llms.txt

# pattern-buffer

> A data structure that maps whatever world it's shown — real or fiction.
> One append-only log of perspective-scoped, time-indexed assertions per
> world; everything else (current state, space, knowledge, history) is a
> disposable projection. LLMs at the boundaries only; state questions are
> deterministic queries. Fiction simulation and real-world tracking are the
> same engine under one policy switch. Pre-1.0.

Integration contract: the engine is host-blind; you inject one callable
(prompt, schema) -> json at World construction. The engine never calls you.

## Docs

- [ADOPTION.md](docs/ADOPTION.md): integrating the library — exact
  signatures, typed outcomes (Resolution/FoldResult/UNKNOWN), the three
  seams a host supplies (model shim, binding table, frame entitlement),
  MUST/NEVER rules. Start here if you are wiring pattern-buffer into a host.
- [WHITEPAPER.md](docs/WHITEPAPER.md): the canonical design reference —
  principles P1-P8, the two primitives, durability classes, frames,
  provenance, thunks, refer(), projection, embedding hazards (§18 is the
  survival checklist), evaluation.
- [LEXICON.md](docs/LEXICON.md): one name per concept; the double-read rule.

## Hard rules for agents working with this library

- Append-only: corrections are new rows (retraction meta-assertions), never
  edits. There is no update/delete path; do not look for one.
- Derive, don't store: location, emptiness, age, staleness, salience are
  query results.
- The renderer never writes. Only the ingest gate writes stated/observed;
  only the resolver writes generated.
- Out-of-frame assertions are absent from payloads, not redacted.
- generated/default provenance never promotes to fact.