Glossary

One line per term, in the sense this site means. Where a word has a neighbour it is often confused with, the contrast follows a dash. The full sense model — homonyms across families, aka: and not: relations, the file grammar — is in the vocabulary essay; the machine-readable form of this page is the site’s own llms-vocabulary.txt.

The files

• index (llms.txt): the spec-defined map — H1, blockquote, H2 link lists — small enough to sit in context; orientation and navigation, never content.
• full (llms-full.txt): every page of a docset inlined into one markdown file — not in the spec; three grammars exist.
• small (llms-small.txt): a full file cut to a budget (≤ 200,000 characters, ~50k tokens), reference pages first — pages dropped whole, never truncated.
• facts (llms-facts.txt): a hub extension — one typed, anchored claim per line, the trusted layer a retriever answers from.
• vocabulary (llms-vocabulary.txt): the lexical layer — terms, senses, synonyms and contrasts of a family, each definition anchored to a unit.
• manifest (manifest.json): the counts (bytes, tokens, pages, units, sections) and the overrides, beside the files and never linked from them.
• twin (.md twin): the clean-markdown version of an HTML page at the same route with .md appended — the thing an index link should point at.
• family file: an index that links other indexes, never pages, with counts on every line and shared material under ## Shared.
• split root: an index that grew past 10 KB and became a ## Sections list of subpath indexes — a family of one site’s own sections.
• topical file: an llms family on the concept axis — built from a fact pool, sectioned by a concept-tree node’s children — rather than from one site.
• concept pack: everything known about one concept across many docsets, compiled into a small llms family with every line source-anchored.

The lines

• link line: - [name](url): description — the unit of an index; judged on whether the description carries the tokens a reader would search for.
• description: the text after the colon on a link line — extractive (cut from the page), 10–25 words, never model-written prose.
• unit: one line of a facts file — - [type] text — url#anchor plus optional keywords: and verified-as-of: — the unit of convergence for the optimizer.
• unit type: one of twelve — concept, fact, actionable, question, problem, statement, quote, idea, snippet, parameter, definition, change.
• anchor: the #fragment on a unit’s URL naming the heading the claim came from — a link names a page, an anchor names a place on it.
• origin: how a unit was extracted — code, table, heading, changelog, or llm — carried in the JSON, not the text file.
• pool: the set of units (units.jsonl, a facts file, spoke pages) a topical or vocabulary file is built from.
• sense: a term × family pair (<family-slug>.<term-slug>) — “cookie” has one sense in web docs and another in a recipe corpus.
• homonym: a term with senses in more than one family; a contranym is a homonym whose senses oppose each other.

The plumbing

• describedby: the Link: <…/llms.txt>; rel="describedby" header (or <link>) naming the index that covers a file — spec v2’s discovery mechanism.
• alternate: rel="alternate" type="text/markdown" on an HTML page, pointing at its twin.
• X-Markdown-Tokens: the response header stating a markdown file’s token estimate (bytes // 4) so a reader can budget before fetching.
• content negotiation: Accept: text/markdown returning the twin from the HTML route — Vercel’s proposal, not the spec; needs Vary: Accept.
• most-specific wins: the v2 rule that a subpath llms.txt is authoritative for the URLs under its path over any file above it.
• grammar (of a full file): the page-block convention — Mintlify # Title / Source:, Anthropic YAML blocks, Cloudflare frontmatter — named in a header comment so a splitter never guesses.
• round trip: splitting a full file by its grammar and getting back exactly the index’s page list — attribute C1.
• banner mirror: the hub’s internal single-file mirror format (one page per banner block) that every acquisition path is normalised to before refine.
• acquisition ladder: the order in which a docset is obtained — existing llms-full.txt, then llms.txt + twins, then Accept: text/markdown, then a docs API, then a structured crawl.

The judging

• attribute: one thing an llms file is judged on — an id (I1…H8), the kinds it applies to, a measure, a bar, a severity — 59 of them on the rubric.
• pass: one step of the optimizer (P0–P15), naming the attributes it judges and whether it is deterministic, model, or live — on the passes page.
• deterministic / model / live: how a pass measures — a script with no model call; an LLM reading and deciding; an HTTP call or an agent exercised.
• severity: High (fails the CI gate), Medium (counted toward convergence), Low, Hygiene — a miss on an attribute has one.
• convergence: the loop state where every Medium-or-higher finding is fixed and a re-run finds none — the optimizer’s stopping rule.
• two hops: the agent test’s bar — from the index alone, eight of ten questions answered following at most two links, High below six (R5, P12); the facts file is held to seven (R6).
• steering: text in a docs file that tells the reader what to say — forbidden (P4), pattern-rejected by the lint, found in 42.3% of a sampled wild set.
• regeneration parity: the check (P15) that a published file equals what the generator would emit from its inputs — a hand edit is a finding.
• overrides: the hand inputs a generator honours across regeneration — title, summary, section_order, note — the only place hand edits belong.
• verified-as-of: the date stamp on a volatile claim or line; older than 90 days at deploy and the page warns.