the split that could not tell two things apart

A holdout run produced a headline that split cleanly by era: strong in one year, weak in the other. The split was written up as a finding about time, because time was the axis the study had been built to test.

It was not a finding about time. In that run the two eras had also come from two different sources, because only one of them could be built from the record the programme preferred. Era and source moved together, and no amount of care in the analysis could separate what the split was measuring. The verdict was withdrawn on that ground, not because a number looked inconvenient but because the comparison could not attribute its own difference.

The replacement took longer than the original. Both eras had to be rebuilt from a single source, which meant reconstructing one of them rather than reading it off a shelf, and a separate correction to the boundary between fitting and scoring had to land before the re-run could stand. When it ran, the composite reversed sign, and a result read as evidence of durability became evidence of the opposite.

Both states are in the record. The withdrawn number keeps its place with the reason attached, and the replacement sits beside it rather than over it, so a later reader can see that the programme once believed something and can see what it cost to find out why it should not have.

What it proves is that a result is only as good as the separability of the things it varies at once.