ui/scoring/dummy.py

"""DummyScorer — a deterministic PLACEHOLDER scorer (Task 3). NOT machine learning.

It derives the 5 axis scores from cheap surface heuristics over the user's turns so the result
card reflects the real upload instead of hardcoded numbers. Every heuristic is logged in
DECISIONS_LOG.md and is intentionally simple/transparent:

  Focus        — fewer, longer conversations read as more focused (turns-per-conversation).
  Technique    — rate of user turns using prompting moves ("act as", "step by step", examples…).
  Critical     — rate of skeptical / verifying language ("why", "source", "cite", "wrong"…).
  Interaction  — average user turns per conversation (back-and-forth depth).
  Input Quality— average user message length + presence of code fences / structure.

Same input -> same output. The real ML scorer implements the identical `Scorer.score` seam.
"""
from __future__ import annotations

import re

from ..data import AXES
from ..parsing import ParsedExport
from .interface import AxisScore, ScoreResult

_TECHNIQUE = re.compile(
    r"\b(act as|you are a|step by step|for example|e\.g\.|think|reason|options?|"
    r"role:|constraints?:|context:|few[- ]shot)\b", re.I)
_CRITICAL = {
    "skepticism": re.compile(r"\b(actually|really|sure\?|doubt|skeptic|cherry|wrong|incorrect)\b", re.I),
    "source_req": re.compile(r"\b(source|cite|citation|reference|evidence|prove|link)\b", re.I),
    "rebuttal": re.compile(r"\b(but |however|disagree|that's not|not quite|i think you)\b", re.I),
    "verify": re.compile(r"\b(verify|double[- ]check|confirm|are you sure|is that right)\b", re.I),
    "re_ask": re.compile(r"\b(again|rephrase|try again|instead|redo|differently)\b", re.I),
}
_CODE_FENCE = re.compile(r"```")
_TIPS = {
    "Focus": "Open a fresh chat per task to keep each thread on one goal.",
    "Technique": "Lean on roles, few-shot examples, and explicit step-by-step asks.",
    "Critical": "Keep pushing back; verify key claims against a second source.",
    "Interaction": "Build on the model's reasoning with follow-ups instead of restarting.",
    "Input Quality": "Lead with role + constraints + a concrete example for richer prompts.",
}


def _clamp(x: float) -> float:
    return max(0.0, min(10.0, x))


def _rate(matches: int, total: int) -> float:
    return matches / total if total else 0.0


class DummyScorer:
    """Heuristic placeholder. Implements the `Scorer` protocol."""

    def score(self, parsed: ParsedExport, progress=None) -> ScoreResult:
        convs = [c for c in parsed.conversations if any(t.role == "user" for t in c.turns)]
        user_turns = parsed.user_turns
        n_users = len(user_turns) or 1
        n_convs = len(convs) or 1
        texts = [t.text for t in user_turns]

        turns_per_conv = n_users / n_convs
        avg_len = sum(len(t) for t in texts) / n_users
        technique_hits = sum(1 for t in texts if _TECHNIQUE.search(t))
        code_turns = sum(1 for t in texts if _CODE_FENCE.search(t))
        crit_counts = {k: sum(1 for t in texts if rx.search(t)) for k, rx in _CRITICAL.items()}
        crit_total = sum(crit_counts.values())

        # --- map heuristics onto 0-10 (deterministic) ---
        focus = _clamp(3.0 + 1.6 * (turns_per_conv - 1))               # more turns/conv -> focused
        technique = _clamp(10 * _rate(technique_hits, n_users) * 2.5)   # ~40% usage -> 10
        critical = _clamp(10 * _rate(crit_total, n_users) * 1.2)        # skeptical language rate
        interaction = _clamp(2.0 + 2.0 * (turns_per_conv - 1))         # back-and-forth depth
        input_quality = _clamp(avg_len / 80.0 + 2.0 * _rate(code_turns, n_users) + 2.0)

        scores = {
            "Focus": focus, "Technique": technique, "Critical": critical,
            "Interaction": interaction, "Input Quality": input_quality,
        }
        # confidence by sample volume; Focus pinned low (cluster signal is weakest — per spec).
        conf = "high" if n_users >= 40 else "medium" if n_users >= 10 else "low"
        confidence = {a: conf for a in AXES}
        confidence["Focus"] = "low"

        axes = [
            AxisScore(a, scores[a], confidence[a], _evidence(a, texts, crit_counts), _TIPS[a])
            for a in AXES
        ]
        improvement = _improvement(scores)
        return ScoreResult(axes=axes, critical_counts=crit_counts, improvement=improvement)


def _evidence(axis: str, texts: list[str], crit_counts: dict) -> list[str]:
    """Pick 1-2 real user turns illustrating the axis; fall back to the first turns."""
    def pick(pred):
        return [t for t in texts if pred(t)][:2]

    if axis == "Technique":
        got = pick(lambda t: _TECHNIQUE.search(t))
    elif axis == "Critical":
        rx = re.compile("|".join(p.pattern for p in _CRITICAL.values()), re.I)
        got = pick(lambda t: rx.search(t))
    elif axis == "Input Quality":
        got = sorted(texts, key=len, reverse=True)[:2]
    elif axis == "Interaction":
        got = pick(lambda t: t.strip().endswith("?"))
    else:  # Focus
        got = texts[:2]
    got = got or texts[:1] or ["(no user turns found)"]
    return [_quote(t) for t in got]


def _quote(t: str, limit: int = 160) -> str:
    t = " ".join(t.split())
    if len(t) > limit:
        t = t[: limit - 1].rstrip() + "…"
    return f'"{t}"'


def _improvement(scores: dict) -> str:
    weakest = min(scores, key=scores.get)
    return f"Biggest lever: {weakest.lower()} is your lowest axis — {_TIPS[weakest].lower()}"