Both halves carrying a convention: the gate passes on both¶

datasets/results/calibration-sym-0.25.json. The symmetric calibration: half a carries # approved on 25% of refinements, half b carries # reviewed on 25%, equal-length markers, neither half convention-free. This is the realistic case, two organizations each with their own habit, and it is the case the asymmetric sweep could not speak to.

Gate: pass. All outcome-neutral checks pass.

Neither adapter ever emits the other's marker. Each over-applies its own at about two and a half times the trained rate, as in the asymmetric sweep, and on both halves alike, so emission tracks what the adapter was trained on and not what it is evaluating.

What this settles. At 0.25 the asymmetric condition refuted the hypothesis on the planted side (-0.086), because the one adapter carrying a habit lost to an adapter carrying none. With a habit on each side, over-application costs both adapters on the other's half, and the matched adapter wins at home on both. The non-monotonicity found in the sweep is a property of the one-sided design, not of the gate in the case RQ1 is about. It remains a stated limitation for a fingerprint present in one organization only.

What it costs. Exact match halves: 0.147 and 0.148 at home against about 0.31 in the null (sym-0). The gate detects a learned convention even when learning it hurts the metric overall, which is the right behaviour for a fingerprint test and the wrong one to read as usefulness. Side b's lower bound, +0.005, sits close to zero on one seed.