VER reference sets — version 1

Ten images, three embedding spaces, every vector digest measured rather than asserted.

1. Overviewreference_set_version 1 · three spaces · ten images · measured 2026-08-25
2. StatusVER 1.1 is not ratified, and neither is §8.2.
3. ContentsOne directory per space, named for the space's last label with its version dropped.
4. The three setsset_manifest_sha256 is the SHA-256 of the RFC 8785 (JCS) canonical form of the manifest's parsed JSON.
5. What a manifest carries(That is clip-vit-b32/manifest.json's first entry with each digest elided after eight characters.
6. The imagesTen PNGs, each drawn by closed-form integer arithmetic over the pixel grid — no seeded RNG, no fonts, no floating point.
7. Measurement blockEvery vector digest came out of a real forward pass through verd.embed's own code path — _load_from_chain, _probe_dim, _to_pil, _forward, l2_normalize — against a checkpoint resolved from the local HuggingFace cache.
8. Vector packingvector_sha256 is SHA-256 over exactly these octets:
9. What a vector digest is and is notA vector_sha256 binds the set.
10. What this does and does not claimClaimed on every commit, by tests/test_record_refset_manifests.py in a tier with no torch and no network:
11. Reproducing this directoryThe generator is `scripts/make_refset_manifests.py`.