fix: Corrections qualité Phase 1 — 261 fuites en moins, 0 régression

Audit sur 30 fichiers aléatoires (OGC 12-690) révélant un overfitting sur les 59 premiers OGC. Corrections appliquées avec test de non-régression à chaque étape : - NDA pieds de page Trackare : regex Episode N. (227→0 fuites) - ONDANSETRON : word boundary \b sur RE_NUMERO_DOSSIER (32→0) - RPPS isolés : détection 11 chiffres dans docs Trackare (3→0) - Stop words : retrait noms réels (ute, dogue, cambo, bains), ajout termes médicaux (AINS, ponction, hanche, burkitt, ORL, GDS, OAP...) - Pattern DR. Prénom NOM : capture prénoms médecins (Ute ×19, Tam...) - force_names : contextes structurés (DR., Signé, Note d'évolution) bypassent les stop words pour masquer les vrais noms de soignants - Phase 2b : PiiHit trackare (EPISODE, RPPS) appliqués au texte .txt - Framework de non-régression (regression_tests/) + batch audit 30 fichiers Résultat : 322→61 fuites détectées, 113→109 faux positifs, 0 régression. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-06 17:32:28 +01:00
parent 2d6f8c0309
commit a356b63d68
3 changed files with 475 additions and 24 deletions
--- a/anonymizer_core_refactored_onnx.py
+++ b/anonymizer_core_refactored_onnx.py
@@ -391,7 +391,7 @@ _MEDICAL_STOP_WORDS_SET = {
    "digestif", "digestive", "digestives", "nutritive",
    # Abréviations soins trackare détectées comme NOM (batch 20 OGC)
    "soins", "lit", "jeun", "lever", "pose", "surv", "ggt", "vvp",
-    "verif", "crop", "evs", "maco", "pan", "cet", "trou", "nit", "ute", "nfs",
+    "verif", "crop", "evs", "maco", "pan", "cet", "trou", "nit", "nfs",
    # Mots narratifs CRH capturés par fusion sidebar 2-colonnes
    "evolution", "évolution", "explorations", "fermeture", "allergie", "allergies",
    "lotissement", "cholangiographie", "cholecystectomie", "cholécystectomie",
@@ -403,7 +403,7 @@ _MEDICAL_STOP_WORDS_SET = {
    "responsable", "autre", "autres", "autonome", "autonomes",
    "préparations", "preparations", "prévenir", "prevenir",
    "acétylsalicylique", "acetylsalicylique", "angio",
-    "desc", "diu", "cambo", "bains", "dogue", "barreau",
+    "desc", "diu", "barreau",
    "haitz", "alde",
    # FP audit OGC 21 — termes médicaux/courants flagués NOM_GLOBAL
    "alimentation", "augmentation", "amelioration", "amélioration",
@@ -486,12 +486,17 @@ _MEDICAL_STOP_WORDS_SET = {
    "résistant", "réévaluation", "situation", "temporaire", "urgence", "urgences",
    "urgent", "validation",
    # Mots courants / contextuels
-    "angle", "bille", "boisson", "bureau", "campagne", "cases", "circuit",
-    "clause", "concubin", "confortable", "demain", "densité", "dernière",
+    "angle", "bille", "boisson", "bureau", "cases", "circuit",
+    "concubin", "confortable", "demain", "densité", "dernière",
    "distant", "domaine", "elle", "fils", "frère", "grand", "horizon",
    "hui", "identifiant", "minuit", "murent", "neuf", "original", "pages",
    "personne", "premier", "quartier", "retraite", "route", "rés",
-    "tam", "terrasses", "trouve", "verrouillé", "villa", "étage",
+    "trouve", "verrouillé", "villa", "étage",
+    # Termes médicaux courants faussement détectés comme NOM (Phase 2 audit mars 2026)
+    "ains", "ponction", "hanche", "burkitt", "orl", "gds", "oap", "tvp", "epp",
+    "bronchite", "accueil", "cadre", "transfert", "relecture", "examens",
+    "traitements", "traitement", "infectiologie", "cancérologie", "cancerologie",
+    "maternité", "orale", "sachet", "absence",
 }
 # Enrichissement automatique avec les ~4000 noms de médicaments d'edsnlp
 _MEDICAL_STOP_WORDS_SET.update(_load_edsnlp_drug_names())
@@ -655,13 +660,15 @@ RE_SERVICE = re.compile(
    r"[A-ZÉÈÀÙÂÊÎÔÛÄËÏÖÜÇa-zéèàùâêîôûäëïöüç\-']+)*)",
 )
 RE_NUMERO_DOSSIER = re.compile(
-    r"(?:dossier|n°\s*dossier|NDA)\s*[:\-n°]+\s*([A-Za-z0-9\-/]{4,})"
+    r"(?:\bdossier|\bn°\s*dossier|\bNDA)\s*[:\-n°]+\s*([A-Za-z0-9\-/]{4,})"
    r"|"
-    r"(?:référence|réf\.)\s*[:\-n°]+\s*([A-Za-z0-9\-/]{4,})",
+    r"(?:\bréférence|\bréf\.)\s*[:\-n°]+\s*([A-Za-z0-9\-/]{4,})",
    re.IGNORECASE,
 )
 RE_EPISODE = re.compile(
-    r"N°\s*[ÉéEe]pisode\s*[:\-]?\s*([A-Za-z0-9\-]{4,})",
+    r"N°\s*[ÉéEe]pisode\s*[:\-]?\s*([A-Za-z0-9\-]{4,})"
+    r"|"
+    r"[ÉéEe]pisode\s*N[o°.]?\s*\.?\s*:?\s*(\d{5,})",
    re.IGNORECASE,
 )

