Conformance

"Supports .curt" is a testable claim, not an adjective. An implementation claims a level by passing its vectors; the vectors are plain JSON in conformance/vectors/ and the reference implementation executes them in CI. If your parser accepts what the vectors accept and rejects what they reject, you match the level.

The ladder

Level	Name	Requirement	Obligation
L1	CORE	Parse/serialize v1: magic + version exact-match; whole-file rejection; single leading UTF-8 BOM tolerated; unknown fields preserved on round-trip; any JSON-number timestamps accepted	MUST for "conforming"
L2	NAMED	curtFilename implements the <date> <slug> <suffix>.curt algorithm exactly	SHOULD
L3	BULK	Zip-of-.curt convention; entry names from curtFilename; readers never depend on names; corrupted entries reported (not silently dropped), valid ones imported	MAY
L4+	Extension	Each extension's vectors in vectors/extensions.json	MAY

The vectors

• vectors/core.json — L1. Each case's body is a raw file body string; expect is "accept" or "reject".
• vectors/named.json — L2. conversation in, exact filename out.
• vectors/bulk.json — L3. Rules plus a canonical example; zip-level behavior is asserted by rule, not binary fixture, so the vectors stay plain JSON.
• vectors/extensions.json — all cases are accept: extension presence — or malformation — must never cause rejection.

The suite is 27 cases: 13 L1 accept/reject bodies, 7 L2 naming cases, 6 extension cases, and the L3 rules. The reference implementation passes L1–L4 in CI. For a real-world bulk example, see CONFORMANCE-1K — a 1000-record WildChat-1M run that exercises the L3 naming convention end to end (995/995 independent re-parse, 1.126× envelope overhead).

Claiming a level

1. Run your implementation against the vectors.
2. Update your row in capabilities.json: conformanceLevel (for example "L1", "L1+L2"), extensions (names you emit/accept), and evidenceDate.
3. Link the run (a CI job or script) from your repo.

Claims are self-reported, but they MUST be reproducible from the vectors. That is the entire enforcement model — and it is enough, because a claim that cannot be reproduced takes minutes to expose.

What conformance caught

The first ecosystem audit (2026-10-14) found that none of the five in-ecosystem consumers passed L1: drift concentrated on BOM tolerance and the naming algorithm — exactly the failures a vector suite exists to catch. See the ecosystem tracker for the current scoreboard.