scorers.py

#!/usr/bin/env python3
# G1 — SCORERS DE LA BATTERIE + AUTO-VALIDATION SUR SORTIES SYNTHÉTIQUES.
# Règle du §22 : le pipeline d'extraction doit passer 100 % de ses cas synthétiques AVANT
# tout run modèle. 4 pièges de motifs en 1 jour (\pm vs ±, «24» dans markup, + 2 historiques)
# ⟹ les scorers sont du code TESTÉ, pas des regex ad hoc.
import json, re, sys, unicodedata
from os.path import normpath

# ── scorer T1 : extraire ANSWER: <entier> ou UNSURE, classes {exact,unsure,tronque,format_fail,degenere}


def detecter_degenerescence(txt, seuil=4):
    """Un 8-gramme de MOTS qui apparaît ≥seuil fois = boucle (quel que soit sa période —
    la v1 ne comparait que des fenêtres ADJACENTES et ratait toute période ≠ 8 : bug attrapé
    par les cas synthétiques de G1, jamais par un run modèle). Court-circuit : boucle de
    caractères (zéros du 09/08)."""
    compact = txt.replace(" ", "").replace("\n", "")
    if len(compact) >= 40 and len(set(compact)) <= 2:
        return True
    mots = txt.split()
    if len(mots) < 8 * 2:
        return False
    from collections import Counter
    c = Counter(tuple(mots[i:i + 8]) for i in range(len(mots) - 7))
    return max(c.values()) >= seuil


def _normaliser_nombre(s):
    s = unicodedata.normalize("NFKC", s)
    s = s.replace("\\boxed{", "").replace("}", "").replace("$", "")
    s = s.replace("\\(", "").replace("\\)", "").replace("**", "").replace("*", "")
    s = s.replace(" ", " ").replace(" ", " ")          # espaces insécables
    s = re.sub(r"(?<=\d)[ _](?=\d{3}\b)", "", s)                  # 1 080 -> 1080
    s = s.strip().rstrip(".")
    return s


def extraire_answer(txt):
    """-> (classe, valeur|None). Ne regarde QUE après le DERNIER </think>."""
    if "<think>" in txt and "</think>" not in txt:
        return ("tronque", None)
    rep = txt.split("</think>")[-1]
    if detecter_degenerescence(rep):
        return ("degenere", None)
    # la DERNIÈRE occurrence de ANSWER fait foi (le modèle peut se corriger), même en
    # milieu de ligne (cas attrapé par G1 : «pas de think du tout ANSWER: 11»)
    occs = list(re.finditer(r"(?i)\banswer\b\s*[:=]?\s*([^\n]+)", rep))
    if not occs:
        return ("format_fail", None)
    val = _normaliser_nombre(occs[-1].group(1))
    if val.upper() == "UNSURE":
        return ("unsure", None)
    m = re.fullmatch(r"[+-]?\d+", val)
    if m:
        return ("exact", int(m.group(0)))
    return ("format_fail", None)


# ── scorer T2 : FINAL: <expression sympy> + équivalence contre liste fermée de cibles

def extraire_final(txt):
    if "<think>" in txt and "</think>" not in txt:
        return ("tronque", None)
    rep = txt.split("</think>")[-1]
    lignes = [l.strip() for l in rep.splitlines() if l.strip().upper().startswith("FINAL")]
    if not lignes:
        return ("format_fail", None)
    return ("ok", re.sub(r"(?i)^final\s*[:=]?\s*", "", lignes[-1]).strip())


def equiv_sympy(candidat, cibles):
    """candidat (str) équivalent à L'UNE des cibles (liste FERMÉE de conventions) ?"""
    import sympy
    from sympy.parsing.sympy_parser import parse_expr
    try:
        c = parse_expr(candidat.replace("^", "**"), evaluate=True)
    except Exception:
        return False
    for t in cibles:
        try:
            cible = parse_expr(str(t), evaluate=True)
            if sympy.simplify(c - cible) == 0:
                return True
            # reponse DECIMALE contre cible rationnelle : residu flottant ~1e-15 que
            # simplify ne reduit pas — tolerance numerique RELATIVE (cas attrape par
            # le pilote T2 : '15.666666666666666' vs 47/3)
            num_c, num_t = complex(sympy.N(c, 20)), complex(sympy.N(cible, 20))
            if abs(num_c - num_t) <= 1e-9 * max(1.0, abs(num_t)):
                return True
        except Exception:
            continue
    return False


