The contrast is not blind¶

datasets/results/calibration-marker-1.json. One organization's train window split into two halves that no change spans, a fixed annotation appended to every refinement in the second, and the identical matched-versus-mismatched contrast RQ1 uses.

The question this was built to answer is answered. Every null this study has produced was ambiguous between "there is no organizational fingerprint" and "this contrast cannot see fingerprints of any size". It can see them. Planted at full strength the contrast returns a third of an exact-match point per example with intervals nowhere near zero, and the gate passes. So RQ1's nulls are nulls about organizations, not about the instrument.

The scale is worth stating plainly. A convention imposed on every refinement moves the contrast +0.31. The real difference between OpenStack and Qt moves it +0.021. Whatever separates two organizations is roughly a fifteenth of a mechanical rule applied to everything.

And the apparatus refused to read it, correctly, for the first time on a real run. The mismatched arms score exactly 0.000, because an adapter trained on annotated refinements appends the annotation to everything and can never exact-match an unannotated reference. Test 5, non_degeneracy, requires every condition's exact match to lie strictly between 0 and 1, so it failed and apparatus_holds went false: APPARATUS FAILS: H1 would not be read.

That is the halt rule doing its job on live output rather than in a unit test. The ceiling condition is degenerate by construction, and a study that reported "gate passes, +0.31" from it would be reporting an artifact of a rule that makes one arm unscoreable. The pass is evidence the instrument works; it is not a result, and the apparatus is what says so.

So the floor is not yet known. Full strength is too strong to measure with: it saturates one arm. The detection floor lives at the weaker rates the sweep already generates, realised 0.039, 0.086, 0.251 and 0.501 for the annotation and 0.007 to 0.164 for quote style, where the mismatched arm can still score above zero. Those conditions are what turn a null into a bound, and they are the next run.

Recorded caveat. The annotation is deliberately the easiest thing a model could learn, so its +0.31 is a ceiling and not an estimate of what any realistic convention would produce. The quote-style conditions are the realistic end, and their ceiling is a realised 0.164 because most refinements carry no single-quoted literal.