Upload README.md with huggingface_hub

README.md

@@ -56,10 +56,14 @@ risk:       low | medium | high
 - `task_type` → model specialty (e.g. design → Claude, systems → GPT, docs → small)
 - `risk` → tier bumper (easy + high-risk still routes capable) and confirmation gate
 
-> **Note on the ONNX export:** the published `.onnx` graphs emit the **three
-> classification heads only** (`complexity`, `task_type`, `risk` logits). The
-> regression sub-dimensions exist in the PyTorch model (`model.pt`) but are not in
-> the ONNX outputs yet — load `model.pt` with the training code if you need them.
+The ONNX graphs emit the three classification logits plus all eight regression
+sub-dimension scores. Output names are:
+
+- `complexity_logits`, `task_type_logits`, `risk_logits`
+- `complexity_sub_reasoning_depth`, `complexity_sub_scope_breadth`,
+  `complexity_sub_domain_knowledge`, `complexity_sub_spec_completeness`
+- `risk_sub_security_surface`, `risk_sub_data_sensitivity`,
+  `risk_sub_production_exposure`, `risk_sub_reversal_cost`
 
 Routing is **kickoff-only**: classify once at task start and lock the model for
 the whole task cycle (no per-turn re-routing → no context thrash).
@@ -125,11 +129,19 @@ def classify(text: str) -> dict:
         result[name.replace("_logits", "")] = {
             "label": labels[i], "confidence": round(float(p[i]), 4),
         }
+    result["complexity_sub"] = {
+        name.replace("complexity_sub_", ""): round(float(out[name][0]), 4)
+        for name in out if name.startswith("complexity_sub_")
+    }
+    result["risk_sub"] = {
+        name.replace("risk_sub_", ""): round(float(out[name][0]), 4)
+        for name in out if name.startswith("risk_sub_")
+    }
     return result
 
 print(classify("refactor JWT key rotation in prod"))
-# {'complexity': {'label': 'medium', ...}, 'task_type': {'label': 'refactor', ...},
-#  'risk': {'label': 'medium', ...}}
+# {'complexity': {'label': ...}, 'task_type': {'label': ...}, 'risk': {'label': ...},
+#  'complexity_sub': {'reasoning_depth': ...}, 'risk_sub': {'security_surface': ...}}
 ```
 
 ## Evaluation
@@ -184,7 +196,11 @@ action without a human gate; languages other than English (trained on English).
   the intended path past this; this checkpoint predates that loop.
 - **Synthetic-label ceiling.** Much training data is LLM-labeled; expect a
   ~75–80% ceiling per head until real disagreement signals are mixed in.
-- **ONNX omits sub-dims** (see note above).
+- **Quantized serving tradeoff.** The fp32 ONNX graph matches the PyTorch model
+  on the locked battery, including sub-dimension scores. The int8 graph is the
+  recommended low-dependency serving artifact and preserves the established v6
+  serving behavior, but dynamic quantization can move borderline labels and
+  regression values.
 
 ## Training