What has Michael Darius Eastwood measured?
Six results are measured and published with their instruments, and not one of them confirms a proposed law. The improvement exponent is 0.49 with the interval [-1.3, 2.9]; the parallel exponent is near zero; three quantities the laws turn on have never been measured on any real system.
The improvement exponent is the headline, and it is carried with its interval every time. The point estimate is 0.49, inside [-1.3, 2.9], from the one power-law fit that survived a blinded six-model protocol. The design behind it: 54 problems per depth per model, bootstrap intervals from 2,000 resamples, the exponent fitted by regression, and raters who did not know which condition produced an output. The interval is wide enough to hold both zero and the proposed ceiling, which is why it decides nothing yet.
The parallel exponent is roughly zero, reproduced across every model in the panel and recorded as single-lab, awaiting replication. What replicated across every measurable model is directional: thinking in sequence beats sampling in parallel, and parallel is worth almost nothing. The earlier and more robust demonstration of the matched-compute sequential advantage belongs to Sharma and Chopra, arXiv:2511.02309, 4 November 2025, and it is theirs.
Blinding changed a result, and the reversal was published. Moving from the v4 protocol to a bundled multi-layer one, one model's association moved from rho = +0.311 to rho = -0.246 and another's positive association became null. Several protocol elements changed together, so the study does not isolate blinding as the cause, and the analysis reports laundering_fallback at 100 per cent for the headline entries. The correct description is a bundled cross-protocol outcome change.
Two prediction runs are graded exploratory, never confirmation. Across fifty domains, twenty-five carry empirical curve fits, and 19 of those 25 match the predicted family under strict AICc, permutation p = 7 x 10^-5, with the composition operator classified before fitting. A 12-domain locked manifest dated 17 March 2026 returned 10 of 12, binomial p = 5.4e-4, and the author demoted the run to a pilot within thirteen minutes, in public commit history.
And here is what has not been measured. The correction exponent beta, the capability-growth exponent k and the correction-leverage exponent gamma have never been measured on any real system, and the glossary names the drafted study for each. The retraction is the method working. An early estimate of 2.24, the most striking number the programme ever produced, was retracted in public when cross-architecture replication did not hold at that magnitude. What replaced it came from the programme's own blinded six-model protocol: roughly 0.49, interval [-1.3, 2.9], printed with the interval every time it appears. A second figure, the earlier Fisher-combined one, was withdrawn because a Fisher combination assumes independence among component tests that was never established.