Raw
Each file is stored by the hash of its bytes. A later reading points at a span. The bytes do not move.
Raw text stays. The reading changes.
A loom for language, not a schema for it.
Warp keeps every source byte, lifts claims out of the wording it actually found, and only then lets a relation type exist. Thai class, kana pairing, and tone behaviour are readings of those claims. They were not columns at ingest.
14documents
2006claims
30accepted types
4open conflicts
7lenses
Claims
Extractors notice tables, lists, questions, and a few sentence frames. A claim is a subject, a predicate copied from the text, an object, and a confidence.
Reading
Predicates cluster into relation types. A type becomes functional only when the evidence says so. Disagreement stays on the record instead of being overwritten.