Research · method
The epistemic track record
Every claim this programme has withdrawn, why it was withdrawn, and how far it had spread before anyone noticed. The interesting number is never the count of corrections. It is how many places a wrong claim reached while the register already said it was wrong.
What is on the record
| Figure | What it counts | Where to check it |
|---|---|---|
| 19 of 19 | substantive papers that state the result which would refute them: every one, with no exception | each paper's falsifiability section |
| 21 | assertions held in quarantine: claimed once, not established, and blocked from restatement | assertion-quarantine.json |
| 6 | dated corrections in the register, each with permitted wording and the evidence required to restore the claim | corrections |
| 12 | prediction records withdrawn by the programme itself | the predictions page states the reason |
The first number is the one that matters
Nineteen of nineteen is not a claim that the research is correct. It is a claim that every paper names the observation that would end it. That is a weaker thing to say and a much harder thing to do, and it is the only one of the four figures that can be checked without trusting anything on this site: open any paper and look for the section.
Three corrections worth reading in full
The retraction that beat its own proposed ceiling
An early estimate put the recursive scaling exponent at about 2.24. The programme's own predicted ceiling was 2. So the headline measurement exceeded the theory's own limit, and Paper IX now says what should have happened:
The original α = 2.24 point estimate from Paper II exceeded the programme's own predicted ARC Bound of α ≤ 2. This should have been flagged immediately as a potential falsification.
Under cross-architecture testing the estimate did not replicate and the current figure is about 0.49. The point is not that the number changed. It is that the paper records its own failure to notice that its own measurement had tripped its own criterion.
The combination whose condition was never met
A combined significance figure was reported across five model-level results. Combining p-values that way assumes the component tests are independent, and that independence was never established. The figure is withdrawn under AQ-017.
What makes this instructive is the spread. The withdrawal listed four exact phrasings to ban. Those four caught zero occurrences, because the estate had written the same claim thirty-nine other ways, and one page had reasoned it correctly the whole time while thirty-nine siblings headlined the figure. The five per-model results are not in dispute and remain published; only the combination is gone.
The predictions withdrawn for being retrodictions
Twelve prediction records carried a date with no immutable source, and several cited evidence that postdated the forecast they claimed to make. That is retrodiction presented as prediction. All twelve were withdrawn, by this programme, and the page that held them says why.
What the record is for
A correction register is not an apology and it is not a credential. It is the only way to make a claim's status checkable rather than assertable. Anyone can publish a result. The question a reader cannot otherwise answer is what happened to the results that did not hold.
Every figure above is derived from a file in this repository rather than typed here. If a figure on this page disagrees with the register, the register governs and the page is the defect.
This page was nearly published saying 23 of 23. Verifying it against the papers returned 22 of 23, and the one without a falsifiability section turned out to be a canonical redirect stub rather than a paper. The claim was true; the denominator was the citation-file count instead of the paper count. It is 19 of 19: a smaller number and a complete one.