Tidy comments, remove serve.py, sharpen teacher prompt

README.md

@@ -36,7 +36,6 @@ src/routelet/
 
   evaluate.py        the Claude LLM baseline (the accuracy ceiling), one call per row
   teacher.py         shared taxonomy prompt + schema + classify(), used by eval + ingest
-  serve.py           optional FastAPI endpoint (stub; Aegis runs the ONNX in-process)
 
 Scripts/             runnable tools (not library code)
   pull_samples.py    pull redacted samples from the proxy's R2 -> one JSONL

pyproject.toml

@@ -13,12 +13,6 @@ train = [
     "datasets",
     "torch",
 ]
-# Only needed if you build out the optional HTTP endpoint (src/routelet/serve.py).
-serve = [
-    "fastapi",
-    "uvicorn",
-    "pydantic",
-]
 
 [build-system]
 requires = ["hatchling"]

src/routelet/export_onnx.py

@@ -156,18 +156,14 @@ def main() -> None:
     print(f"ST pooling mode: {pooling_mode}")
     print(f"ST modules: {[type(m).__name__ for m in body]}")
 
-    # Build the exportable wrapper.
     wrapper = EncoderWithPooling(body)
     wrapper.eval()
-
-    # Export fp32 ONNX.
     _export_embedder(wrapper, EMBEDDER_OUT)
 
-    # Quantize to int8.
     quantize_dynamic(EMBEDDER_OUT, EMBEDDER_INT8_OUT, weight_type=QuantType.QInt8)
     print(f"[quantize] saved {EMBEDDER_INT8_OUT}")
 
-    # Load fitted temperature T from the file written by train_setfit.py.
+    # Temperature is fitted and written by train_setfit.py; absent means no scaling.
     temp_path = Path(TEMPERATURE_FILE)
     if temp_path.exists():
         temperature = json.loads(temp_path.read_text())["temperature"]
@@ -176,7 +172,7 @@ def main() -> None:
         temperature = 1.0
         print(f"[temperature] {TEMPERATURE_FILE} not found, using T=1.0 (no scaling)")
 
-    # Dump LR head to JSON. classes_ order is what argmax maps to.
+    # classes_ order is the column order the Rust consumer's argmax maps back to labels.
     labels = head.classes_.tolist()
     head_data = {
         "coef": head.coef_.tolist(),
@@ -194,11 +190,9 @@ def main() -> None:
     shutil.copy(f"{SETFIT_DIR}/tokenizer.json", TOKENIZER_OUT)
     print(f"[tokenizer] copied -> {TOKENIZER_OUT}")
 
-    # Print I/O signature.
     fp32_sess = ort.InferenceSession(EMBEDDER_OUT, providers=["CPUExecutionProvider"])
     _print_session_info(fp32_sess)
 
-    # File sizes.
     fp32_bytes = Path(EMBEDDER_OUT).stat().st_size
     int8_bytes = Path(EMBEDDER_INT8_OUT).stat().st_size
     print("\nfile sizes:")

src/routelet/teacher.py

@@ -23,20 +23,26 @@
 
 # Condensed from docs/taxonomy.md. Kept stable so it caches across the
 # per-command calls.
-SYSTEM = """You are an intent router. Classify each user command into exactly one of five intents:
+SYSTEM = """You are an intent classifier for short voice commands. Classify each into one of five
+intents. First strip filler words (um, uh, like) and self-corrections, then classify.
 
-- find_action: locate or operate a UI element on the current screen.
+- find_action: locate or operate a named UI element on the current screen.
 - integration: one discrete action against an app or service.
-- chat: answer from general knowledge or conversation. No app action, no personal data.
-- memory: store or recall a personal fact (storing needs an explicit remember/note/save).
-- agent: a task needing two or more chained steps or a plan.
+- chat: general knowledge or conversation. No app action, no personal data. The default.
+- memory: store or recall a personal fact. Storing needs an explicit remember/note/save.
+- agent: a task needing two or more steps or a plan.
 
-Tie-breakers for the tricky cases:
-- Two or more chained actions ("X and then Y") is agent, not integration.
-- A question answered from a stored personal fact is memory; from world knowledge it is chat.
-- A UI verb (click/tap/scroll) on a named element is find_action, even if an app is named.
+Apply these boundary rules first for the tricky cases:
+- Steps: two or more chained actions ("X and then Y"), or an implied multi-step task
+  needing a plan ("book me a restaurant"), is agent, not integration.
+- Source: answered from a stored personal fact is memory; from world knowledge it is chat.
+- Storing: "remember/note/save X" is memory even when it looks like another intent.
+- A UI verb (click/tap/scroll/select) on a named element is find_action, even if an app is named.
 - Playback (skip, pause, next, volume) is integration, not find_action, unless a button is named.
-- "explain how to..." or "talk me through..." is chat, even when it names an app action."""
+- "Explain how to..." or "talk me through..." is chat, even when it names an app action.
+
+If a command still fits more than one after these rules, pick the first match in this order:
+agent, memory, integration, find_action, chat."""
 
 INTENT_SCHEMA = {
     "type": "object",

src/routelet/train_setfit.py

@@ -70,8 +70,8 @@ def calibrate_temperature(
 ) -> float:
     """Fit a scalar temperature T on calib_examples.
 
-    Minimizes NLL of softmax(logits / T) against true labels using
-    scipy.optimize.minimize_scalar with Brent's method (bounded to [0.05, 20]).
+    Minimizes NLL of softmax(logits / T) against true labels with
+    scipy.optimize.minimize_scalar (bounded method) over T in [0.5, 20.0].
 
     Returns T > 0.
 
@@ -176,7 +176,7 @@ def threshold_analysis(
     n = len(test_examples)
 
     def _print_table(conf: np.ndarray, label: str) -> None:
-        print(f"\n--- confidence threshold analysis (124-row holdout, {label}) ---")
+        print(f"\n--- confidence threshold analysis ({n}-row holdout, {label}) ---")
         print(
             f"{'threshold':>10}  {'deferred':>10}  {'defer%':>8}  "
             f"{'kept_acc':>10}  {'defer_acc':>10}"