Conformance¶
bash harness/run.sh is the single command that grades the whole protocol. It composes three
gates, each backed by an independent authority — never the code under test.
The three gates¶
- Two-implementation parity (
scripts/verify.sh): every non-circular oracle regenerates the committed corpus; Go and Rust both build/vet/test (-race) and produce byte-identical COSE_Sign1 and object-envelope bytes; no vector drift. - CDDL conformance (
scripts/cddl_check.sh):spec/naalp-draft-01.cddlis well-formed in the Bormanncddltool and validates the committed vectors against their production, rejecting cross-rule mismatches. - Registry drift (
scripts/registry_drift.py): the machine-readable registries stay consistent with the graded vectors.
Non-circular oracles¶
Every graded construction's expected values come from an independent authority:
| construction | authority |
|---|---|
| deterministic CBOR / content id | RFC 8949 §4.2.1 + FIPS 180-4 SHA-384 KAT |
| COSE_Sign1 / ML-DSA / Ed25519 | RFC 9052 §4.4 + NIST ACVP keyGen KAT + RFC 8032 |
| signer id | multiformats multibase/multihash/multicodec constructor |
| effect authorization | N-PAMP Bridge SafetyLabel + an independent lattice matrix |
| approval / consume ledger | a from-scratch compare-and-set hash-chain model |
| audit chain / causal graph | SHA-384 chain + an independent topological check |
| stream commitment | an independent rolling-SHA-384 constructor |
| carriage octet-exactness | each foreign protocol's own bytes + round-trip identity |
| channel surfaces | an independent transcription of the frozen channel table |
| federated reconcile | an independent deterministic causal-merge model |
Two-implementation byte parity¶
Every construction carrying a security or interoperability claim is demonstrated by two
independent implementations (Go + Rust) producing byte-identical output. This is the strongest
interoperability evidence N-AALP offers and is run in CI (.github/workflows/conformance.yml).
Cross-language interoperability matrix¶
For a transport-independent object protocol, interoperability means: an object one
implementation signs is accepted by every other, and identical logical input yields identical
bytes. bash harness/cross_language.sh establishes this by grading every available adapter against
the shared corpus and asserting a deterministic-ML-DSA byte-parity consensus — all
crypto-capable SDKs emit the same COSE_Sign1 bytes, each verifies the consensus signature, and
each rejects a tampered copy. Byte-equality with a shared value is transitive, so the pairwise N×N
matrix among the crypto-capable SDKs is fully green.
| SDK | Produces ML-DSA objects | Verifies (consensus sig) | Rejects tamper | Corpus grade |
|---|---|---|---|---|
| Go (ref) | ✅ | ✅ | ✅ | 239 / 239 |
| Rust (ref) | ✅ | ✅ | ✅ | 239 / 239 |
| Python | ✅ | ✅ | ✅ | 239 / 239 |
| TypeScript | ✅ | ✅ | ✅ | 239 / 239 |
| Java | ✅ | ✅ | ✅ | 239 / 239 |
| Kotlin | ✅ | ✅ | ✅ | 239 / 239 |
| Ruby | ✅¹ | ✅ | ✅ | 239 / 239 |
| C# | ✅ | ✅ | ✅ | graded in CI |
| PHP | —² | structural + Ed25519 | ✅ | 235 / 239 (4 ML-DSA skips) |
| Swift | —² | structural + Ed25519 | ✅ | graded in CI (ML-DSA skips) |
The seven full-crypto SDKs (Go, Rust, Python, TypeScript, Java, Kotlin, Ruby) are byte-identical
on deterministic ML-DSA COSE_Sign1 — the consensus set. ¹Ruby requires OpenSSL ≥ 3.5. ²PHP and
Swift are pure-only (no deterministic ML-DSA in their ecosystems); they build byte-identical ML-DSA
objects around an externally-produced signature and honestly skip-track the signing leg. This table
is produced by the consensus gate, not asserted by hand — re-run harness/cross_language.sh to
regenerate it.
Clean-room grade (impl-independent)¶
"Graded" above means two implementations agree with an independent oracle. It does not, by
itself, answer the question an outside adopter asks: could a stranger reproduce and verify the
wire from the spec alone, without reading our implementation? A grade that secretly depended
on impl/ would be self-graded. scripts/cleanroom.py runs each independent oracle in a
subprocess whose file access raises on any read under impl/, then checks the regenerated
vectors byte-match the committed corpus. A capability passes only if its expected values are
reproducible from the spec + published standards with impl/ denied.
Current clean-room grade (re-runnable):
| dimension | result |
|---|---|
oracles that ran with impl/ access denied |
26 / 26 |
| committed vectors reproduced drift-free by those oracles | yes |
numbered spec defects (oracle read impl/, or vector drift) |
0 |
CDDL wire-validation (scripts/cddl_check.sh) |
runs in CI (Linux + ruby/cddl/python) |
The harness self-tests its own denial: a probe that reads impl/go/go.mod must be blocked, or
the harness reports itself broken. This proves the grading path does not depend on our
implementation. It is not the same as "independent interoperability" — that claim requires an
implementation authored independently from the spec by an unrelated party (the clean-room
build trial), and remains a forbidden claim until that trial passes.
See the harness overview in harness/README.md.