VER 1.0 — The Visual Embedding Recordstable
§8 Layer 3 — Contextual Embeddings
Contextual embeddings are optional, additive, and provenance-graded. Each carries role: "contextual", a text-space space_id, and a source object: { class, text_sha256, lang, provenance }.
| Source class | Origin | Determinism grade |
|---|---|---|
ocr_pixels | Text extracted from the Canonical Buffer | A — pixel-derived |
caption_auto | Model-generated caption of the Canonical Buffer (captioner pinned like a space) | A− — pixel-derived, model-dependent |
caption_human / alt_text / title / description | Embedded metadata or platform fields | B — contextual, mutable |
surrounding_text / page_context | Acquisition environment | C — environmental, untrusted |
- Contextual vectors MUST NOT alter, contribute to, or be fused into any
canonical_visualembedding. - Fusion is permitted only as a distinct
role: "fused"embedding in akind: "fused"space with a declared recipe{ inputs, weights, method }. - Consumers MUST be able to filter by role, source class, and grade — grade-aware retrieval is the point of grading.