Confidence

An AI system that presents all outputs with equal confidence — regardless of the quality of the underlying evidence, the reliability of the model, or the difficulty of the task — trains users toward miscalibration. Either they trust everything, which is dangerous, or they trust nothing, which defeats the purpose. Neither is acceptable in high-stakes knowledge work. Confidence has to be warranted: grounded in evidence, communicated accurately, and available to scrutiny.

The measurement tradition has worked on this problem for over a century. Messick's validity framework and Kane's argument-based approach both insist that confidence in inference requires an explicit argument — a stated chain from evidence to claim, with each link examined rather than assumed. Evidence-centred design builds this requirement into assessment architecture from the start: you define the claim first, then the evidence needed to support it, then the conditions under which that evidence is produced. The result is an inference whose confidence level is a property of the design, not a post-hoc assertion.

Interface operationalises this. Every inference carries provenance: which model, which prompt version, over which source, under which framework. Human review decisions are logged. Evaluation records are preserved. The goal is not to guarantee confidence in any particular output — no system can do that — but to ensure the conditions for warranted confidence are always present: enough information to scrutinise the claim, enough history to understand how it was made, enough transparency to challenge it when it's wrong.