ONNX quantization inspection

Read the serialized quantization contract before inferring execution.

ONNX can serialize QDQ boundaries, QOperator compute nodes, and bound scale and zero-point tensors. Those facts do not by themselves prove runtime fusion or executed integer arithmetic.

Reproduce the public fixture

npx -y deepbom@1.97.4 audit \
  "./corpus/onnx-extension-contract-corpus/deepbom-static-per-axis-qdq.onnx" \
  --expected-sha256 c19f7155f030dc396aad04909b7429b8a4d6d1cb51fcd53a84e3c97ae9678d0e \
  --section quantization --json

The 289-byte source fixture is committed in the public repository. Release 1.97.4 classifies its representation as static_qdq_representation, binds one QuantizeLinear and one DequantizeLinear node, decodes scale values [0.125, 0.25, 0.5] and zero points [128, 127, 126], and derives per-axis parameterization on axis 1.

Interpret the result

Observed	Artifact bytes, initializer values, data types, shapes, and operator attributes.
Derived	Parameter binding, granularity, axis-cardinality consistency, and static representation classification.
Not asserted	Runtime graph fusion, execution-provider selection, kernel selection, and executed integer arithmetic.

Use an assistant

After connecting the DEEPBOM Skill or plugin, ask: Check how this ONNX model is quantized without executing it, and include the file hash. A successful answer should preserve the observed/derived boundary above.