Expertise earns its edges

Experts often omit what has become automatic. They reconstruct a clean procedure after a messy event. They describe what should happen rather than what did happen. Different experts can agree on visible steps while disagreeing sharply about the hidden decision points.

The working paper on expert judgment documents the consequence: no elicitation method is a truth machine. Interviews, the Critical Decision Method, think-aloud, observation, artifact review and performance traces expose different slices of judgment. The practical response is triangulation—multiple incidents, experts, methods and held-out cases—not faith in one articulate account.

This essay reuses the paper’s expert-capture sequence at the point where validation begins: decision situation, cue, interpretation, expectation or action, and boundary or error. The sequence prevents a cue from being detached from the relation that made it useful. Even with provenance attached, it remains a hypothesis.

The companion essay, “What the expert sees before the student knows to look”, gives a revealing example. Pathology residents and attending physicians both located the critical region of a slide about 94% of the time. The trainees’ problem was often not looking in the wrong place; it was describing and interpreting what they saw incorrectly. “Look here” would have captured the location while missing the judgment (Brunyé et al., 2023).

Expertise is relational. A cue becomes useful because it changes an expectation, differentiates two cases, or alters what should happen next.

An expertise model can contain many heuristics. A particular episode may make some of them available, while the resulting reasoning visibly uses, adapts, rejects or coordinates only a subset. Availability is an opportunity condition. Enactment is an observation about use in a defined activity. Neither is learner ownership.

The distinction mirrors a familiar problem in expertise research. People and institutions possess many espoused rules: consult the customer, protect downside, follow the evidence. Their presence in a handbook or interview says little about when they govern a real decision. Comparison with concrete incidents, observations and held-out cases is needed to determine whether an account describes practice or only an idealized account of it.

Even validly detected enactment would establish only that a resource was used, adapted or rejected in that episode. Non-use can reflect irrelevance, restraint, efficiency or missing observability. Joint use of two heuristics supplies a co-use observation, not proof that the relation was correct, causal, expert, successful or learned. Those interpretations require other criteria.

A heuristic association is a probabilistic, revisable representation of accumulated evidence that reasoning resources are used in relation under stated observation conditions. At the public construct level, HeuriSight uses Beta–Bernoulli updating with time decay to represent uncertainty in such associations while allowing older observations to carry less influence. The update is a design choice for representing uncertainty, not a result about expertise, reasoning quality or learning.

The mathematics can discipline uncertainty. It cannot determine whether the observations were detected correctly, whether both resources were genuinely relevant, whether the same person selected them, whether their conjunction improved a decision or whether a learner could use them without assistance. Those are questions about the observation model, provenance, opportunity and independent criteria.

An association is therefore not a semantic truth, causal prerequisite or mastery relation. It earns an educational interpretation only when it predicts or explains something outside the observations that created it. Auditability is valuable because a representation can be inspected and revised. It is not a validity certificate.

If heuristics are to be tested rather than merely accumulated, cases and observations must be chosen so competing interpretations can fail. Near-neighbor cases can test where a heuristic should and should not apply. Predictions made before feedback can be compared with outcomes. Held-out cases can reveal whether the captured relation travels beyond the incidents that produced it. Learner tasks can then test whether instruction produced independent application rather than dependence on the representation.

Table 1. An expertise model earns stronger interpretations only through new evidence

Source and note: Authors’ construct synthesis of HS-WP-2026-02, HS-WP-2026-05 and HS-WP-2026-09. The rows are evidentiary burdens, not a product workflow or a report of completed HeuriSight results. Each stronger interpretation requires evidence not used to create the preceding representation.

Case-learning research shows why discrimination matters. In a management negotiation experiment, 88 master’s students formed 44 dyads and studied the same two cases. One week later, 14 of 22 comparison-trained dyads (64%) used the target contingency-contract strategy, versus 5 of 22 dyads (23%) who analyzed the protagonists separately, χ²(1, N = 44) = 7.503, p < .01; no confidence interval was reported (Thompson, Gentner, & Loewenstein, 2000). Overall monetary gain differed by only 2% and not significantly. The result supports explicit analogical encoding of one strategy, not a general advantage for any collection of cases.

The same discipline should govern a heuristic association. Co-occurrence is not relationship. Enactment is not success. Acceptance is not transfer. Each step requires its own evidence.

An “expert mind” metaphor encourages premature compression. It asks the system to synthesize several accounts into one confident voice. A testable model should often do the opposite: preserve who believed what, in which incident, under which constraints, and with what counterevidence.

Disagreement can reveal hidden boundary conditions. One expert may treat a cue as decisive in a high-reliability environment; another may discount it where the base rate differs. Averaging them can produce a rule no one actually uses. A versioned expertise model can preserve both claims, their provenance and the evidence gathered under each context.

This also protects instruction. Students should not be taught that expertise is universal certainty. They should see where experts agree, what their judgments depend on, and which evidence would make a reasonable expert change course. The case becomes a place to test a model of judgment, not a theatre for imitating authority.