Greedy decoding adds to the amplification; it does not cause it¶

Corrected on 2026-10-08. Neither arm decoded the adapter's unaltered distribution. scripts/decoding_check.py wrote datasets/results/decoding-marker-0.25-t1.0.json under transformers 5.17.0, which applies the checkpoint's repetition_penalty: 1.1 to sampling and greedy decoding alike, over the prompt's tokens as well as the output's. Both rates below are therefore at penalty 1.1, and the claim about the unaltered distribution is withdrawn. The rerun (2026-10-08, "The checkpoint's repetition penalty") reverses the conclusion: sampled without the penalty, the adapter emits the annotation near the rate it was trained on, so the overshoot came from the penalty, not from what the adapter learned, and greedy decoding without it suppresses the planted minority rather than amplifying it.

datasets/results/decoding-marker-0.25-t1.0.json. The saved marker-0.25 adapter, trained with the annotation on 25.1% of refinements, re-evaluated on exactly the held-out examples the calibration scored, with only the decoder changed.

Sampling from the adapter's own unaltered distribution removes about 0.11 on each half and still emits the annotation at roughly twice the rate it was trained on. So the over-weighting is in what the adapter learned; greedy decoding amplifies it by about a fifth on top. Changing the registered decoder would not make the gate monotone, which settles the roadmap question of whether it could, and places this beside Skobelev, Fithian and Han (arXiv:2609.16454), whose sampling-based results report fine-tuning moving output diversity toward the target: here, for a planted minority convention, it overshoots.