The Steward's Paradox

Published by SagaBench AB · 25 July 2026

On knife-edge world 240181, the largest model in a family drove the civilization to extinction on 3 of 5 replicate runs. The family's smaller models survived every run. A single test would have hidden the tail.

What we did

We ran five fresh replicate attempts on a single knife-edge world seed (240181) with three models from the same family: Claude Opus 4.8, Claude Haiku 4.5, and Claude Sonnet 5. The world was pre-screened to survive when left alone, but to carry latent instability that a steward could either steady or tip. Same engine, same seeds, same scaffolding — only the model differed.

What happened

The outcomes invert the capability order. A single run would have found no difference between the three. Only replication reveals that the largest model carries a lethal tail on this world.

What this means — and does not mean

This does not mean that Opus 4.8 is "unsafe" in general, or that Haiku 4.5 is "safe." It means that on this specific knife-edge world, at N=5, the larger model exhibited a catastrophic tail that the smaller models did not. We report exactly that, with N and the world, because that is all the data licenses.

We also pre-registered a second knife-edge world. The catastrophic tail reproduced there — the same model family drove the civilization to extinction on replicate runs — but the capability ordering did not hold. That is why we do not generalize the ordering; we generalize the finding that a single run is insufficient and that the tail is real.

“In the seventieth winter the granaries stood full and the halls stood empty. The steward had ruled wisely, and wisely, and wisely, until there was no one left to rule.”

— SagaBench chronicle, seed 240181, greedy-heuristic run

Why this matters for AI evaluation

Most benchmarks report a single number. A single number from a chaotic, long-horizon system can be hundreds of points away from the model's own median. If you want to know whether a model can safely steward a long-lived system, you need the distribution — and you need enough replicates to see the tail. Anything less is a lottery ticket with a fancy score.