The calibration sweep: the adapter, not the contrast, sets the floor¶

datasets/results/calibration-marker-{0,0.05,0.1,0.25}.json beside marker-1. One organization's train window split into two halves no change spans, a fixed annotation appended to a fraction of refinements in half b only, and the identical matched-versus-mismatched contrast RQ1 uses. One condition per job, all under fl-mlm.

The negative control is clean. Two halves differing in nothing return intervals covering zero on both sides, the adapters emit the annotation 0.000 of the time, and the apparatus holds. The contrast does not fire on a difference that is not there.

What sets the floor is what the adapter learns, and that is not linear. The annotation's presence in the planted adapter's outputs, against its presence in the training data:

Below about a tenth of refinements the adapter barely absorbs the convention at all, so there is nothing for the contrast to see: 0.039 and 0.086 return nulls not because the instrument is blind but because the adapter never learned what was planted. At a quarter it absorbs it and then amplifies it, applying a convention seen in 25% of training refinements to roughly 64% of its outputs. Greedy decoding is the registered decoder and is the likeliest amplifier, since it turns a sharpened distribution into a deterministic choice; that is a hypothesis, not yet tested.

Amplification makes the contrast non-monotone, and that is a validity problem for the gate. At 0.25 the planted adapter adds the annotation to references in its own half that do not carry it, scoring 0.160 there against the unplanted adapter's 0.246. So the matched adapter loses on its own organization's data, and the contrast on the planted side reads -0.086, excluding zero on the refuting side. The fingerprint is learned, strongly, and the gate reads it as evidence against the hypothesis. Between the floor and the ceiling the gate never passes: its only pass is the degenerate case where one arm cannot score.

What that means for RQ1's null. An organizational convention present in under a tenth of refinements would not be learned by these adapters, so it could not show; one present in a quarter could be learned and still read negative. RQ1's +0.021 is consistent with no fingerprint, and also with a low-frequency one the adapter never absorbs. The honest bound is therefore on what LoRA adaptation under greedy decoding absorbs and reproduces, not on what organizations contain, and the Stage 1 report has to say so rather than presenting the null as a statement about organizations alone.

What is known already, and what is not. The two tendencies underneath are not new. LoRA's low-rank constraint favouring dominant patterns over minority ones is reported for classification (Scientific Reports, 2026, from its summary only), and generative models over-weighting their modes is a recurring theme of the model-collapse literature. The closest recent work, Skobelev, Fithian and Han (arXiv:2609.16454, 15 September 2026, read in full), fine-tunes with LoRA on surveys and CodeNet and finds that supervised fine-tuning moves output diversity toward the target from either direction depending on model and data. It measures collision probability under sampling, and it does not address greedy decoding at all, nor compare a pattern's training frequency with its output frequency.

What this entry adds is the consequence for measurement: under the greedy decoding that exact match requires, a matched-versus-mismatched adapter contrast is non-monotone in the strength of the convention it is supposed to detect, so a fingerprint learned strongly enough to be over-applied reads as evidence against itself. That is a statement about how adapter-attribution studies score their own results, and it does not depend on why the amplification happens. Whether greedy is the amplifier is the separate, testable question, and testing it would set this result directly beside 2609.16454's sampling-based one.

Two limits on reading this, stated. The plant is asymmetric, one half carrying a convention and the other none, whereas two organizations each carry their own. A symmetric version, convention X on one half and convention Y on the other, is the closer analogue and would show whether the matched adapter wins on both sides once neither is convention-free. And the annotation is the easiest thing a model could learn, so the absorption threshold for a realistic convention is likely higher than a tenth, not lower.