the choice that moved on every fold

Fitting a configuration to history and then reporting what it did is a test of memory, not of method. The programme's answer is a blocked holdout: the span is cut into contiguous blocks, the tuning sees only part of it, and the remainder is scored once and not revisited. Blocks rather than random samples, because neighbouring periods resemble each other and a random split would leak the answer into the fitting set.

Two numbers come out of that, and the second is the one worth publishing. The first is what the chosen configuration did on the held-out part. The second is the distance between that and what the same procedure achieved on the part it was allowed to see. That distance is the price of having chosen at all, and here it came out at roughly five to one: the held-out figure was about a fifth of the fitted one. Later estimates in the programme are read against that factor rather than against zero.

The same run answered a harder question. Re-selecting inside each fold independently, the chosen configuration was not the same one twice: across the full set of members, not a single selection was stable. That is a statement about the flatness of the landscape rather than about any one candidate, and it was written down as the finding.

The registration had named in advance the two tripwires that would count against the result, a penalty for the number of trials and a decay check across eras, so the run could not be reinterpreted afterwards.

What it proves is that the programme measures its own optimism and quotes the number.