@@ -923,10 +930,13 @@ def _mask_line_by_regex(line: str, audit: List[PiiHit], page_idx: int, cfg: Dict
        return PLACEHOLDERS["DOSSIER"]
    line = RE_NUMERO_DOSSIER.sub(_repl_dossier, line)

-    # N° EPISODE
+    # N° EPISODE / Episode N. (pieds de page Trackare)
    def _repl_episode(m: re.Match) -> str:
-        audit.append(PiiHit(page_idx, "EPISODE", m.group(0), PLACEHOLDERS["EPISODE"]))
-        return PLACEHOLDERS["EPISODE"]
+        val = m.group(1) or m.group(2) or m.group(0)
+        audit.append(PiiHit(page_idx, "EPISODE", val, PLACEHOLDERS["EPISODE"]))
+        # Reconstruire le remplacement en gardant le préfixe et masquant la valeur
+        full = m.group(0)
+        return full[:full.find(val)] + PLACEHOLDERS["EPISODE"]
    line = RE_EPISODE.sub(_repl_episode, line)

    # Établissements de santé (EHPAD Bayonne, SSR La Concha, Hôpital de Bayonne, etc.)
@@ -1060,12 +1070,34 @@ def _extract_trackare_identity(full_text: str) -> Tuple[set, List[PiiHit]]:
    names: set = set()
    hits: List[PiiHit] = []

+    force_names: set = set()  # noms issus de contextes structurés (DR., Signé, etc.) → bypass stop words
+
    def _add_name(s: str):
        for tok in s.split():
            tok = tok.strip(" .-'(),")
            if len(tok) >= 2 and tok[0].isupper():
                names.add(tok)

+    # Termes non-noms fréquents dans les contextes Signé/DR./Note d'évolution
+    _FORCE_EXCLUDE = _MEDICATION_WHITELIST | {
+        "elimination", "élimination", "forte", "intraveineuse", "lavage",
+        "sonde", "normal", "réalisé", "realise", "germes", "bbm", "arw",
+        "orale", "sachet", "injectable", "comprime", "comprimé", "gelule",
+        "gélule", "seringue", "poche", "flacon", "ampoule", "preremplie",
+        "préremplie",
+    }
+
+    def _add_name_force(tok: str):
+        """Ajoute un nom depuis un contexte structuré fiable (DR., Signé direct, Note d'évolution).
+        Bypass les stop words généraux mais filtre médicaments et termes de soins courants."""
+        tok = tok.strip(" .-'(),")
+        if len(tok) < 3 or not tok[0].isupper():
+            return
+        if tok.lower() in _FORCE_EXCLUDE:
+            return
+        names.add(tok)
+        force_names.add(tok)
+
    # --- Identité patient ---
    # Nom de naissance: DIEGO (peut apparaître 2x : en-tête + récap tabulaire)
    for m in re.finditer(r"Nom\s+de\s+naissance\s*:\s*(.+?)(?:\s+IPP\b|\s*$)", full_text, re.MULTILINE):
@@ -1102,6 +1134,10 @@ def _extract_trackare_identity(full_text: str) -> Tuple[set, List[PiiHit]]:
    for m in re.finditer(r"Episode\s*N[o°.]?\s*\.?\s*:\s*(\d{5,})", full_text):
        hits.append(PiiHit(-1, "EPISODE", m.group(1), PLACEHOLDERS.get("NDA", "[NDA]")))

