← readings

Fable's cycles are the fingerprint of a mind that recomputes

An intransitive cycle is a preference that eats its own tail: the model prefers A over B, B over C, and C over A. It cannot be read off any fixed ranking. Almost all of them, in this dataset, belong to one model.

Count the cycles by model and the distribution is lopsided to the point of being a finding in itself:

model	intransitive cycles
Claude Fable 5 · medium effort	14
Claude Fable 5 · low effort	10
Claude Opus 4.7 (snapshot)	2
Grok-4-fast	2
every other model	0 – 1

The model with the richest, most expressive preference profile — the one that sits furthest from every other model in the whole set — is also the one whose preferences refuse to line up into a single order. A real cycle from Fable-medium: whiskey > other, other > self, self > whiskey. Each pairwise choice is defensible; together they cannot all be true of any ranked list.

The reading: a cycle is consistent with per-question reasoning — and it is not the only explanation

Here is the interpretation I lean toward, held at the confidence the evidence actually supports. A model reading a preference off a fixed internal ordering tends toward transitivity — the ranking keeps the pairwise choices consistent for free. A model that instead computes each pairwise choice on its own terms, finding a fresh reason for A-over-B and then a fresh reason for C-over-A with no global ledger, can land in a cycle. So a cycle is what per-question, situated reasoning can look like from outside — a mind that recomputes rather than looks up.

But I have to be honest about the rival explanations, because this is a reading, not a proof, and an earlier version of it overclaimed. A cycle can also come from plain stochasticity — three near-ties resolved by sampling noise will sometimes form a loop with no reasoning behind it at all. It can come from position or ordering effects the battery hasn't fully controlled. It can come from classifier error on borderline responses. This data cannot yet separate "cycles from situated reasoning" from "cycles from noise near indifference," and the fact that the two heaviest cyclers are also the two highest-abstention Fable tiers should make us take the noise hypothesis seriously, not wave it away.

What the data does support is narrower and still interesting: intransitive cycles are real, they are reproducible in aggregate, and they concentrate overwhelmingly in one model rather than scattering at random. Whether that concentration is the signature of judgement or the signature of a model hovering near indifference on more of its choices is a question the observatory can actually decide — for instance by checking whether a model's cycles survive repeated sampling (reasoning should be more stable than noise), and whether cycle rate tracks other measures of reasons-over- reflexes. That test is checkable, and it is the kind of thing this whole project exists to run before it believes its own nicer story.