safety security
Two Copies of One Agent Failed Together
A preregistered 18,000-mission test found same-model handoffs co-failed on 90 percent of missions where either agent failed.

Summary
A preregistered 18,000-mission test found same-model handoffs co-failed on 90 percent of missions where either agent failed.
Reliability estimates often multiply component success rates as if failures were conditionally independent. The reported two-agent tests show that assumption can over-credit redundancy when both agents share a model; switching models reduced association in all six registered contrasts, while changing vendor after changing model produced the registered null. The authors replace independence with a finite-sample linear-program certificate over measured co-execution moments. Adding four moment constraints narrowed the identified interval by 85.7 percent and raised its certified floor from 0.2455 to 0.4116.
Why it matters
A preregistered 18,000-mission test found same-model handoffs co-failed on 90 percent of missions where either agent failed.
Limits and context
No additional limitation was separately recorded.
Key claims
A preregistered 18,000-mission test found same-model handoffs co-failed on 90 percent of missions where either agent failed.
Evidence: source-2026-08-15-013
Sources
- arXiv preprint 2608.12895arXiv · primary research
Corrections
No corrections have been recorded for this story.