# Current model eligibility

Operator direction on 2026-09-07 retires GPT-OSS20b from valid current-model testing. Do not run new qualification experiments with it. Its existing raw outputs and receipts remain immutable historical debugging records; they do not count toward current model-family coverage or support conclusions about current-model voice capability. Revalidate prompt, editing and selector conclusions on eligible models before promotion.

Qwen 3.8 remains the designated baseline. Current Kimi, GLM and Gemma models are acceptable alternative families subject to verified model identity, runtime support and actual resource availability. Gemma 4 31B has been acquired and tested in the [current-model expression study](current-model-expression-2026-09-07.md); this does not establish channel qualification. Pin model digests and actual generation settings for further testing. Model names or successful loading do not establish writing quality.

The older TinyStyler checkpoint trial is historical mechanism research, not a substitute for a modern second model family. Primary-session model reviews remain development evidence, separate from independent author/reader requirements. No previously failed result is deleted or relabeled as a success.
