{
  "validator": "onnx_genai_metadata::load_metadata + validate_metadata",
  "onnx_genai_ref": "PR #1723 branch feat/native-workflow-backend @ 6746effa1",
  "supersedes": "evidence/rust_validation.json (validated against PR #1716 @ ca8d9d1)",
  "finding": "Under the #1723 validator this package's inference_metadata.yaml was rejected: 'pipeline.workflow.steps.nodes[0].outputs is missing port hidden_states.34'. The decoder component declares hidden_states.34 for a truthful graph contract, but standalone generation does not route it; only the speculative package binds it as folded_carry_seed. The validator requires every non-optional declared output port to be bound by every invoke.",
  "fix": "Declared the hidden_states.34 output port optional: true. The port stays declared, so the assistant and speculative packages keep a truthful target contract, while this workflow may legitimately leave it unbound.",
  "results": {
    "target_decoder": {
      "file": "inference_metadata.yaml",
      "result": "valid"
    },
    "generic_variant": {
      "file": "inference_metadata.mobius.yaml",
      "result": "valid"
    }
  },
  "load_check": {
    "command": "onnx-genai show <package_dir>",
    "before": "Error: Invalid argument: pipeline.workflow.steps.nodes[0].outputs is missing port 'hidden_states.34'",
    "after": "pipeline: 1 component(s) \u2014 decoder: model.onnx"
  },
  "execution_note": "This is an fp16 CUDA export. Generation was not re-measured here: the available onnx-genai build exposes only the CPU execution provider, and plain ONNX Runtime CPU running model.onnx directly (no onnx-genai involved) already returns a degenerate top token for a trivial prompt, so CPU fp16 numerics \u2014 not the workflow metadata \u2014 govern that outcome. The package's own CUDA evidence under evidence/ remains the execution record."
}