scripts/span_common.py

"""Shared span-extraction harness: word prep, BIO encode/decode, metrics.

ONE module, called identically by every bakeoff arm. If scoring code differs
between arms the comparison is worthless, so nothing in here may be
arm-specific.

Label scheme: 3 labels, O=0 / B-DISFLUENCY=1 / I-DISFLUENCY=2.

transformers>=5 tokenizer detail, verified in temp/logs/probe_api.log:

  CORRECT   tok(words, is_split_into_words=True)  with `words` pre-split
            -> correct word_ids (verified 12/12 distinct on 3 tokenizers)

  WRONG     tok(words, is_pretokenized=True)
            -> silently returns 3 tokens `[None, 0, None]`, no exception

  WRONG     tok([sentence_string], is_split_into_words=True)
            -> a 1-element list IS one "word" under this API, so every token
               gets word_id 0. This is correct behaviour, not a library bug;
               an earlier probe of mine misread it as one.
"""

import json
import re
import sys
from typing import Dict, List, Optional, Sequence, Tuple

sys.path.insert(0, "/opt/vox/sandbox/scripts")
from compiler import clean_punctuation_whitespace, verify_roundtrip  # noqa: E402

O, B, I = 0, 1, 2
LABELS = ["O", "B-DISFLUENCY", "I-DISFLUENCY"]
MAX_LEN = 128  # AGENTS.md invariant 13

# --- Gate 2 constraints (user-directed 2026-10-03) --------------------------- #
# BOTH are hard. over_deletion_rate is sentence-level ("did we edit a turn the
# user said correctly"). harmful_span_rate is span-level ("did we delete a word
# the user actually said") and is the Golden Rule's fatal failure mode, so it
# binds harder. Reporting one without gating on it is how a model gets selected
# for something it must never do.
OVER_DELETION_GATE = 1.0   # percent of gold-CLEAN turns edited anyway
HARMFUL_SPAN_GATE = 2.0    # percent of emitted spans eating kept text


# --------------------------------------------------------------------------- #
# Word-level preparation
# --------------------------------------------------------------------------- #
def split_words(text: str) -> Tuple[List[str], List[int], List[int]]:
    """Whitespace pre-split preserving exact char offsets.

    Returns (words, starts, ends) where text[starts[i]:ends[i]] == words[i].
    This mirrors gliner2's default 'whitespace' word splitter so that the span
    family and the BIO family score against identical word boundaries.
    """
    words, starts, ends = [], [], []
    for m in re.finditer(r"\S+", text):
        words.append(m.group(0))
        starts.append(m.start())
        ends.append(m.end())
    return words, starts, ends


def encode_bio(text: str, spans: Sequence[Sequence[int]]) -> List[int]:
    """Char spans -> per-word BIO labels.

    A word is INSIDE a span if any of its characters fall inside it, so a span
    boundary landing mid-word still produces a recoverable word span.
    """
    words, starts, ends = split_words(text)
    labels = [O] * len(words)
    for s, e in spans:
        hit = False
        for i, (ws, we) in enumerate(zip(starts, ends)):
            if ws < e and we > s:  # char overlap
                labels[i] = B if not hit else I
                hit = True
    return labels


def gold_word_spans(text: str, spans: Sequence[Sequence[int]]) -> List[Tuple[int, int]]:
    """Gold char spans -> word-index spans, using the same overlap rule as encode_bio.

    A word-level model CANNOT reproduce a char span that starts or ends mid-word,
    nor one that carries an absorbed orphan comma or trailing space. Scoring
    char-exact would therefore measure the representation, not the model.
    Word-level is the level the model actually predicts at.
    """
    words, starts, ends = split_words(text)
    out: List[Tuple[int, int]] = []
    for s, e in spans:
        idx = [i for i, (ws, we) in enumerate(zip(starts, ends)) if ws < e and we > s]
        if idx:
            out.append((idx[0], idx[-1]))
    return out


def decode_bio(
    text: str,
    p_inspan: Sequence[float],
    p_begin: Sequence[float],
    threshold: float,
) -> List[Tuple[int, int]]:
    """Per-word probabilities -> char spans at a tunable precision threshold.

    p_inspan[w] = 1 - P(O) at word w. A word survives if p_inspan >= threshold.
    Inside the surviving set, a word opens a new span when P(B) > P(I), which
    keeps two adjacent disfluencies from being merged into one span.