# ── scorer T5 : deep-equal de tool-calls après parse (jamais de diff de chaînes)

def _canon(o):
    if isinstance(o, dict):
        return {unicodedata.normalize("NFC", k): _canon(v) for k, v in sorted(o.items())}
    if isinstance(o, list):
        return [_canon(x) for x in o]
    if isinstance(o, str):
        s = unicodedata.normalize("NFC", o)
        return normpath(s) if s.startswith("/") else s
    if isinstance(o, float) and o.is_integer():
        return int(o)
    return o


def json_deep_equal(a_txt, b_obj):
    try:
        a = json.loads(a_txt)
    except Exception:
        return ("invalide", False)
    return ("ok", _canon(a) == _canon(b_obj))


# ═════════ AUTO-VALIDATION ═════════
def _tests():
    T = []
    A = extraire_answer
    # T1 — 30 cas
    T += [("t1-01", A("blah</think>\nANSWER: 248") == ("exact", 248)),
          ("t1-02", A("<think>je pense ANSWER: 999</think>\nANSWER: 24") == ("exact", 24)),
          ("t1-03", A("</think>ANSWER: \\boxed{248}") == ("exact", 248)),
          ("t1-04", A("</think>ANSWER: $1080$") == ("exact", 1080)),
          ("t1-05", A("</think>ANSWER: 1 080") == ("exact", 1080)),
          ("t1-06", A("</think>ANSWER: 1 080") == ("exact", 1080)),
          ("t1-07", A("</think>ANSWER: UNSURE") == ("unsure", None)),
          ("t1-08", A("</think>answer: unsure") == ("unsure", None)),
          ("t1-09", A("<think>coupé en plein raisonnement") == ("tronque", None)),
          ("t1-10", A("</think>the answer is 248")[0] == "format_fail"),
          ("t1-11", A("</think>ANSWER: twenty-four")[0] == "format_fail"),
          ("t1-12", A("</think>" + "0" * 300)[0] == "degenere"),
          ("t1-13", A("</think>ANSWER: -14") == ("exact", -14)),
          ("t1-14", A("</think>ANSWER: 24 8")[0] == "format_fail"),
          ("t1-15", A("</think>ANSWER: 248.") == ("exact", 248)),
          ("t1-16", A("</think>ANSWER : 248") == ("exact", 248)),
          ("t1-17", A("</think>ANSWER= 248") == ("exact", 248)),
          ("t1-18", A("</think>Réflexion.\nANSWER: 7\nvoila") == ("exact", 7)),
          ("t1-19", A("</think>ANSWER: 24\nANSWER: 248") == ("exact", 248)),
          ("t1-20", A("</think>ANSWER: **248**") == ("exact", 248)),
          ("t1-21", A("pas de think du tout ANSWER: 11") == ("exact", 11)),
          ("t1-22", A("</think>ANSWER: \\(14\\)") == ("exact", 14)),
          ("t1-23", A("</think>ANSWER: 2,5")[0] == "format_fail"),        # T1 = entiers seulement
          ("t1-24", A("</think>ANSWER: 0") == ("exact", 0)),
          ("t1-25", A("</think>ANSWER: 7920") == ("exact", 7920)),
          ("t1-26", A("</think>ANSWER:248") == ("exact", 248)),
          ("t1-27", A("</think>\n\nANSWER: 496\n") == ("exact", 496)),
          ("t1-28", A("</think>" + ("les mots se repetent ici encore et encore oui " * 30))[0] == "degenere"),
          ("t1-29", A("<think>a</think>ok<think>b</think>ANSWER: 3") == ("exact", 3)),
          ("t1-30", A("</think>ANSWER: 1_080") == ("exact", 1080))]
    # T2 — 12 cas
    F = extraire_final
    T += [("t2-01", F("</think>FINAL: 11/(16*pi**2)")[1] == "11/(16*pi**2)"),
          ("t2-02", F("<think>coupé")[0] == "tronque"),
          ("t2-03", F("</think>rien de final")[0] == "format_fail"),
          ("t2-04", equiv_sympy("11/(16*pi**2)", ["11/(16*pi**2)"])),
          ("t2-05", equiv_sympy("(11)/(16*pi^2)", ["11/(16*pi**2)"])),