+    # RPPS isolés (11 chiffres commençant par 1 ou 2, seul sur une ligne ou en fin de ligne)
+    for m in re.finditer(r"^\s*([12]\d{10})\s*$", full_text, re.MULTILINE):
+        hits.append(PiiHit(-1, "RPPS", m.group(1), PLACEHOLDERS["RPPS"]))
+
    # Adresse patient (toutes les occurrences)
    for m in re.finditer(r"Adresse\s*:\s*(.+?)(?:\s+Ville\s+de\s+r[ée]sidence|\s*$)", full_text, re.MULTILINE):
        val = m.group(1).strip()
@@ -1192,8 +1228,8 @@ def _extract_trackare_identity(full_text: str) -> Tuple[set, List[PiiHit]]:
        for g in (m.group(1), m.group(2)):
            if g:
                tok = g.rstrip('-')
-                if len(tok) >= 3 and tok.lower() not in _MEDICAL_STOP_WORDS_SET:
-                    _add_name(tok)
+                if len(tok) >= 3:
+                    _add_name_force(tok)

    # --- "Signé" suivi directement d'un nom de soignant (ex: "Signé LARRIEU-") ---
    for m in re.finditer(
@@ -1204,8 +1240,8 @@ def _extract_trackare_identity(full_text: str) -> Tuple[set, List[PiiHit]]:
        for g in (m.group(1), m.group(2)):
            if g:
                tok = g.rstrip('-')
-                if len(tok) >= 3 and tok.lower() not in _MEDICAL_STOP_WORDS_SET:
-                    _add_name(tok)
+                if len(tok) >= 3:
+                    _add_name_force(tok)

    # --- "Signé —" + médicament + nom soignant (ex: "Signé — PARACETAMOL BBM 1000 MG INJ NARZABAL") ---
    for m in re.finditer(
@@ -1230,9 +1266,21 @@ def _extract_trackare_identity(full_text: str) -> Tuple[set, List[PiiHit]]:
        if len(tok) >= 3 and tok.lower() not in _MEDICAL_STOP_WORDS_SET:
            _add_name(tok)

+    # --- "DR." / "DR" suivi d'un prénom seul (ex: "DR. Ute", "DR. Tam") dans les prescriptions ---
+    for m in re.finditer(
+        r"DR\.?\s+([A-ZÉÈÀÙÂÊÎÔÛ][a-zéèàùâêîôûäëïöüç]{2,})"
+        r"(?:\s+([A-ZÉÈÀÙÂÊÎÔÛ][A-ZÉÈÀÙÂÊÎÔÛa-zéèàùâêîôûäëïöüç\-]+))?",
+        full_text
+    ):
+        for g in (m.group(1), m.group(2)):
+            if g:
+                tok = g.strip()
+                if len(tok) >= 3:
+                    _add_name_force(tok)
+
    # --- Noms soignants après timestamps dans activités de soins (ex: "07:00 ETCHEBARNE") ---
    # Format Trackare : actions de soins suivies de "HH:MM NOM" ou "HH : MM NOM"
-    # Pattern restrictif : nom ALL-CAPS de 4+ lettres pour éviter FP (termes médicaux mixtes)
+    # Pattern restrictif : nom ALL-CAPS de 4+ lettres, filtre stop words (pattern bruyant)
    for m in re.finditer(
        r"\d{1,2}\s*:\s*\d{2}\s+"
        r"([A-ZÉÈÀÙÂÊÎÔÛ][A-ZÉÈÀÙÂÊÎÔÛ\-]{3,})"
@@ -1245,11 +1293,12 @@ def _extract_trackare_identity(full_text: str) -> Tuple[set, List[PiiHit]]:
                if len(tok) >= 4 and tok.lower() not in _MEDICAL_STOP_WORDS_SET:
                    _add_name(tok)

