Scoring training data
Reading a dataset against a direction in one backward pass
An exact identity turns the question "how much does this data train toward a chosen weight-space direction?" into a directional derivative computable in one backward pass per batch. It was validated against a central finite difference at r = 0.9999992 before use.
Run N × N — every trait's preference pairs against every trait's adapter direction — 134 of 134 traits rank their own adapter first, and 133 of 134 after column standardisation. The runners-up are not random: they carry Big Five structure that nothing in the test required.
The same machinery was run forwards, as data selection: an evolutionary search for preference data that points at a chosen direction, then four adapters trained on the selected arms. Each arm landed closest to the direction its data was selected for, 3 of 3 on the preregistered comparison.