On Friday the U.S. government directed Anthropic to suspend access to Claude Fable 5. I'd spent that same day running it — not casually, but as a controlled experiment I'd set up in advance.
Two bounded, pre-registered runs. I fixed the success criteria before the model touched anything, made it leave a provable evidence trail, and then re-verified every claim myself in a clean sandbox. The point was never to watch an AI do impressive things. It was to find out whether the work was real, and to be able to prove the answer either way.
What it did, on real code: it rebuilt a continuous-integration board that actually reproduces, converged a hashing inconsistency across three separate codebases — proven by identical digests — and restored three corrupted source files, one of them byte-for-byte against a hash I had frozen before the corruption existed. An integrity check it added partway through then caught three more corrupt files I didn't know I had, out of eight hundred and thirteen. Every figure in the write-up resolves to a content hash, so you don't have to take my word for any of it.
I want to be careful about what this is and isn't. It's two runs and one operator — not a verdict on the model's general ability, and I don't claim one. The suspension is contested and, by Anthropic's account, possibly temporary; that part isn't mine to litigate. What I can do is leave a clean, reconstructable record of what the model actually did while it was here.
This is the same order the trade taught me. You wire it, then you test it, then you energize it — in that order, every time, because what's behind the wall has to do what the paperwork says. I pointed the same discipline at a model instead of at code. Calibrated honesty, applied to the thing doing the work.
The preprint and the full data are open. If you want to argue with a number, the number is public.
→ The Fable 5 case study (preprint + data, CC BY) · the code · the methodology · the experience report · the self-audit · the math case study