tgs-verify

tgs-verify is the open verifier for The Gray Systems custody manifests — the per-asset provenance record (origin, generator model and license, generation parameters, SHA-256 content fingerprint) that ships with every TGS training-data delivery. Point it at a manifest and the delivered files; it recomputes every fingerprint and tells you, loudly, whether the data you received is exactly the data that was documented.

Single file. Zero dependencies. Python 3.8+. Fully offline — verification never phones home and requires nothing from the supplier.

Or skip the clone: pip install tgs-verify then tgs-verify <manifest.json>.

30-second demo

git clone https://github.com/thegray-systems/tgs-verify && cd tgs-verify
python3 tgs_verify.py examples/manifest.json

tgs-verify 1.0.0 — 3 asset(s), root: examples
  ✓ assets/scene_0001.png  c0c4002a61f8…
  ✓ assets/scene_0002.png  87226cc1ea67…
  ✓ assets/scene_0003.png  be0eb7f19036…
✓ VERIFIED — 3/3 assets match their custody record

Now watch it catch tampering — one pixel changed in one asset:

python3 tgs_verify.py examples/manifest.json --root examples/tampered

  ✓ assets/scene_0001.png  …
  ✗ ALTERED   assets/scene_0002.png
      expected 87226cc1…
      actual   a41647e1…
  ✓ assets/scene_0003.png  …
✗ FAILED — 2/3 assets match their custody record

Usage

python3 tgs_verify.py <manifest.json> [--root DIR]     # verify a JSON custody manifest
python3 tgs_verify.py --csv PN-123_manifest.csv --root ./delivery   # verify a delivery CSV
python3 tgs_verify.py manifest.json --json             # machine-readable report

Exit codes: 0 verified · 1 failed · 2 manifest unreadable. Wire it into CI and a bad delivery fails your pipeline before it touches training.

What a custody manifest records

Every asset in a TGS delivery carries: a SHA-256 fingerprint of the exact delivered bytes, a declared origin (synthetic / recorded / collected), for synthetic assets the generator model, version, and the license it was operated under, optional generation parameters, and the license class the data ships under (evaluation-only, non-exclusive, training-only, exclusive, full-ip-buyout, custom). The full format is defined in SPEC.md.

The verifier checks the certificate, not the factory: it proves delivery integrity and documented provenance without exposing anything about how data is produced.

Why this exists

Training-data provenance fails quietly. A dataset changes hands, a file gets re-encoded, a folder gets "helpfully" cleaned up — and the record no longer matches reality, which surfaces months later in diligence, an audit, or a lawsuit. A custody manifest plus an open verifier makes the failure loud and immediate instead: anyone in the chain — buyer, counsel, auditor, acquirer — can re-check the delivery in seconds, forever, without trusting the supplier's word.

Roadmap

• v1.2 — Ed25519 manifest signatures (tamper-evident, attributable manifests) per SPEC §5.
• v1.1 — project.produced_by (renamed from project.supplier, which remains a deprecated alias); project block now validated; two new origin values — authored, derived (with required derived_from).
• v1.0 — license.class expanded to six classes — evaluation-only, non-exclusive, training-only, exclusive, full-ip-buyout, custom — aligned with the TGS delivery system.

License

MIT. The manifest format (SPEC.md) is open for anyone to implement — verifiers, generators, or both.