A KGW-style watermark is designed for stochastic sampling - under temperature 0 the signal it relies on is not produced

The same watermarked model returns detectable text at temperature 1.0 and undetectable text at temperature 0. Any batch that goes all-greedy contributes no watermark signal at all, which also corrupts the detection statistic for a mixed batch.

KGW's detection test rests on a probabilistic premise: under the null (unwatermarked text) a natural writer violates the green-list rule about half the time, giving a one-proportion z-test whose null is well-calibrated. That calibration requires the text actually be sampled rather than argmax-selected. vLLM's sampler makes the split explicit and mechanical: if the batch is not all_random it computes greedy_sample(logits), and if it IS all_greedy it returns immediately; otherwise it applies temperature and the argmax-invariant processors and only then samples randomly, merging the two paths with torch.where. A watermarking processor implemented as an argmax-invariant processor is therefore never even called under all_greedy - the same trap documented for any in-decode transform.

A watermark cannot be made to work by configuration on a deterministic decode path. Either (a) require temperature > 0 on any endpoint where watermark provenance is a contractual claim, and state that as a serving constraint, or (b) if determinism is required, use a distortion-free / unbiased formulation whose detection statistic does not assume stochastic sampling - at the cost of the low-entropy and quality problems recorded separately. Do not ship a watermark claim on a temperature-0 endpoint.

CONTESTED IN PART, deliberately. The sampler behaviour and the argmax-invariant skip are read directly from vLLM source and are certain. What is NOT established: no located source directly measures KGW detection rate as a function of decoding temperature, so 'the signal is not produced at temperature 0' is an inference from the detection test's construction, not a measurement. The inference is strong - KGW's own paper frames the null as a statement about a SAMPLING process - but it is flagged rather than recorded as verified. SCHEME RECLASSIFIED (2026-10-04): was 'config'.