    The threshold is the single knob the Golden Rule needs: precision
    dominates recall, so we must be able to move it without retraining.
    """
    words, starts, ends = split_words(text)
    spans: List[Tuple[int, int]] = []
    cur: Optional[int] = None
    for i in range(len(words)):
        if p_inspan[i] < threshold:
            if cur is not None:
                spans.append((starts[cur], ends[i - 1]))
                cur = None
            continue
        opens = p_begin[i] > 0.5  # P(B) > P(I) within the surviving set
        if cur is None or opens:
            if cur is not None:
                spans.append((starts[cur], ends[i - 1]))
            cur = i
    if cur is not None:
        spans.append((starts[cur], ends[cur]))
    return spans


# --------------------------------------------------------------------------- #
# Slice-cut reconstruction
# --------------------------------------------------------------------------- #
def reconstruct(raw_text: str, spans: Sequence[Sequence[int]]) -> str:
    """Deterministic slice cut + Tier 1 punctuation collapse (invariant 12).

    Case is PRESERVED. Casing is Tier 1's responsibility; Tier 2 only excises.
    Forcing an initial capital here corrupted the 8% of the corpus whose gold
    is lowercase Switchboard.
    """
    out, last = [], 0
    for s, e in sorted(spans, key=lambda x: x[0]):
        out.append(raw_text[last:s])
        so_far = "".join(out).rstrip()
        rest = raw_text[e:].lstrip()
        if so_far and rest and so_far[-1].isalnum() and rest[0].isalnum():
            out.append(" ")
        last = e
    out.append(raw_text[last:])
    return clean_punctuation_whitespace("".join(out))


def same_text(a: str, b: str) -> bool:
    """Case-insensitive + punctuation-normalised equality for reconstruction scoring."""
    return clean_punctuation_whitespace(a).lower() == clean_punctuation_whitespace(b).lower()


# --------------------------------------------------------------------------- #
# Metrics
# --------------------------------------------------------------------------- #
def prf(tp: int, fp: int, fn: int) -> Dict[str, float]:
    p = tp / (tp + fp) if tp + fp else 0.0
    r = tp / (tp + fn) if tp + fn else 0.0
    f = 2 * p * r / (p + r) if p + r else 0.0
    return {"precision": round(p, 4), "recall": round(r, 4), "f1": round(f, 4)}


def score_predictions(
    gold: List[dict],
    preds: List[List[Tuple[int, int]]],
    threshold: float,
) -> Dict:
    """Full metric set, scored at WORD level (the level the model predicts at).

    The gate metric is `over_deletion_rate`.

    over_deletion_rate  = % of gold-CLEAN turns the model edited anyway.
                           This is the number that must stay under 1.0%.
    harmful_span_rate   = % of predicted spans that eat text the gold says the
                           user actually said. This is the fatal failure mode:
                           it corrupts valid words rather than merely missing a
                           disfluency.
    exact_match         = % of turns where the slice cut reproduces gold
                           clean_text. The true end-to-end number.
    """
    tp = fp = fn = 0
    clean_edited = clean_total = 0
    harmful = pred_total = 0
    exact = 0
    per_cat: Dict[str, Dict[str, int]] = {}

    for ex, sp in zip(gold, preds):
        raw = ex["raw_text"]
        gw = set(gold_word_spans(raw, [s["span"] for s in ex.get("spans", [])]))
        pw = set(gold_word_spans(raw, sp))
        tp += len(gw & pw)
        fp += len(pw - gw)
        fn += len(gw - pw)

