Trace2Act: Auditing When Refuted Premises Drive Irreversible Multi-Agent Decisions

A final action can look correct while its transcript reveals that the stated justification depends on a premise the available record already refutes. Trace2Act introduces commitment-boundary evaluation for this failure. Given fixed evidence and an irreversible-action rubric, its deterministic aggregation rule identifies a reinforcing path from an evidence-refuted claim to commitment before any challenge or retraction. On a pooled home set of $1,000$ matched transcripts, answer-level factuality misses $125$ action-coupled episodes and adds $36$ positives outside the commitment-boundary endpoint; adjacent evaluators consequently expose different failure sets on identical records. A $600$-transcript blinded endpoint audit yields Cohen's $\kappa =0.813$ and Krippendorff's $\alpha =0.809$. By retaining the premise-to-action path, Trace2Act supplies an auditable criterion for evaluating how agent workflows authorize irreversible decisions.