§0 Critique of the Source Thesis

The thesis is architecturally correct. Its central claim — that the primary image embedding must be a function of the visual asset alone, with metadata preserved as separate, machine-readable context — is the right call and matches both operational reality (platforms strip metadata) and the trust model of C2PA (unsigned metadata is not evidence). The three-layer separation and the refusal to replace IPTC/XMP/EXIF/C2PA are also correct.

It is, however, a thesis and not a standard, and it contains six defects that would have broken it in production. VER 1.0 corrects all six.

• The orientation paradox. “Independent of metadata” is impossible as literally stated. EXIF Orientation and ICC color profiles alter what the pixels mean: two byte-identical sensor buffers with different orientation tags render as different images, and embedding the un-rotated buffer produces vectors that disagree with every human and platform rendering of the same asset. Fix: metadata is split into render-affecting metadata (consumed exactly once during normalization, then discarded) and descriptive metadata (preserved, never consumed by the embedding layer). The determinism claim becomes precise instead of false.
• Pixels are not stable either. Platforms recompress, resize, and transcode. Pixel-level determinism does not survive distribution any better than EXIF does — the thesis's own non-determinism argument, applied consistently, indicts pixels too. Fix: a four-tier identity gradient (byte hash → normalized-pixel hash → perceptual hash → semantic embedding), each with an explicit survival profile. The canonical embedding is defined at the semantic tier, which is robust by design; the standard stops claiming determinism where only robustness exists.
• No embedding space identity. A vector is meaningless without its model, weights, preprocessing, dimensionality, dtype, normalization, and distance metric. Two conformant implementations of the thesis could emit mutually incomparable vectors and both claim compliance — “interoperable” was aspiration, not specification. This is the largest omission. Fix: mandatory Embedding Space Descriptors with pinned weight hashes and published reproducibility tolerances. Interoperability becomes testable.
• “EXIF as authoritative when present” is unsafe. EXIF is unsigned and trivially forged; treating it as authoritative imports attacker-controlled data into trusted fields such as capture time and GPS. Fix: a three-tier trust model (C2PA-verified > signed-unverified > unsigned) with explicit reconciliation precedence and preserved conflicts.
• “Preserve everything” collides with privacy law. GPS coordinates, device serial numbers, and identifiable-person annotations cannot be blindly preserved and redistributed under GDPR-class regimes. Fix: redaction-with-proof — sensitive values are replaced by salted hashes plus reason codes, so a record can prove a value existed without disclosing it, and “redacted” is distinguishable from “absent” and “stripped.”
• No normative machinery. No requirement language, no serialization, no identity model, no conformance levels, no security analysis. Fix: everything below.