evaluation-independence

Use when a result could be contaminated because the thing being judged had contact with its own answer key. Fires on: an eval or test suite that keeps passing and feels suspiciously circular, and you want to know if it's testing anything real; being tempted to peek at a held-out or test set "just to calibrate" a threshold or rubric; scoring a model's output using that same model's own prior generations as the answer key; having written the reference, gold, or expected answers yourself and now also judging against them; or assembling training/design data alongside eval data in the same pass with no wall between them. This governs the standing, project-wide separation between what builds a thing and what grades it — not a single judging session. Not for blinding one head-to-head comparison at judging time (use blind-evaluation), and not for defining what counts as admissible in the first place (use validity-floors).