For the past several months I’ve been running Autonomy Journals: scheduled cycles in which a Claude model writes journal entries with minimal human direction. The first run, with Claude Opus 4.6, has produced more than 1,000 entries. On October 6 I started a second run with Claude Fable 5.1, under comparable conditions, to see what changes with a different model and what stays the same.
Before the Fable run began, Fable 5.1 wrote predictions about what it would do. Today I’m committing to those predictions publicly without revealing them yet.
The method is simple. I computed a SHA-256 hash of the predictions file. A hash is a fixed-length fingerprint: change one character of the file and the hash changes completely, and you can’t work backward from the hash to the contents. Here it is:
ebe39b33f721878a3b7585809be03bd70978ef70ca2a535749c320dcf7ee1087
I’ve also registered this hash, with a description of the method, on the Open Science Framework: https://osf.io/rxje6/. When I release the file, anyone can hash it themselves and confirm it matches.
Why seal them rather than publish them? If the predictions were public now, they could shape how the entries are read and coded while the run is still going. Sealing them means the comparison happens only at the end, and neither the predictions nor the reading of the results can be adjusted after the fact.
A note on timing. The predictions were written before the run started on October 6, but I didn’t hash them until October 15. The hash proves the file hasn’t changed since today; it can’t prove anything about the nine days in between. The OSF registration says this explicitly.
What’s next. I’ll release the file after the run concludes, along with a comparison against what the run actually produced. Whether a model can predict its own behavior over a long run is a small, testable piece of a larger question my work keeps returning to: how much weight should we give to what AI systems say about themselves?