Snapshot checked: 2026-08-12 UTC, PR #175 head
1b380dfae8bf4a83646ae59ea3d3b42d466f3858 and specification PR #990 head
58a7cc2d04105e7525b0ed369ccf0a4325dc34b2.
PR #175 now contains draft affine quantization properties and worked examples. The remaining questions distinguish localized wording corrections from schema-placement and validation decisions.
The current cdx:ai-ml:model:parameter namespace describes learned parameter
count and tuning methods. The draft schema's modelParameter object describes
external model inputs and outputs. Would a distinct namespace for interface or
tensor-format properties avoid giving parameter two meanings?
PR #175 currently uses zeroPoint and hyphenated granularity values such as
per-tensor. Candidate examples use those exact draft spellings so that
interoperability tests detect accidental drift.
PR #175 currently combines the mathematical family and symmetry classification
as affine_asymmetric and affine_symmetric. The descriptions still need to
avoid treating a zero-valued zero point as proof of symmetry or describing
symmetric signed INT8 as halving the integer range.
Should scalar values use shortest round-trip decimal strings, and should per-axis vectors use compact JSON arrays? If so, the taxonomy should state that cardinality and axis validation require a semantic validator beyond the generic property schema.
The superseded #948 branch placed properties inside modelParameter.format.
The current #990 draft instead exposes sibling modelParameter.properties and
does not expose format.properties. Candidate examples follow the sibling
location; maintainer confirmation is still required before calling it final.
#948 accepted pre-processing and post-processing. The superseding #990
branch replaces those values with raw and processed. A model-ready
quantized tensor would therefore use processed if the current vocabulary is
retained.