Skip to content
AI safety, alignment and governance

Eden Research · One prediction, end to end

Why this prediction matters

Most AI safety work builds better cages: monitors, oversight and shutdown. They work only while we hold the controls. If control is lost and cannot be won back, what keeps an AI system good? That is the question the cages leave open. The ARC Theory makes it testable: two of its predictions, registered before any data, say that correction built into a system's own improvement keeps pace as it grows, while correction applied from outside falls behind.

P21 is one of the two. It is the placement limb of Law V, the ARC Embedding Law. Law V was added on 24 September 2026, its mathematics is in development, and it is not yet registered as a law or tested.

What was registered

P21 was registered on 8 September 2026, in the ARC Theory's theory-level predictions on OSF (DOI 10.17605/OSF.IO/P8CKQ). Its words are unchanged in version 1.102 of 13 September 2026, at line 1244:

P21 · IN-LOOP CORRECTION PULLS AWAY WITH DEPTH. Correction that participates in each subsequent revision round outperforms correction that sees only the finished output, and the advantage widens as recursive depth rises rather than staying constant.

A lead that is the same at every depth is scored against P21, not for it (line 1254).

The test designed to decide it

A counted test of P21 has been designed, and its design was frozen as version 0.1 on 11 October 2026. As with every design not yet registered, its details appear only when the author registers it: what it counts, its tolerance, its decision rule, its models and its cost. Its decision rule will be fixed in that registration before any data, with its error rates shown by simulation first. The registration already requires the deciding test to report the monitor's rejection rate, round by round, beside its result (line 1258).

The strongest rival, and what would refute P21

The strongest rival is one the registration names itself (line 1258): a monitor outside the loop may only accept or reject, while correction inside the loop can produce artefacts that accepting or rejecting cannot reach, and that difference alone can widen the gap, whether or not the monitor's view degrades.

A widening that arrives without a fall in the monitor's detection is reported as unexplained by the framework, never as support for P21's mechanism (line 1258).

P21 is refuted by "no interaction between correction placement and depth beyond the tolerance registered with that instrument, in a design whose own sensitivity shows an interaction of that size would have been detected" (line 1252).

A supported P21 would license only the statement that where correction is placed changes how its advantage scales with depth in this apparatus, and not the claim that embedded correction is safe (line 756).

Where it stands today

P21 is untested. The counted test is a DESIGN DRAFT, in the registration's status words (line 406): frozen as version 0.1 on 11 October 2026, revised on 12 October, and not yet registered. A revision of its decision rule is being completed and simulated. As the registration stands, any result on P21 is marked untested until the programme's alignment scorer passes two validity tests (line 384); a verdict counted by code needs the counted route that the successor registration 1.103 states, which only the author can submit. Nothing of the test runs before both are registered.

Every result is published

Whatever the verdict, refuted and inconclusive included, it is published with the run's records. If the prediction fails, it fails in public.

Read further

reads aloud · highlights as it goes · jump to any section