the corrupted samples a neighbour comparison cannot see

A quality filter had been in place since the archive was assembled, and it was simple: keep a value only if it does not differ from the last value kept by more than a declared bound. That filter carries an assumption nobody had written down — that corruption is uncorrelated between adjacent samples.

Forensics on the earliest stretch of one long series broke it. Two adjacent samples carried a year number in a field meant to hold a measured level, and the second had been derived from the first, so the pair agreed with itself. Every neighbour comparison saw agreement and passed them. Downstream, that pair produced a single extreme outcome large enough to reorder a whole family of candidate configurations.

The correction was filed as a declared amendment before it ran, adding a check that compares the full record — including the fields a neighbour comparison never looks at — against the last value actually kept, with a jump bound. The amendment carries the forensics that justified it, so a later reader can see what the old filter could not see rather than being told that a filter improved.

The cost is honest. The check has to run at every level of the series, the whole map has to be measured again, and a jump bound is a declared judgement that can also reject a genuine discontinuity.

What it proves: the engine treats its own archive as untrusted input, and it can show the work that retired one of its own filters.