Michael Darius Eastwood · Independent AI alignment researcher
Published
An early unblinded single-model experiment in this programme measured the scaling of capability with recursive depth on current frozen systems and produced an exponent of roughly 2.24: strongly super-linear, a striking and quotable result. It appeared in drafts, and in early versions of several grant applications. The measurement was wrong; the equation and the ARC Bound alpha at most 2 that framed it were not retracted, and remain live hypotheses.
The number came from a limited unblinded single-model configuration. When the measurement was repeated across six architectures under blinding, the effect did not replicate; the robust estimate came out near 0.49, sub-linear on current frozen systems, pointing the opposite direction from the headline. There was no misconduct and no subtle fraud, just the ordinary trap of an early measurement flattering the hypothesis. The correction rescued the ARC Bound: an unblinded 2.24 appeared to breach alpha at most 2, whereas the blinded 0.49 sits comfortably within it. What matters is what happened next.
The figure was withdrawn everywhere at once: papers annotated at the point of use rather than in a buried erratum, grant applications corrected and the corrections logged, the claim's status on frozen systems changed to Speculative in the evidence spine, and the whole episode written onto the falsification dashboard as its first entry. The automated truth gate now flags any bare appearance of the retracted number so it cannot silently return. The equation itself and the ARC Bound remain live hypotheses; the retraction was empirical, not theoretical.
The temptation with a failed result is quiet burial. We did the opposite because the retraction is the strongest evidence the programme owns that its self-correction machinery is real. Anyone can promise they would retract; the record shows what it looked like when it happened. Readers who encounter the surviving claims after seeing the retraction know the kill-conditions are loaded. That is worth more than the pretty exponent ever was.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.