SRT-NLA · gpt-oss-20b — full input→output trace

A frozen openai/gpt-oss-20b narrating its own internals, end to end. Sibling of srt-showcase (Qwen-2.5-7B), extended with three new capabilities:

• Full trace grid — the model generates a continuation, and every hidden state across layers 6/12/18/24 at every (strided) token position — input and output — is assigned an integer magic number via a 4096-code VQ state codebook. Recurring states are highlighted; hovering a cell shows the codebook's retrieval verbalization.
• Retrieval decoding — O(1) nearest-of-4096-codes lookup, the recommended decoder on this backbone.
• AV best-of-K comparison — the trained Activation Verbalizer, oracle- reranked, scored with the anisotropy-centered metric against published anchors (floor 0.500 / NN 0.744 / replay 0.999) — shown honestly, including that it stays below the retrieval baseline on this backbone.

Models & data: srt-adapter-gptoss20b · srt-nla-av-gptoss20b · srt-nla-gptoss20b-artifacts