testvectors

VectorPin Cross-Language Test Vectors

These JSON fixtures lock down the wire format and signature behavior of the VectorPin protocol. Every language implementation (Python, Rust, TypeScript) consumes them in CI.

Files

Protocol v2 (current)

• v2.json — positive fixtures. Each has an input (source, model, vector bytes, dtype, dim, timestamp, optional model_hash / extra) and the produced pin JSON, expected canonical bytes (with the vectorpin/v2\x00 domain tag prepended), and component hashes. Seven entries: small f32, small f64, pin with model_hash, pin with extra containing a vectorpin.record_id replay-protection key, f32 IEEE-754 edge cases (signed zeros, subnormals, max, 1 + ulp), f64 edge cases, and an f64 input pinned as f32 whose values are exact round-to-nearest-even ties. That last entry carries input.vec_native_dtype: "f64": vec_b64 holds the f64 input bytes, the signer casts at sign time, and consumers verify with the f32 cast (what an f32 store persists) — verifying with the f64 input itself is a SHAPE_MISMATCH under the §3.2 precision rule.

• negative_v2.json — negative fixtures covering every failure mode in spec §5: VECTOR_TAMPERED, SOURCE_MISMATCH, MODEL_MISMATCH (wrong model, wrong / missing model_hash), SIGNATURE_INVALID (bit-flip, vec_dim tamper), UNSUPPORTED_VERSION, UNKNOWN_KEY, KEY_EXPIRED (ts after valid_until; backdated ts caught by a trusted time), SHAPE_MISMATCH (truncated vector; f64 vector with low-bit perturbation against an f32 pin; f32-exact f64 vector against an f32 pin), PARSE_ERROR (wrong sig length, padded sig, non-canonical trailing bits, unknown top-level field, non-string extra, extra: null, v: 2.0, NaN at verify, NFD string, lone surrogate, malformed and calendar-invalid timestamps, oversize pin, kid / model over the field caps), RECORD_MISMATCH, COLLECTION_MISMATCH, and TENANT_MISMATCH. Each entry carries a description and an expected_failure code matching a v2 VerifyError enum value.

  Optional per-entry inputs consumers must honor: tampered_vec_b64 / nan_vec_b64 with vec_dtype (the dtype of the supplied bytes, which may be wider than the pin's) and vec_dim; tampered_source; expected_model; expected_model_hash; expected_record_id / expected_collection_id / expected_tenant_id; key_valid_from / key_valid_until (register the key with that window); and trusted_time (pass as the verifier-trusted instant).

Protocol v1 (legacy)

• v1.json — positive fixtures from the v1 protocol. These verify ONLY under vectorpin.LegacyV1Verifier (the opt-in migration verifier). The strict v2 verifier rejects them as UNSUPPORTED_VERSION.

• negative_v1.json — one v1 negative fixture (tampered vector). Retained for migration regression testing.

These v1 files are frozen — do NOT regenerate them. The v1 wire format is closed and the existing signatures must remain byte-for-byte verifiable.

Reproducing

The v2 signing key is deterministic:

• seed: AAECAwQFBgcICQoLDA0ODxAREhMUFRYXGBkaGxwdHh8 (bytes 0x00..0x1f in URL-safe base64, no padding)
• key id: test-vectors-v2-2026-05

The v1 fixtures used the same seed under key id test-vectors-2026-05.

Re-running scripts/generate_test_vectors.py regenerates v2.json and negative_v2.json byte-for-byte from the deterministic seed. If your port disagrees, the canonicalization or signing algorithm is off.

The deterministic seed contract is load-bearing for cross-language parity: every Rust/TypeScript test that consumes these fixtures re-derives the signature from the same seed and compares against the published bytes. Do not change the seed without coordinating with every implementation and bumping the cross-language compat-test expectations.

The seeds are published intentionally — these fixtures are public test data, not production keys.

Domain tag (v2)

v2 pins are signed over DOMAIN_TAG || canonical_json(header) where DOMAIN_TAG = b"vectorpin/v2\x00" (13 bytes — 12 ASCII characters "vectorpin/v2" plus one NUL terminator). Ports MUST match those exact bytes. See src/vectorpin/attestation.py for the canonical constant.