Synthetic image/text pair provenance¶

SyntheticPairManifest links image, text and label artifacts to generator version/digest, source SPDX declarations, source digests, reproducible transformation chains and an explicit validation outcome/digest. Missing or nonpermissive declarations and disconnected chains are rejected before a manifest can be constructed. Each transform records input/output, tool and configuration digests; prompts, payloads and private paths are not stored.

to_json() is canonical and byte-stable. manifest_digest detects metadata changes; it is not a signature, authorization, license determination or evidence that the declarations are truthful. require_publishable() refuses a failed validation outcome, and verify_pair_artifacts() checks the bytes of all three members offline against their declared hashes. Passing either check does not publish anything or replace owner review.

Supply synthetic assets only. SPDX normalization reuses the maintained synthetic license-identifier policy. Provenance, source rights, validator adequacy and the actual reproducibility environment must be established outside this manifest. Artifact verification accepts at most ten megabytes per member and never returns source content in diagnostic messages. No model downloads or network calls occur.