Measured 142,400 model calls on fifty coupled instances
Coupling a fleet adds at most 0.013 to error correlation, against +0.447 from shared weights; the rise in errors came from longer prompts.
Registration and verificationRegistered at commit 94d6a34 on 2026-09-13, before launch. Every figure recomputed from the raw ledgers and independently rescored. The registered positive control saturated; the support for the result comes from injection tests run afterwards in verification.
Fleets of AI agents increasingly read each other's output. If fifty copies of one model can see what the others just did, do they start making the same mistakes at the same time, the way a crowd herds? Error Independence showed that shared weights already correlate errors. This campaign asked how much coupling adds on top.
Fifty instances of one model, wired so that each reads a shared number the others write. The registration fixed seven hypotheses with numeric thresholds, three coupling definitions, two control arms, a positive control that drives every agent onto one commanded answer, a gate block on a tuning seed and a sealed held-out block of equal size to the primary, with a rule that no threshold moves after a reading.
142,400 model calls, a sealed held-out block of equal size, and zero error codes, scored for excess concordant error: how often instances went wrong together beyond what their individual error rates predict.
Coupling added +0.0040 excess error correlation (90% interval −0.0048 to +0.0124). The useful form of that result is a ceiling of about 0.013 at 90% confidence, far below the +0.447 that shared weights produce. Errors did rise, from 0.152 to 0.296, but they rose with prompt length: median prompt tokens went from 279 to 614, and a decoy roster of the same length reproduced 89 percent of the rise.
The registered positive control does not license that null. Arm X drives every agent onto one commanded answer, so the count wrong loses its variance and the correlation statistic returns its degenerate code, while the concordant-error excess reads exactly 0.0000 because observed and baseline agreement both saturate at 1.000. The warrant was built afterwards, in verification, by injecting a common shock into a known fraction of the real coupled ledger's ticks and running the frozen pipeline over it unchanged: an analytic +0.1922 came back as +0.1749, and the rule fires from about 0.18 upward.
| Excess error correlation added by coupling (90% interval −0.0048 to +0.0124) | +0.0040 |
|---|---|
| Ceiling on that excess at 90% confidence | about 0.013 |
| Error correlation from shared weights, for comparison | +0.447 |
| Error rate as median prompt length rose from 279 to 614 tokens | 0.152 to 0.296 |
| Share of that rise reproduced by a same-length decoy roster | 89% |
“Better questions lead to better worlds.”Dinand Tinholt