One record for an image and its embeddings.
VER defines a serialization format, a canonical pixel normalization pipeline, an identity model, an embedding-space declaration mechanism, a metadata preservation and trust model, and four conformance levels.
VER does not define, mandate, or prefer any particular embedding model.
The L0 floor, verbatim from the fixture corpus: identity, image facts, one fully described visual space and the canonical visual embedding it anchors. conformance/valid/minimal-l0.json
Four tiers of identity, each answering a different question
Every record carries up to four tiers. Each states what it survives and what it fails under, so no consumer has to guess which one to trust for which job.
| Tier | Artifact | Survives | Fails under / Role |
|---|---|---|---|
| Byte | content_hash — SHA-256 of Asset bytes | Exact copies only | Any re-encode or metadata rewrite. Role: dedupe, C2PA binding. |
| Pixel | pixel_hash — SHA-256 of Canonical Buffer | Metadata stripping and rewriting | Recompression, resize, crop. Role: render identity. |
| Perceptual | PDQ / pHash digests | Mild recompression, resizing | Crops, heavy edits, adversaries. Role: near-duplicate clustering. |
| Semantic | Canonical visual embedding | Most benign transforms | Model change, adversarial perturbation. Role: similarity, retrieval. |
Perceptual digests and embeddings are similarity signals. They MUST NOT be used for authentication, integrity, or provenance claims — those belong to the byte tier and to signatures.
Four layers, and a determinism claim that survives them
Layers 1–3 carry embeddings, metadata and context. Layer 0 is what makes Layer 1's determinism claim true: metadata availability is never a precondition for computing a canonical visual embedding.
Identity and normalization
CPNP-1 resolves render-affecting metadata into the Canonical Buffer and computes the identity gradient.
Layer 1Canonical visual embeddings
Vectors computed from the Canonical Buffer only, each bound to a declared Embedding Space.
Layer 2Metadata and provenance
Byte-exact preservation, a reconciled view, availability states, trust tiers, redaction and C2PA binding.
Layer 3Contextual embeddings
Optional text-derived vectors, provenance-tagged by source class, never fused into Layer 1.
What a record must carry, level by level
L0 is deliberately minimal and is the thesis's own rule made testable: nothing but pixels is required to produce a conformant record.
| Capability | L0 | L1 | L2 | L3 |
|---|---|---|---|---|
| CPNP-1, content_hash, pixel_hash | M | M | M | M |
| ≥ 1 canonical_visual embedding + full space descriptor | M | M | M | M |
| Byte-exact raw metadata preservation | — | M | M | M |
| Availability states per family | — | M | M | M |
| Reconciled view, precedence, conflicts[] | — | — | M | M |
| Trust tiers; redaction-with-proof | — | — | M | M |
| Acquisition + chain; C2PA validation | — | — | — | M |
| Contextual embeddings w/ source provenance | — | — | — | M |
| Record signature (JWS/COSE over JCS) | — | — | — | S |
The profile is executable, not illustrative
Every issue code has exactly one fixture registered against it, or an explicit exemption saying why it cannot have one. A validator either reproduces the manifest exactly or it disagrees with the standard.
Numbers the working group can be held to
A registry snapshot addressed by its own digest, a decoder corpus re-derived on every commit, and reference sets whose every vector digest was measured rather than asserted.
The VER Registry
Four value spaces. 16 provenance action tokens are registered; one space is open and three are seeded-closed.
Decoder reference corpus
21 Assets that pin what a conformant decoder must produce, each with its content digest, dimensions and pixel hash.
Reference sets
10 images per set, 30 measured vector digests. A tolerance claim now has something to be measured against.
Two planes, versioned separately
The record format and the artifact release move independently. A corrective release never edits a published document in place; it annotates it.
The first published artifact set, frozen with its defects. Nothing in the directory may be modified.
A corrective release over the 1.0.0 artifact set: three schema constraints, five example edits, and the normative errata annex.
Unratified. Adds the registry, lineage, model bundle manifests and redaction commitments. Producers must not emit 1.1 records to third parties.
A record that passes is well-formed, never verified
A signature attests that a producer published a record — not that what the record claims about the world is so. The vector suite checks that a record round-trips, not that a model is accurate.
The v1.1-draft track is unratified: producers MUST NOT emit 1.1 records to third parties outside an agreed ratification trial.
ver_versionMUSTconstcpnp_versionMUSTconstrecord_idMUSTstring (uuid)created_atMUSTstring (date-time)generatorMAYobjectconformance_levelMAYenum