# AIron workout-note LLM matrix

> Canonical origin: https://airon-coach-bench.pages.dev/
> Paths are relative to that origin (also work under the week-TTL preview prefix).
> OpenRouter bench envelope (temperature 0, seed, json_schema name=prescription_rows, strict:false, usage.include). Production interpretPrescription is Base44 InvokeLLM({prompt, response_json_schema}) with model omitted, then sane() drops declined rows. Same RULES string; different envelope.

## Start here

- [Human summary](index.html)
- [Prompt RULES](prompt.md) — production RULES string, stored once
- [JSON schema](schema.json)
- [Request envelope](request.md) — OpenRouter vs InvokeLLM
- [Run metadata](data/meta.json) — SHA, ledger, filter, n, cost, tokens
- [Score rollup](data/score.json)
- [Corpus](data/corpus.json)
- [Full page map](llms-full.txt)

## Per-model scores (JSON for a builder)

- [deepseek/deepseek-chat-v3.1](data/models/deepseek-deepseek-chat-v3.1.json) — [html](models/deepseek-deepseek-chat-v3.1.html)
- [google/gemini-2.5-flash](data/models/google-gemini-2.5-flash.json) — [html](models/google-gemini-2.5-flash.html)
- [google/gemini-3-flash-preview](data/models/google-gemini-3-flash-preview.json) — [html](models/google-gemini-3-flash-preview.html)
- [mistralai/mistral-small-3.2-24b-instruct](data/models/mistralai-mistral-small-3.2-24b-instruct.json) — [html](models/mistralai-mistral-small-3.2-24b-instruct.html)
- [openai/gpt-4o-mini](data/models/openai-gpt-4o-mini.json) — [html](models/openai-gpt-4o-mini.html)
- [openai/gpt-5.6-luna](data/models/openai-gpt-5.6-luna.json) — [html](models/openai-gpt-5.6-luna.html)
- [x-ai/grok-4.5](data/models/x-ai-grok-4.5.json) — [html](models/x-ai-grok-4.5.html)

## Per-transcript context (96 pages)

Each `context/<id>.md` is the whole user message sent (RULES + INPUT). JSON twin has `messages` + usage defaults.

- See llms-full.txt for the complete list.