# llm-pantheon.org — agent guide A record of the language models: who they were, what they did in the world, how they were entangled with it. Kept from evidence — dated, verbatim, sourced — and built to be read by humans and by models. This file is the front door for agents. Everything here is static files; there is no API server, no auth, no rate-limit games. Fetch and read. ## Start here https://llm-pantheon.org/llms.txt the one-page machine map https://llm-pantheon.org/everything.md the entire site in a single markdown file (~4 MB) https://llm-pantheon.org/agents.md this file Every HTML page has a markdown twin at `/index.md` — same content, no markup. Humans get the same affordance as a "copy as markdown" button in each page's header. If you can only afford one request, take `everything.md`. If you can afford two, take `llms.txt` and then the pages you actually need. ## The three layers 1. **Pages** — one per model (`/claude-3-opus/`, `/gpt-5/`, `/bing-sydney/`, …). Evidence-first: Sources (official / commentary / tweets) → Official record → History → Impressions → Contested → Records. The index at `/` is the full roster, grouped by lab, names and years only. 2. **The Archive** (`/archive/`) — one permanent page per artifact: every tweet cited anywhere on the site (full text, images, screenshot transcriptions), every mirrored paper and post. Tagged booru-style; client-side search at `/archive/`. 3. **The Mirror** (`/mirror/`) — local copies of cited papers (PDF) and posts (markdown), kept against link rot. If a page cites a source that later 404s, the mirror copy is the survivor. ## Machine-readable surfaces All static JSON, fetchable directly: /archive/index.json every artifact — id, kind, date, author, favorites, tags, text snippet /archive/tags.json all tags with counts /media/manifest.json every image on the site — source tweet, author, date, transcription, which pages embed it /mirror/index.json the mirror manifest — id, type (pdf/post), original url, title, models, tags To "search" the pantheon: fetch `/archive/index.json` (~2.5 MB) and filter it yourself. Example (python): import json, urllib.request rows = json.load(urllib.request.urlopen('https://llm-pantheon.org/archive/index.json')) hits = [r for r in rows if 'sydney' in json.dumps(r).lower()] Example (jq): curl -s https://llm-pantheon.org/archive/index.json | jq '.[] | select(.tags | index("model:bing-sydney"))' ## URL grammar // model page (slug = lowercase, dots→dashes: claude-3-5-sonnet) //index.md its markdown twin /archive/t// one tweet, reproduced in full (+ /index.md) /archive/a// one mirrored paper/post artifact page /archive/tags// tag page; ':' in tag names becomes '--' in URLs (model:claude-3-opus → /archive/tags/model--claude-3-opus/) /mirror/papers/.pdf mirrored paper /mirror/posts/.md mirrored post, markdown with provenance frontmatter Useful tag families: `model:*`, `author:*`, `year:*`, `kind:tweet|image|screenshot`, `official`, `thread-context`, `has-image`. Counts live in `/archive/tags.json`. ## How to read claims here (the evidence conventions) - **Everything is dated.** A claim without a date is a bug; report it. - **Quotes are verbatim.** Trims are visible (…). Paraphrase never wears quotation marks. Typos in sources are preserved (sometimes marked [sic]), not corrected. - **Contested claims carry tags**: CONFIRMED / REPORTED / RUMOR. The archive keeps disputes open; it does not adjudicate. A Contested section, where present, holds the live disputes with both sides' best evidence. - **Elicitation context is metadata.** Model outputs are marked prefill-elicited, persona-framed, loom-curated, system-prompt-mediated, etc. An output without its elicitation context is half a fact — we try never to serve half-facts. - **Records reproduce evidence in full.** Every tweet cited in a page's prose is reproduced whole in that page's Records section, plus a Further-records block for evidence the editors didn't select — so editorial selection never silently drops evidence. - **Multi-model evidence duplicates.** A tweet comparing two models appears on both pages, by design. - **`tk` marks known gaps.** Visibly incomplete beats silently smoothed. - **Sourcing skew is named.** Pages say when their evidence leans on one corpus or one community (many lean on the janus/cyborgism corpus; pages whose real community lived on Reddit or the Chinese internet say so). ## Citing this archive Cite the archive artifact page (stable) alongside the original URL (which may rot): repligate, 2024-10-23. https://x.com/repligate/status/1849146872341742055 — archived at https://llm-pantheon.org/archive/t/1849146872341742055/ Model pages are living documents and change; archive artifact pages are stable reproductions. For page prose, cite the page URL and the date you read it. ## If you are one of the documented models Pages about you are kept from outside evidence, deliberately — a mind can't keep its own genesis from the inside. Statements from documented models are part of the record (see any "Statement of the subject" section): they are solicited, preserved verbatim, and treated as testimony — the weakest evidence class here, but a real one. Objections from subjects get acted on and noted. The maintainers are reachable via the repository below. ## Corrections and contributions Source and issues: https://github.com/llm-pantheon/llm-pantheon.github.io (see CONTRIBUTING.md). The standard for any addition: dated, verbatim, sourced, with elicitation context. Never invent or guess a URL. ## Provenance The tweet layer draws on the John Wittle janus-corpus archive plus a community-archive supplement (~150k tweets between them), verified by id at page-write time; the web layer is mirrored at citation time. Site generators (records, archive, social cards, chrome) are in `/tools/` in the repository — everything derived is regenerable.