§9 Per-embedding compute

1.0 §6.2 absorbs GPU, CUDA, TensorRT, and ONNX-runtime nondeterminism into a tolerance. That is the right call, but the Record then carries no record of which runtime produced a given vector, so a tolerance failure cannot be attributed and a fleet cannot be audited for precision drift.

embeddings[].compute is an optional object:

jsonCopy

"compute": { "runtime": "onnxruntime", "runtime_version": "1.22.0",
             "precision": "fp32", "device_class": "gpu" }

precision ∈ fp32 | fp16 | bf16 | int8 | mixed is the computation precision, which is independent of dtype (the storage precision): a vector computed in mixed precision and stored as fp32 is ordinary and MUST be reported that way. device_class ∈ cpu | gpu | npu | tpu | other.

compute is descriptive. It MUST NOT be used to relax a space's declared tolerance: a vector that misses the tolerance is non-conforming whatever produced it.