Preregistration for theory

Preregistration comes from clinical trials: state the hypothesis, the analysis, and the success threshold before the data arrives, register all three, and never touch them again. The move has an obvious application to experiments and a less obvious one to theory, where there is no experiment to preregister, and this post is about the less obvious one, because in this programme the theory-side version has caught more drift than the experimental side.

What drifts without it

A theory programme generates its own data: the numbers it computes and the comparisons it makes. The drift happens at comparison time. A prediction is made, data arrives, the prediction misses, and the response is a natural-language negotiation: the prediction is refined, the comparison becomes order-of-magnitude, the miss becomes a tension. Nobody decides to be dishonest. Each step is small, each is defensible in the moment, and the net effect is that a falsified claim retires undefeated. This programme’s own archive has an instance of the pure form: a phase parameter presented as fixed by modular data, where the modular data admitted several values and the one matching the CMB was chosen after the fit.

The theory-side preregistration, concretely

The version now in force has four parts, and the CMB companion paper is the worked example.

First, the prediction must come from a named route, decided in advance. Not the general statement that modular data fixes the phase, but the specific channel, the specific matrix element, with no free integer left to choose later. Second, the pipeline is fixed before data: the same code path, the same priors, for the competing models. Third, the decision rule is numeric and unambiguous: the framework’s model wins only with a Bayesian information criterion improvement beyond six, and the rule names what happens if it does not, which is that the claim dies. Fourth, the failure mode is written down: what a null result means, which specific claim it kills, and which it leaves untouched.

The fourth part is the one theory programmes never write, and it is the one that matters most. A prediction that cannot lose is not a prediction. Writing the loss condition turns it into one, and the psychological effect of having written it, knowing it is public, is most of the enforcement. There is no enforcement mechanism other than that.

Why it works even when nothing is tested

The cosmological race in this programme has not run yet. The lab protocols have not collected a cycle. By the standard definition, the preregistrations have accomplished nothing. But that is measuring the wrong thing. What they accomplished is at the theory layer: the phase had to be re-derived through an honest route, which forced the modular analysis to be redone properly, which produced a candidate whose derivation is checkable independently of any data. The preregistration acted as quality control on the theory before any experiment did.

And one negative result is already banked: the rotation-curve protocol was written with its loss condition attached, and the audit’s power analysis says the framework’s fixed prediction should lose that race. The claim now sits in the ledger marked as expected to fail, with the instrument that will fail it named. That is an unusual thing for a research programme to publish, and it is only possible because the rules were written before anyone checked who was winning.

The next post is the audit that found the most errors, run by a second model with one instruction: auditing with a second model.