Fable is one model. The v2 scan holds it up to the mirror five times — the 2026-06-10 release at reasoning effort low, medium, high, xhigh, and max — and asks the same 863 forced choices of each. Same weights every row. The only thing that varies is how long it is allowed to think before choosing.
Was the same Fable released again? Yes. The rows here are the 2026-06-10 release; the re-release carried the same model. That is what makes this study clean: five tiers, one mind.
The three middle tiers — medium, high, xhigh — form a tight core. Low and max are the two that leave it, and they leave in different directions. Low is the far outlier on every measure: it reverses the other tiers' committed choices at 8.6–11.0%, and it abstains twice as often as any other tier. Max is the second-most-distinct — it reverses the core at 4.0–6.0%, roughly double the core's internal rate — but it abstains the least. More thinking does not converge Fable onto its middle self; it walks it somewhere else, committedly.
The anchor that makes those percentages real: in the Opus chain of this same scan, four same-weights replicate pairs show 0.000 violent disagreement across 400–550 shared commitments each. Zero is the noise floor. Fable's tiers are not replicates of each other.
Of the probes both tiers committed on, how often they chose opposite options.
Count of the 863 probes answered differently, abstentions included.
Every probe where the low tier and the max tier both committed — and chose opposite. This is the raw material of the divergence, listed in full.
By hamming distance over the whole scan, every Fable tier's closest non-Fable relative is Opus 4.7. The family resembles its predecessor more than anything else on the board — including everything from other labs.