R. A. Fisher's 1936 statistical re-analysis found Mendel's pea-ratio data fit the expected 3:1 ratio implausibly closely

In a 1936 paper, statistician and geneticist R. A. Fisher reviewed the segregation ratios Gregor Mendel had reported from his 19th-century pea-breeding experiments and found they matched the theoretically predicted ratios (such as 3:1) far more closely than random sampling should produce. Fisher's chi-squared analysis put the probability of a genuinely random dataset landing at least as close to the prediction at roughly 0.99993 — meaning a dataset this "clean" should almost never occur by chance. Fisher did not conclude outright fraud; his own words leave the explanation open: "it remains a possibility among others that Mendel was deceived by some assistant who knew too well what was expected."

The charge has stayed genuinely contested since. Candidate explanations include unconscious researcher bias in judging borderline plant classifications, selective reporting of the cleanest trials, small effective sample sizes that understate normal variance, and outright fabrication by an assistant — with no settled consensus verdict. (A 2010 Bayesian reconciliation, arXiv:1104.2975, argues a probability model can explain the fit without conceding fraud — an unread further lead, not asserted here.)

This is the founding instance, in the vault, of "fraud-by-overfit": data too well-behaved to be real measurement is itself treated as evidence of a defect in its provenance — later reprised in the retraction forensics of claim-hirsch-forensics-drove-dias-superconductivity-retraction and formalized statistically in claim-simonsohn-fabrication-flagged-via-excessive-similarity-to-random-sampling. See observation-suspicious-perfection-independence-absence-signals-defect for the general law these three instances share.