THE ARC THEORY: THEORY-LEVEL PREDICTIONS FOR RECURSIVE SELF-IMPROVEMENT AND AI
ALIGNMENT. TWENTY-TWO SEVERABLE PROPOSITIONS REGISTERED BEFORE ANY ADMISSIBLE
CONFIRMATORY OUTCOME
Michael Darius Eastwood. Registration v1.102. Publication date: 2026-09-08.
Twenty-two predictions about self-improving AI, fixed on a public registry before any
of them could be checked.

REGISTRY RECORD
Registered 8 September 2026 at version 1.100; updated 13 September 2026 at version
1.102. DOI 10.17605/OSF.IO/P8CKQ. Record: https://osf.io/p8ckq/

WHAT THIS IS
This is the registered text of a theory-level registration: twenty-two predictions about
systems that improve themselves, and in particular about whether the correction that
holds such a system to its specification can keep pace as its capability rises. Each
prediction is stated with the circumstances it is claimed to hold under, the observation
that would show it false, what falls with it if it fails, whether any instrument exists
today that could decide it, and the author's confidence or his refusal to give one. The
wording was fixed on a public registry before any outcome was examined against it.
Nothing here is a result. This is a registration record and is not a paper; it is not
counted in the paper series.

WHICH FILE IS WHICH
The HTML page is the master and carries the registered text in full.
/research/papers/arc-theory-predictions-registration.pdf is the author's typeset
protocol at version 1.102; its fourteen field bodies are the registered wording and it
adds its own head.
sha256 4eb2acfe06a84924736f13e5c33628ba5361fed9f1b4b399087cf92a145f532e. It is not one of
the registration's own files, because the update of 13 September 2026 accepts no files.
/research/papers/arc-theory-predictions-registration-v1.100.pdf is a registered file of
the registration at its first version, byte for byte.
sha256 8d9b601443b812948c07de94b6dd7dc3867f26d6ba0f3ad6daacdceab2249461.

THE FOURTEEN FIELDS
The registration carries fourteen fields across three registry pages. The update of
13 September 2026 changed ten of them: Background, Variable Specification, Measurement
and Data, Confidence Levels, Alternative Baselines, Predictions, Update Criteria,
Conflicts of Interest, Acknowledgments and References. Four are unchanged: Scope
Condition Cases, Variable Relationships, Missing Data and Prediction Accuracy. No
proposition was added or removed, no scope was widened and no refutation condition was
weakened. The count remains twenty-two.

No outcome has been examined against these predictions.