Add Rust-validator evidence (rust_validation.json); note authoritative #1716 validate_metadata pass

README.md

@@ -46,8 +46,9 @@ hybrid cache as two state-service groups — `full_attention` (real owner layers
 **4/9/14**) and `sliding_attention` (the 12 sliding owners, `evictable_prefix:
 true`) — over the **real** graph ports (`past_key_values.N.key` /
 `present.N.key`), with the 20 shared-KV layers correctly owning no cache. It is
-**validated against the onnx-genai PR #1716 schema**
-(`schema/inference_metadata.schema.json`). Per-layer head_dim (256/512) lives in
+**validated with the authoritative onnx-genai PR #1716 `validate_metadata` Rust
+validator** (cross-references + real ONNX graph ports, not just JSON-Schema; see
+`evidence/rust_validation.json`). Per-layer head_dim (256/512) lives in
 the graph IO. `inference_metadata.mobius.yaml` is the raw `mobius --runtime
 onnx-genai` emitter output (its own `schema_version: 1.0`) and `policies/` holds
 that workflow's sampler / termination / cache-update graphs.

evidence/rust_validation.json

@@ -0,0 +1,10 @@
+{
+  "validator": "onnx_genai_metadata validate_metadata (authoritative Rust cross-reference + graph-port validator)",
+  "onnx_genai_ref": "PR #1716 branch copilot/gemma4-e2b-metadata @ ca8d9d1",
+  "note": "jsonschema (draft 2020-12) checks structure only; this Rust validator additionally cross-checks that state-service alias keys are real state cells, layer pairing, rollback coverage, AND that declared component ports match the actual ONNX graph inputs/outputs.",
+  "results": {
+    "target_decoder": {"file": "onnx-genai-example-gemma4-e2b/inference_metadata.yaml", "result": "valid"},
+    "assistant_speculative": {"file": "onnx-genai-example-gemma4-e2b-assistant/inference_metadata.yaml (combined-package layout: target/model.onnx + model.onnx)", "result": "valid"}
+  },
+  "combined_package_layout": "The speculative workflow spans components {target, assistant}; the Rust validator enforces that all artifacts live within one package root (no symlink/.. escape) and opens each ONNX to check ports. Validated with target graph placed at target/model.onnx alongside the assistant model.onnx. The two hosting repos keep model.onnx at their roots for standalone ORT use."
+}

file_sizes.tsv

@@ -1,5 +1,5 @@
 path	bytes
-README.md	4072
+README.md	4171
 SOURCE_LICENSE.md	321
 chat_template.jinja	18569
 evidence/gen_target_v1.py	7202
@@ -11,6 +11,7 @@ evidence/l5_generation.py	6960
 evidence/onnx-genai-1716-reference-decoder.yaml	11784
 evidence/probe_hf_shapes.py	2110
 evidence/runtime_limitations.json	2371
+evidence/rust_validation.json	1178
 inference_metadata.mobius.yaml	72071
 inference_metadata.yaml	47524
 model.onnx	766755

package_size_bytes.txt

@@ -1 +1 @@
-10297033757
+10297035034