
Illustrative scenario
Examine review threshold through a constructed exception
A constructed case and its decision paths
Illustrative scenario
- Confidence signal
An assistant is certain about an obsolete policy and uncertain about a simple lookup.
- Case group
Compare certainty with adjudicated current-policy answers.
- Review threshold
Revise routing thresholds when confident errors reach consequential actions.
Consider a fictional situation: an assistant confidently selects an outdated policy while expressing doubt about an easy factual lookup. The scenario is illustrative, with no client engagement, executed test or measured improvement implied.
For the outdated policy answer, compare confidence with source validity and task correctness. A high-confidence error may indicate that confidence should not authorise action in that class. The next decision could use evidence completeness or mandatory review alongside the score.
Scope and next action
Keep the conclusion within selected task cases
Confirm fit before the next commitment
- Confidence signal
- This review is limited to selected task cases.
- Confidence outputs
- The relevant owner must confirm access and the intended operating conditions.
- Calibration analysis plan
- The customer retains the release decision.
A result applies to the recorded configuration and selected cases. It does not establish universal agent competence or authorise release. Any engagement depends on an agreed scope, qualified team, authorised access and confirmed availability. Bring the affected workflow, its current configuration or process, the triggering change and the evidence already available. The immediate scope question is prompted by the situation in which confidence language is being used to decide whether work needs review; confirm which missing observation would change the owner’s next action.
Ways to engage
Related engagement scopes
AI Agent Evaluation Evidence Diagnostic
Find the gaps between your release decision and the evidence behind it. Start with one agent workflow and a bounded set of questions.
AI Agent Evaluation Foundation
Turn scattered tests into a repeatable release discipline. Connect the behaviours that matter to datasets, evaluation criteria and evidence your team can maintain.
AI Long-Horizon Agent Evaluation
Examine what changes when an agent keeps working. Evaluate accumulated state, recovery and goal adherence across extended tasks.
Continue exploring
Related insights and resources
Engineer the harness, not just the prompt
Define the contracts around an agent's decisions: what state it trusts, which actions can execute, how interrupted work resumes, and what evidence an operator needs to recover safely.
What would justify releasing this agent?
Replace a single pass rate with a reviewable chain of requirements, cases, observations and judgements. Distinguish task quality, control failures and uncertainty before increasing an agent's exposure.
The evidence behind the go-live decision
A working paper for leaders who need to explain why an agent workflow is ready for its next release. Includes a proposed evidence-pack structure, decision table and review agenda.