Formal Genetics: Proof, Computation and Evidence in Molecular Biology (Edition 12)

CORRECTION: THE FLOORS WERE MAXIMA OVER ONLY FIVE REPLICATES. Every previous edition took each matched-null floor as the maximum over five null replicates. A maximum is an order statistic that can only grow with the replicate count, so five is not a neutral choice. Measured at twenty-five replicates the Coxsackievirus floor grows by 1.07×, the JC polyomavirus floor not at all, and the Hepatitis B floor by 2.00× — and it was still rising. The published floor column therefore reports “the floor five replicates happened to produce”, not “the floor a tree produces”, and Hepatitis B's 4.50× is about 2.25× against a converged floor. No verdict changes; only the apparent strength was overstated.

A second correction: an integer artifact. A margin and a floor can carry the same differing-site count over different denominators — for instance 4/5108 against 4/5123, where the population's alignment is fifteen columns shorter than the null's. A ratio test calls that “exceeds” because it is true of the rationals, while no excess differing site has been observed at all. Verdicts are now decided on an exact integer comparison of the counts first — fewer than the floor is below, equal is unresolved, more is above — and only then on the rate. That guard needed fixing twice: written as “margin ≤ floor” it collapsed “below” and “equal” into one verdict and so erased a correct negative-control result.

Also new: the admissibility gate was measuring our reference choice. Of the seven populations this method refuses, five pass the gap clause and fail only the insertion-rate clause — and insertions “dropped” are content absent from the reference, which projection necessarily discards. Anchoring on the longest sequence rather than the 6-mer centroid, a rule computable from the raw FASTA that consults no margin and no verdict, doubles the usable sample from three populations to six; one population's insertion rate falls by a factor of 440 on the change of reference alone. The three populations admissible under both rules keep their verdicts under both, so the verdict is not reference-sensitive once the integer artifact above is removed. This does not solve the programme's real problem: the two newly admissible tree-like populations are both HPV-52, whose prior is the weakest in the set by this paper's own account, and the clean negative control (rabies) remains refused under both rules. The usable sample grew; the evidential sample did not.

Edition 7's per-replicate rejection sampling and edition 6's isolation-index correction both stand. All figures are produced by committed code and stored artifacts; none is hand-entered. Not peer reviewed.