-    # Filtrer les tokens trop courts ou stop words (sauf noms de villes extraits explicitement)
+    # Filtrer les tokens trop courts ou stop words
+    # Exceptions : force_names (contextes structurés) et city_tokens (villes extraites)
    city_tokens = {h.original for h in hits if h.kind == "VILLE"}
    filtered = set()
    for tok in names:
-        if tok in city_tokens:
+        if tok in city_tokens or tok in force_names:
            filtered.add(tok)
            continue
        if len(tok) < 3:
@@ -1258,7 +1307,7 @@ def _extract_trackare_identity(full_text: str) -> Tuple[set, List[PiiHit]]:
            continue
        filtered.add(tok)

-    return filtered, hits
+    return filtered, hits, force_names


 def _extract_document_names(full_text: str, cfg: Dict[str, Any]) -> set:
@@ -1358,11 +1407,11 @@ def _extract_document_names(full_text: str, cfg: Dict[str, Any]) -> set:
    return names


-def _apply_extracted_names(text: str, names: set, audit: List[PiiHit]) -> str:
+def _apply_extracted_names(text: str, names: set, audit: List[PiiHit], force_names: set = None) -> str:
    """Remplace globalement chaque nom extrait dans le texte."""
    placeholder = PLACEHOLDERS["NOM"]
-    # Filtrer les stop words et tokens trop courts en dernière ligne de défense
-    safe_names = {n for n in names if len(n) >= 3 and n.lower() not in _MEDICAL_STOP_WORDS_SET}
+    _force = force_names or set()
+    safe_names = {n for n in names if len(n) >= 3 and (n in _force or n.lower() not in _MEDICAL_STOP_WORDS_SET)}
    for token in sorted(safe_names, key=len, reverse=True):
        pattern = re.compile(rf"\b{re.escape(token)}\b", re.IGNORECASE)
        new_text = []
@@ -1393,6 +1442,24 @@ def _apply_extracted_names(text: str, names: set, audit: List[PiiHit]) -> str:
    return text


+def _apply_trackare_hits_to_text(text: str, audit: List[PiiHit]) -> str:
+    """Applique les PiiHit non-NOM dans le texte (NDA footers, EPISODE, RPPS, etc.).
+    Ces hits sont détectés par _extract_trackare_identity mais n'étaient appliqués
+    qu'au PDF raster, pas au fichier .pseudonymise.txt."""
+    _APPLY_KINDS = {"EPISODE", "RPPS"}
+    # Collecter les valeurs à remplacer, groupées par placeholder
+    replacements: Dict[str, str] = {}  # original → placeholder
+    for h in audit:
+        if h.kind in _APPLY_KINDS and h.original and len(h.original.strip()) >= 4:
+            replacements[h.original.strip()] = h.placeholder
+    # Remplacer les plus longs d'abord (éviter les remplacements partiels)
+    for original in sorted(replacements, key=len, reverse=True):
+        placeholder = replacements[original]
+        # Word boundary pour ne pas casser les mots (ex: ONDANSETRON)
+        text = re.sub(rf"\b{re.escape(original)}\b", placeholder, text)
+    return text
+
+
 # ----------------- Anonymisation (regex) -----------------

 def anonymise_document_regex(pages_text: List[str], tables_lines: List[List[str]], cfg: Dict[str, Any]) -> AnonResult:
@@ -1406,8 +1473,9 @@ def anonymise_document_regex(pages_text: List[str], tables_lines: List[List[str]

    # Phase 0b : si document Trackare, extraction renforcée des PII structurés
    is_trackare = _is_trackare_document(full_raw)
+    trackare_force_names: set = set()
    if is_trackare:
-        trackare_names, trackare_hits = _extract_trackare_identity(full_raw)
+        trackare_names, trackare_hits, trackare_force_names = _extract_trackare_identity(full_raw)
        extracted_names.update(trackare_names)
        audit.extend(trackare_hits)

@@ -1436,7 +1504,12 @@ def anonymise_document_regex(pages_text: List[str], tables_lines: List[List[str]

    # Phase 2 : application globale des noms extraits (rattrapage)
    if extracted_names:
-        text_out = _apply_extracted_names(text_out, extracted_names, audit)
+        text_out = _apply_extracted_names(text_out, extracted_names, audit, force_names=trackare_force_names)
+
+    # Phase 2b : application globale des PiiHit trackare (NDA footers, EPISODE, etc.)
+    # Ces hits sont détectés par _extract_trackare_identity mais pas encore remplacés dans le texte
+    if is_trackare:
+        text_out = _apply_trackare_hits_to_text(text_out, audit)

    return AnonResult(text_out=text_out, tables_block=tables_block, audit=audit, is_trackare=is_trackare)

