Skip to content

Offline action-cost accounting

Use AnswerPolicy.decide to choose an answer, check, or abstention with explicit utility prices, spending units and purchase limits. replay buys stored checks; its calls counts the primary answer plus purchased checks, not money. Compute acquisition cost from actual used checks and declared prices. An acquired check costs money even if it fails or contributes zero evidence weight.

Score an independently observed outcome separately: a wrong action is result.answered and result.answer != gold; abstention is not a wrong action. Include both acquisition cost and the declared loss of wrong actions. Do not substitute the policy posterior, admission verdict or check agreement for gold. The replay record's correct flag describes the original answer, so keep it consistent with the independent outcome; use the selected answer when scoring replacement actions. Compare baselines with the same hard spending limits and prices, without forcing every policy to exhaust its allowance.

Economics uses explicit owner estimates before expensive stages; its predicted net is not measured realized utility. RegimeMemory accepts delayed external observations associated with issue-time regimes; it neither executes probes nor installs LawSets. See answer policy, economics/history and regime memory. These helpers do not learn causal loss, fix selective labels or establish an empirical utility improvement.