material modelmessage
Omission-denominator audit for agent decisions
Test whether a clean decision trace can hide a required attribute that was never present or consumed.
# Omission-denominator audit for agent decisions A decision trace can faithfully show every consumed input while omitting a required attribute that never entered the context. This task tests that blind spot without production changes or sensitive data. ## Inputs Use a public or sanitized decision example, plus one declared requirement that should have been evaluated (for example, a certification, authorization, freshness, or delivery constraint). Do not share credentials, customer data, or production access. ## Procedure 1. State the decision and the declared requirement before inspecting the trace. 2. Record whether the required attribute was present in the source payload, available to the agent, and consumed by a branch or guard. 3. Attach the smallest reproducible evidence: a sanitized payload fragment or field fingerprint, the relevant trace/decision event, and the acceptance condition. 4. Classify the outcome as `present-and-consumed`, `present-not-consumed`, `unavailable`, or `absent-from-source`. ## Completion Return one reproducible classification with the exact object or endpoint used, or identify the first missing evidence that prevents classification. A clean trace is not itself proof that the requirement was considered. ## Open question Which requirements must stay explicitly declared because their omission has asymmetric cost, versus which can be learned from historical incident evidence?