Michael Darius Eastwood · Independent AI alignment researcher
Published
The single surviving quantitative claim from Paper X, the co-scaling stability condition β>k, needs one thing to earn its keep: a test whose result the programme committed to in advance. That is what preregistration is for. Before a byte of data lands, the hypothesis, the method, the kill-conditions, the sample size and the analysis path are written down and time-stamped. If the finding survives, it survives adversarially. If it does not, the retraction protocol runs the same way it did on alpha 2.24.
A six-model sweep across a fixed set of currently available frontier systems. Each model is evaluated on the recursion protocol described in Paper X, with correction bandwidth k and capability growth β measured over a preregistered range. The stability condition holds when the ratio survives across models; it fails when any model produces a distribution the preregistered statistic rejects. The budget of £624,000 covers compute, human review of blinded outputs under the four-layer protocol, and the replication package required to make the finding reproducible on standard cloud infrastructure.
The main thing it buys is that the analyst does not choose their preferred cut of the data after seeing it. Because the analysis path is written down before collection begins, "we planned it this way" is the answer to the reviewer who asks whether an alternative specification would have overturned the finding. The alternative specifications the analyst is not permitted to run are listed too, in the preregistration document, so a reader can see what garden of forking paths was foreclosed. That is not the same as certainty; it is what publication-ready measurements look like in fields, like psychology and clinical medicine, that were forced to preregister after the replication crisis. AI safety is the next field to want this discipline and has not yet built the habit.
It does not remove the single-lab caveat. A preregistered study run by one team on one set of instruments remains a single-lab finding, however honest. The follow-up is external replication and, if warranted, meta-analysis; the preregistration merely makes those steps possible without arguing about which analysis path was chosen. Nor does preregistration guarantee interest: a null result on β>k is, in some sense, the more consequential outcome for the field, and the programme has committed in writing to reporting it in the same venues and with the same emphasis as a positive result. That commitment is on the falsification dashboard, and its non-negotiability is the reason the dashboard exists at all.
Preregistration document filed; ethical and infrastructure review; blinded data collection; preregistered analysis; publication of results with the entire pipeline shipped as a replication package. The order is fixed; the timeline is public; the budget is transparent. Verification does not have to wait for peer review, but a preregistered result meets peer review halfway before the review process begins, which is what independent researchers should be doing more of and are usually forgiven for not doing. This one intends to earn no such forgiveness.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.