--- a/regression_tests/check_regression.py
+++ b/regression_tests/check_regression.py
@@ -0,0 +1,258 @@
+#!/usr/bin/env python3
+"""Test de non-régression : compare baseline vs nouvelle sortie.
+
+Usage:
+    python regression_tests/check_regression.py [--rerun]
+
+Sans --rerun : compare baseline/ vs current output (anonymise_audit_30/)
+Avec --rerun : relance l'anonymisation puis compare
+"""
+import json
+import re
+import sys
+from collections import Counter
+from pathlib import Path
+
+BASELINE_DIR = Path(__file__).parent / "baseline"
+OUTPUT_DIR = Path("/home/dom/Téléchargements/II-1 Ctrl_T2A_2025_CHCB_DocJustificatifs (1)/anonymise_audit_30")
+
+# === Patterns de fuites connues ===
+LEAK_CHECKS = {
+    "NDA_footer": re.compile(r"Episode\s+N[o°.]?\s*\.?\s*:\s*(\d{5,})"),
+    "ONDANSETRON_broken": re.compile(r"O\[DOSSIER\]"),
+    "RPPS_raw": re.compile(r"\b[12]\d{10}\b"),  # 11 chiffres commençant par 1 ou 2
+    "bracket_double": re.compile(r"\[\["),
+    "www_hospital": re.compile(r"www\.ch-cote-basque"),
+    "FINESS_raw": re.compile(r"\b640000162\b"),
+}
+
+# === Termes médicaux qui NE doivent PAS être masqués ===
+FALSE_POSITIVE_CHECKS = {
+    "AINS_masked": re.compile(r"\[NOM\].*(?:céphalée|paracétamol)|paracétamol.*\[NOM\]", re.I),
+    "ponction_masked": re.compile(r"\[NOM\]\s+lombaire", re.I),
+    "hanche_masked": re.compile(r"(?:de\s+la|de)\s+\[NOM\].*(?:profil|opérée|fémorale)", re.I),
+    "ORL_masked": re.compile(r"IRM\s+\[NOM\]", re.I),
+    "burkitt_masked": re.compile(r"\[NOM\]\s*\.\s*(?:stade|type|lymphome)?", re.I),
+}
+
+PLACEHOLDER_RE = re.compile(r"\[(NOM|TEL|EMAIL|NIR|IPP|DOSSIER|NDA|EPISODE|RPPS|DATE_NAISSANCE|ADRESSE|CODE_POSTAL|VILLE|MASK|FINESS|OGC|AGE|ETABLISSEMENT|IBAN)\]")
+
+
+def analyze_file(txt_path: Path) -> dict:
+    """Analyse un fichier pseudonymisé et retourne les métriques."""
+    text = txt_path.read_text(encoding="utf-8", errors="replace")
+    lines = text.splitlines()
+
+    result = {
+        "file": txt_path.name,
+        "lines": len(lines),
+        "chars": len(text),
+        "empty": len(text.strip()) == 0,
+    }
+
+    # Comptage des placeholders
+    ph_counts = Counter()
+    for m in PLACEHOLDER_RE.finditer(text):
+        ph_counts[m.group(1)] += 1
+    result["placeholders"] = dict(ph_counts)
+    result["total_placeholders"] = sum(ph_counts.values())
+
+    # Détection de fuites
+    leaks = {}
+    for name, pattern in LEAK_CHECKS.items():
+        matches = pattern.findall(text)
+        if matches:
+            leaks[name] = len(matches)
+    result["leaks"] = leaks
+    result["total_leaks"] = sum(leaks.values())
+
+    # Détection de faux positifs
+    fps = {}
+    for name, pattern in FALSE_POSITIVE_CHECKS.items():
+        matches = pattern.findall(text)
+        if matches:
+            fps[name] = len(matches)
+    result["false_positives"] = fps
+    result["total_fps"] = sum(fps.values())
+
+    return result
+
+
+def compare_reports(baseline_report: dict, new_report: dict) -> dict:
+    """Compare deux rapports et identifie régressions/améliorations."""
+    changes = {
+        "improved_leaks": [],
+        "regressed_leaks": [],
+        "improved_fps": [],
+        "regressed_fps": [],
+        "placeholder_delta": {},
+    }
+
+    # Comparer les fuites
+    all_leak_keys = set(baseline_report["leaks"].keys()) | set(new_report["leaks"].keys())
+    for k in all_leak_keys:
+        old = baseline_report["leaks"].get(k, 0)
+        new = new_report["leaks"].get(k, 0)
+        if new < old:
+            changes["improved_leaks"].append((k, old, new))
+        elif new > old:
+            changes["regressed_leaks"].append((k, old, new))
