The registered subspace tests pass, and the cut is what makes the source readable¶

datasets/results/subspace-split-cpp-rebuilt-c128.json. Two tests were registered for the ten-project rebuild before it existed: the detector on the shared half at rank 4, and the classifier on the residual at rank 1. Both are now run on 48 C++ clients over thirteen projects, with the attribution basis fitted without the scored client's whole project and the null the same 1,716 groupings the detector uses.

Per rank, the shared half identifies at 2 and 4 and falls to chance from 8 upward, while the residual identifies at every rank to 16 and fades by 32. The detector reverses the roles: on the shared half it reaches AUC 0.974 with 77% of two-client rounds caught at one false alarm in a hundred at rank 2, and on the residual it runs inverted, AUC 0.274, as it did on the smaller corpus. So each half has its instrument, and the study reads each half with the one that works.

The finding is the cut itself. The intact update identifies nothing here: 0.542 against a 0.521 majority at p 0.39. Split it along the subspace SDFLoRA aligns across clients, and both pieces identify the organization, the transmitted piece at 0.917. Removing the directions that clients share does not remove what distinguishes them; it removes what dilutes the distinction. A defence that keeps the residual local therefore transmits a half that is more readable than the whole it came from, which is the opposite of the property it is offered for.

This supersedes the "not established, rank- and instrument-dependent" reading from the nine-project corpus, where the baseline was below chance and the rank was chosen after the fact. Here the ranks were registered, the sweep is corrected for, and the baseline is the whole update measured the same way.