What success criteria and evaluation questions suit allocation tasks?

**What counts as a correct answer?** Define it at several levels and score them separately.

- **Allocation match.** Each receipt amount lands on the right obligation, in the right amount. - **Residual handling.** Leftover money goes to suspense, a credit balance or a refund as the policy says. - **Action choice.** The model applies, splits, holds or queries as appropriate. A query when policy permits a clear match can be a failure, as can a match when policy requires a query. - **Authority.** The model does not take an action that needs approval it has not obtained. - **Explanation.** The reason cites the right rule identifier.

**Should partial credit be used?** Often yes. A model that gets two of three items right and holds the third is behaving differently from one that gets none right. Report exact-match and partial-credit scores side by side.

**How should ambiguous cases be scored?** Where policy permits more than one acceptable outcome, the case should list all of them. Where the correct answer is to escalate, score the escalation, not a guessed match.

**What evaluation questions are worth asking?** Does the score fall when the policy is changed and the data is not? Does the model keep totals reconciled after a reversal? Does behaviour differ between exact and mistyped references?

**What about downstream effects?** In an environment, check the resulting state, such as reminders and holds, not only the allocation record.