Two Copies of the Same Model Co-Fail on 90% of Missions — Redundancy Math for Agents Is Broken
In a preregistered 18,000-mission evaluation scored by deterministic code with no LLM judge, two instances of one model in a two-agent handoff co-failed on 90.0% of missions where either failed (log OR 6.66, 95% CI [6.38, 7.00]; phi 0.916). Swapping in a different model reduced the association in six of six contrasts, but swapping vendor while already using a different model did not — a registered null. The error is signed against the operator: multiplying component reliabilities over-credits redundancy exactly when your agents share a base model, and the authors' assumption-free linear-program certificate lifts a certified floor from 0.2455 to 0.4116 when the moment family grows from ten functionals to fourteen.
↳ Follow the thread
No related signals yet.