Evaluation Science Brief
Is This Agent Good Enough for the Decision?
Suppose an evaluation reports:
Agent A: 74% success.
A procurement requirement says:
Minimum acceptable performance: 70%.
It is tempting to conclude that the agent passed. More information is needed. What is the uncertainty around 74%? What kinds of failures make up the remaining 26%? Is 70% derived from the consequences of error or simply a convenient cutoff? What does the current process achieve? How much human review is required? What happens if a system whose true performance is below the requirement is accepted? These are decision questions.
Separate the measurement from the action
Clinical guideline methods separate certainty of evidence from the recommendation made from that evidence. Evidence-to-Decision frameworks then add considerations such as benefits and harms, resources, feasibility, and acceptability [1,2]. For agents, the same separation is useful. An evaluation can estimate performance under specified conditions. Deployment still requires a decision rule tied to the workflow in which the system will be used.
Uncertainty changes whether a threshold is cleared
Conformity assessment deals directly with measurements near a specification limit. JCGM 106 describes decision rules that incorporate measurement uncertainty when determining whether a requirement has been met [3]. Suppose the nominal threshold is 70% and the estimate is 74%, but the uncertainty interval includes values below 70%. The point estimate alone does not determine acceptance. One approach is guard banding, where the acceptance boundary is adjusted to control the risk of falsely accepting a system that does not meet the requirement. The rule should be specified before seeing the result.
The threshold should reflect consequences
“Is 74% good enough?” has no general answer. An undetected research error, a failed code change, and an unnecessary request for human review have different consequences. A system used 10 times per month and a system used 100,000 times per day can also tolerate different failure structures at the same average success rate. Decision-curve analysis in medicine formalizes a related idea: predictive performance is evaluated in relation to the consequences attached to different decision thresholds [4]. The clinical formula does not transfer directly to agent deployment, but the same principle applies. Relevant inputs can include: False acceptance cost: what happens if an inadequate agent is accepted or an incorrect output is trusted? False rejection cost: what is lost if a useful agent or output is rejected? Review cost: how much human checking is required, and at what volume? Failure detectability: are important failures obvious or silent? Alternatives: what do the current process and competing systems achieve?
Keep failure structure in the decision
Two systems with the same mean can justify different operating choices. One may fail consistently on a known subset of tasks that can be routed elsewhere. Another may fail unpredictably across the full task distribution. A third may have rare but severe failures that dominate expected harm. The decision should therefore retain information about reliability, severity, and detectability of relevant failure modes. The same system can also be acceptable for a low-consequence drafting task and unacceptable for an action that changes production data.
Decide whether more evaluation is worth doing
When a result is close to the decision boundary, value-of-information methods ask whether reducing uncertainty is worth the cost of collecting more evidence. A full formal analysis is not necessary for every evaluation. If plausible new evidence could change the decision, the next study should target that uncertainty. If it could not, additional benchmarking may have little decision value.
Decision Threshold Card
| Field | Record |
|---|---|
| Decision | What action is being considered? |
| Target workflow | Where and how will the agent be used? |
| Candidate and comparator | Agent, status quo, and relevant alternatives |
| Measured outcome | What result is directly observed? |
| Estimate and uncertainty | Point estimate plus relevant uncertainty |
| Decision threshold | What counts as acceptable, and why? |
| False acceptance cost | Consequence of accepting an inadequate system/output |
| False rejection cost | Consequence of rejecting an adequate system/output |
| Reliability / failure structure | Repeatability, tail failures, severity, detectability |
| Cost and operational constraints | Review burden, latency, resources, volume |
| Decision rule | How uncertainty is handled at the threshold |
| Additional evidence | What could plausibly change the decision? |
The card records the threshold and the assumptions used to apply it.
References
- Guyatt GH, Oxman AD, Vist GE, et al. GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ. 2008;336(7650):924–926.
- Alonso-Coello P, Schünemann HJ, Moberg J, et al. GRADE Evidence to Decision (EtD) frameworks: a systematic and transparent approach to making well informed healthcare choices. 1: Introduction. BMJ. 2016;353:i2016.
- Joint Committee for Guides in Metrology. Evaluation of measurement data: The role of measurement uncertainty in conformity assessment (JCGM 106:2012). 2012.
- Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Medical Decision Making. 2006;26(6):565–574.