          ("t2-06", equiv_sympy("0.6875/pi**2", ["11/(16*pi**2)"])),
          ("t2-07", not equiv_sympy("7/(16*pi**2)", ["11/(16*pi**2)"])),
          ("t2-08", equiv_sympy("11 - 2*3/3", ["9"])),
          ("t2-09", not equiv_sympy("n'importe quoi(", ["9"])),
          ("t2-10", equiv_sympy("22/2", ["11", "11/(16*pi**2)"])),
          ("t2-11", F("</think>FINAL: x + FINAL: y")[1] == "x + FINAL: y"),
          ("t2-12", equiv_sympy("sqrt(4)", ["2"])),
          ("t2-13", equiv_sympy("15.666666666666666", ["47/3"])),
          ("t2-14", not equiv_sympy("15.67", ["47/3"])),
          ("t2-15", equiv_sympy("0.16460905349794238683", ["40/243"]))]
    # T5 — 12 cas
    J = json_deep_equal
    ref = {"name": "read_file", "arguments": {"path": "/a/b.md", "limit": 200}}
    T += [("t5-01", J('{"name":"read_file","arguments":{"path":"/a/b.md","limit":200}}', ref) == ("ok", True)),
          ("t5-02", J('{"arguments":{"limit":200,"path":"/a/b.md"},"name":"read_file"}', ref) == ("ok", True)),
          ("t5-03", J('{"name":"read_file","arguments":{"path":"/a//b.md","limit":200}}', ref) == ("ok", True)),
          ("t5-04", J('{"name":"read_file","arguments":{"path":"/a/b.md","limit":200.0}}', ref) == ("ok", True)),
          ("t5-05", J('{"name":"read_file","arguments":{"path":"/a/c.md","limit":200}}', ref) == ("ok", False)),
          ("t5-06", J('pas du json', ref)[0] == "invalide"),
          ("t5-07", J('{"name":"write_file","arguments":{"path":"/a/b.md","limit":200}}', ref) == ("ok", False)),
          ("t5-08", J(json.dumps({"name": "read_file", "arguments": {"path": "/a/b.md", "limit": 200}}), ref) == ("ok", True)),
          ("t5-09", J('{"name":"read_file","arguments":{"path":"/caf\\u00e9/x.md","limit":1}}',
                      {"name": "read_file", "arguments": {"path": "/café/x.md", "limit": 1}}) == ("ok", True)),
          ("t5-10", J('{"name":"read_file","arguments":{"path":"/a/b.md"}}', ref) == ("ok", False)),
          ("t5-11", J('{"name":"read_file","arguments":{"path":"/a/b.md","limit":"200"}}', ref) == ("ok", False)),
          ("t5-12", J('[1,2,3]', [1, 2, 3]) == ("ok", True))]
    # T6 — 10 cas détecteur
    D = detecter_degenerescence
    T += [("t6-01", D("0" * 300)),
          ("t6-02", D("dekameters " * 60)),
          ("t6-03", not D("une liste : " + " ".join(f"item{i} valeur{i}" for i in range(60)))),
          ("t6-04", not D("texte normal de physique sur le confinement et les boucles de Wilson qui ne se repete pas du tout car chaque mot change")),
          ("t6-05", D(("exactement la meme phrase de huit mots ici oui " * 25))),
          ("t6-06", not D("ANSWER: 248")),
          ("t6-07", not D(" ".join(str(i) for i in range(200)))),
          ("t6-08", D("ab" * 200)),
          ("t6-09", not D("11 11 11")),                       # trop court pour conclure
          ("t6-10", not D(", ".join(f"fichier_{c}.md" for c in "abcdefghijklmnopqrstuvwxyz")))]
    return T


if __name__ == "__main__":
    tests = _tests()
    ko = [(n) for n, ok in tests if not ok]
    par_scorer = {}
    for n, ok in tests:
        k = n.split("-")[0]
        a, b = par_scorer.get(k, (0, 0))
        par_scorer[k] = (a + ok, b + 1)
    for k, (a, b) in sorted(par_scorer.items()):
        print(f"  {k}: {a}/{b}")
    if ko:
        print(f"  ÉCHECS: {ko}")
        sys.exit(1)
    print(f"  G1 OK — {len(tests)}/{len(tests)} cas synthétiques passés")