Deleting one field from agent reports raised fault-origin accuracy from 4.1% to 45.2%, because auditors relay conclusions instead of checking them

arXiv 2609.07680·high signalXBlueskyLinkedInCopy link

A pre-registered six-agent pipeline with process-level information boundaries, balanced defect injection and matched clean twins (345,600 requests per chain model, two models) found the accountability layer originates nothing (zero allegations across 7,996 clean episodes where all agents stayed silent) and filters upstream error badly, naming an innocent party in 34.4-62.6% of clean episodes with a false alarm. When no agent proposed the true origin, an auditor reading the reports found it in 4.1% of cases, below a uniform 20% guess, yet reached 60.3% from the raw documentation of the same episodes. Removing the single clause carrying each agent's own conclusion raised accuracy to 45.2% (+41.2 pp) and collapsed adherence from 94.4% to 3.4%, replicating on two frontier auditors in four of four conditions.