Class StructuralSignal

uncommon_route/signals/structural.py:47–83

Predict tier from text structure using v1's learned classifier.

class StructuralSignal:
    """Predict tier from text structure using v1's learned classifier."""

    def predict(self, row: dict[str, Any]) -> TierVote:
        messages = row.get("messages", [])
        text = _extract_last_user_message(messages)
        if not text.strip():
            return TierVote(tier_id=None, confidence=0.0)

        # Signal B is intended to classify the current user ask. Claude Code
        # injects a large system prompt and tool scaffold into every request;
        # including all of that text makes trivial prompts like "hello" look
        # complex. Short system constraints still matter, e.g. JSON output.
        system_prompt = _compact_system_prompt_for_classification(messages)
        result = classify(text, system_prompt=system_prompt)
        tier_id = _map_v1_to_v2_tier_id(result.tier, result.complexity)
        confidence = max(0.0, min(1.0, result.confidence))

        # Dampen confidence for very short prompts — structural features
        # are unreliable when there's little text to analyze.
        # Use word count for Latin scripts, character count for CJK.
        length_text = text if not system_prompt else f"{system_prompt}\n\n{text}"
        word_count = len(length_text.split())
        if word_count <= 2 and len(length_text) > 15:
            # CJK or no-space script: use character length instead
            word_count = len(length_text) // 3  # rough CJK word estimate
        if word_count <= 8:
            confidence = min(confidence, 0.50)
        elif word_count <= 15:
            confidence = min(confidence, 0.70)
        if system_prompt and tier_id is not None and tier_id >= 1:
            confidence = max(confidence, 0.70)

        if tier_id is None:
            return TierVote(tier_id=None, confidence=confidence)

        return TierVote(tier_id=tier_id, confidence=confidence)