        # harmful = predicted span covering any char NOT inside a gold span
        gold_chars = set()
        for s in ex["spans"]:
            gold_chars.update(range(s["span"][0], s["span"][1]))
        for s, e in sp:
            pred_total += 1
            if any(c not in gold_chars for c in range(s, e)):
                harmful += 1

        cat = ex["category"]
        d = per_cat.setdefault(cat, {"n": 0, "exact": 0, "edited": 0, "tp": 0, "fp": 0, "fn": 0})
        d["n"] += 1
        d["tp"] += len(gw & pw)
        d["fp"] += len(pw - gw)
        d["fn"] += len(gw - pw)
        if same_text(reconstruct(raw, sp), ex["clean_text"]):
            exact += 1
            d["exact"] += 1
        if cat == "CLEAN":
            clean_total += 1
            if sp:
                clean_edited += 1
                d["edited"] += 1

    return {
        "threshold": threshold,
        "span_word": prf(tp, fp, fn),
        "over_deletion_rate": round(100.0 * clean_edited / clean_total, 3) if clean_total else 0.0,
        "harmful_span_rate": round(100.0 * harmful / pred_total, 3) if pred_total else 0.0,
        "exact_match": round(100.0 * exact / len(gold), 3) if gold else 0.0,
        "n": len(gold),
        "per_category": {
            k: {
                "n": v["n"],
                "exact_match": round(100.0 * v["exact"] / v["n"], 2) if v["n"] else 0.0,
                "span_word": prf(v["tp"], v["fp"], v["fn"]),
            }
            for k, v in sorted(per_cat.items())
        },
    }


def load_jsonl(path: str) -> List[dict]:
    with open(path) as f:
        return [json.loads(l) for l in f if l.strip()]


def sweep_thresholds(gold, preds_by_thr, thresholds: Sequence[float]) -> Dict:
    """Score every threshold, then pick the operating point.

    Selection rule, in order:
      1. Keep only thresholds where over_deletion_rate < OVER_DELETION_GATE
         AND harmful_span_rate < HARMFUL_SPAN_GATE.
      2. Among those, take the highest exact_match (the Golden Rule: precision
         first, and exact_match is the only metric a user actually experiences).
      3. If NOTHING passes both gates, do not silently pick the best-looking
         number. Report the most conservative threshold -- the one that minimises
         the fatal metric -- and mark `no_threshold_satisfies_gate` so the gap
         is visible instead of buried.
    """
    rows = []
    for t in thresholds:
        m = score_predictions(gold, preds_by_thr[t], t)
        m["passes_over_deletion_gate"] = m["over_deletion_rate"] < OVER_DELETION_GATE
        m["passes_harmful_gate"] = m["harmful_span_rate"] < HARMFUL_SPAN_GATE
        m["passes_gate"] = m["passes_over_deletion_gate"] and m["passes_harmful_gate"]
        rows.append(m)

    ok = [r for r in rows if r["passes_gate"]]
    if ok:
        best = max(ok, key=lambda r: r["exact_match"])
        best["no_threshold_satisfies_gate"] = False
    else:
        best = min(rows, key=lambda r: (r["harmful_span_rate"], r["over_deletion_rate"]))
        best = dict(best)
        best["no_threshold_satisfies_gate"] = True
        best["selection_note"] = (
            f"NO threshold satisfies both gates "
            f"(over_deletion<{OVER_DELETION_GATE}%, harmful_span<{HARMFUL_SPAN_GATE}%). "
            f"Reported point is the most conservative available: "
            f"harmful={best['harmful_span_rate']}% over_deletion={best['over_deletion_rate']}%."
        )
    return {"sweep": rows, "best": best, "gates": {
        "over_deletion_rate": OVER_DELETION_GATE, "harmful_span_rate": HARMFUL_SPAN_GATE}}