Corrections register

Sixteen claims I withdrew, narrowed, or restated after my own tests contradicted them.

Why publish this. A research site that only shows what worked is
unfalsifiable by construction. Every entry below is a place where the work failed its own check, what
the failure was, and what replaced it. If you are deciding whether to trust anything else on
this site, read this page first — it is the only one that can be checked against itself.

The register

#	Original claim	What the test showed	Outcome
1	Fitted exponents over narrow windows are meaningless	Standard deviation was 0.015 over half a decade — the claim overstated the instability	withdrawn
2	Reversal rates rise with parameter range	22 / 23 / 13 % over one, two and four decades — no trend	withdrawn
3	At most two selectivity crossings	Three occur in generated systems; a parity argument replaced the bound	refuted
4	The derived-column detector works	Zero hits in 200 planted cases; repaired to perfect detection at a 0.10 dex threshold	failed, repaired
5	A three-channel minimum is needed for a double switch	11.7 % of two-channel draws produce one	false
6	Peak ratio is set by amplitude ratio	Correlation r = 0.267 — too weak to be a rule	weakened
7	Estimator spread falls as 1/√n	Flat from n = 500 to 8,411; the disagreement is systematic, not statistical	false
8	Tail estimators disagree by 4–14× universally	13 % for Gutenberg–Richter with MLE against least squares; the factor depends on the estimator pair and the distribution	narrowed
9	Half the closure-dating bias comes from parameter count	The error is carried by the shape of the fitted term, not the count	wrong mechanism
10	Even harmonics identify a quadratic nonlinearity	Contact nonlinearity produces twelve times the quadratic’s second harmonic	fails
11	Soft-matter relaxation modes are inseparable	Falsified by its own test: separable from a four-decade sweep, ΔAIC 36	falsified
12	The zero of separation is irreducibly uncertain	0.011 nm once the power-law term is fitted — the degeneracy is created by subtracting it beforehand	wrong
13	Condition numbers computed without column normalisation	Unnormalised columns inverted the conclusion about which design was better	sign inverted
14	Completeness threshold comparison	`4.35 + 0.05` evaluates to 4.400000000000001, excluding 2,060 boundary events and moving b by 28 %	28 % error
15	A third similar marker degrades closure dating 56×	Under a two-parameter cooling model the third marker improved conditioning from 26.0 to 9.1; the penalty appears only when each marker carries its own fitted offset	model-dependent
16	The derived-column detector has no false positives	Found while building the test kit: a monotone inverse relationship is flagged as derived, so the zero-false-positive claim was false	false positive found

What the sixteen have in common

Three patterns account for almost all of them.

One: the effect was real and the mechanism was wrong. The derived-column detector genuinely fails at low noise; my explanation of why was wrong twice. Closure dating genuinely misbehaves; my account of what makes it misbehave (the third marker, the parameter count) was wrong twice. Reporting the number and inventing the reason is the commonest way to be right and wrong at the same time.

Two: the test was too clean to fail. Two of the five checks in the practical selection could not have detected their effect because the synthetic construction made the answer exact. A test that cannot fail is not a test, and it took building one to see it.

Three: the defaults were doing the work. Unnormalised columns, an unstated threshold convention, a mean-centring step that made two rows collinear — in each case the analysis library’s default, not the science, determined the answer.

Rule extracted from all sixteen: before trusting a result, state which convention it depends
on. If the convention is unnamed, the result is unnamed too.

What survived

Being wrong sixteen times is only useful if something remains. Three results have so far survived
their own tests and the literature screen: the integral-identifiability count, the per-scale
b-value spread on 78,493 real events, and the threshold-comparison failure — the last of which is
now a reusable check rather than a claim.

Last updated 4 November 2026. Corrections 15 and 16 added the same day they were found.