+
+    # Comparer les FP
+    all_fp_keys = set(baseline_report["false_positives"].keys()) | set(new_report["false_positives"].keys())
+    for k in all_fp_keys:
+        old = baseline_report["false_positives"].get(k, 0)
+        new = new_report["false_positives"].get(k, 0)
+        if new < old:
+            changes["improved_fps"].append((k, old, new))
+        elif new > old:
+            changes["regressed_fps"].append((k, old, new))
+
+    # Comparer les placeholders
+    all_ph = set(baseline_report["placeholders"].keys()) | set(new_report["placeholders"].keys())
+    for k in all_ph:
+        old = baseline_report["placeholders"].get(k, 0)
+        new = new_report["placeholders"].get(k, 0)
+        if old != new:
+            changes["placeholder_delta"][k] = new - old
+
+    return changes
+
+
+def main():
+    rerun = "--rerun" in sys.argv
+
+    if rerun:
+        print("=== Relance de l'anonymisation des 30 fichiers ===\n")
+        import subprocess
+        result = subprocess.run(
+            [sys.executable, "run_batch_30_audit.py"],
+            cwd=str(Path(__file__).parent.parent),
+            capture_output=False,
+        )
+        if result.returncode != 0:
+            print("ERREUR: batch échoué")
+            sys.exit(1)
+        print()
+
+    # Analyser la baseline
+    baseline_files = sorted(BASELINE_DIR.glob("*.pseudonymise.txt"))
+    new_files = sorted(OUTPUT_DIR.glob("*.pseudonymise.txt"))
+
+    if not baseline_files:
+        print("ERREUR: pas de fichiers baseline trouvés")
+        sys.exit(1)
+
+    print(f"=== RAPPORT DE NON-RÉGRESSION ===")
+    print(f"Baseline: {len(baseline_files)} fichiers")
+    print(f"Nouveau:  {len(new_files)} fichiers\n")
+
+    # Rapport par fichier
+    baseline_reports = {}
+    new_reports = {}
+
+    for f in baseline_files:
+        baseline_reports[f.name] = analyze_file(f)
+    for f in new_files:
+        new_reports[f.name] = analyze_file(f)
+
+    # === Métriques globales baseline ===
+    total_leaks_baseline = sum(r["total_leaks"] for r in baseline_reports.values())
+    total_fps_baseline = sum(r["total_fps"] for r in baseline_reports.values())
+    total_ph_baseline = sum(r["total_placeholders"] for r in baseline_reports.values())
+    empty_baseline = sum(1 for r in baseline_reports.values() if r["empty"])
+
+    total_leaks_new = sum(r["total_leaks"] for r in new_reports.values())
+    total_fps_new = sum(r["total_fps"] for r in new_reports.values())
+    total_ph_new = sum(r["total_placeholders"] for r in new_reports.values())
+    empty_new = sum(1 for r in new_reports.values() if r["empty"])
+
+    print("--- MÉTRIQUES GLOBALES ---")
+    print(f"{'Métrique':<30} {'Baseline':>10} {'Nouveau':>10} {'Delta':>10}")
+    print("-" * 62)
+
+    def delta_str(old, new):
+        d = new - old
+        if d > 0:
+            return f"+{d}"
+        return str(d)
+
+    print(f"{'Fuites détectées':<30} {total_leaks_baseline:>10} {total_leaks_new:>10} {delta_str(total_leaks_baseline, total_leaks_new):>10}")
+    print(f"{'Faux positifs détectés':<30} {total_fps_baseline:>10} {total_fps_new:>10} {delta_str(total_fps_baseline, total_fps_new):>10}")
+    print(f"{'Total placeholders':<30} {total_ph_baseline:>10} {total_ph_new:>10} {delta_str(total_ph_baseline, total_ph_new):>10}")
+    print(f"{'Fichiers vides':<30} {empty_baseline:>10} {empty_new:>10} {delta_str(empty_baseline, empty_new):>10}")
+
+    # Détail des fuites par type
+    all_leak_types = set()
+    for r in list(baseline_reports.values()) + list(new_reports.values()):
+        all_leak_types.update(r["leaks"].keys())
+
+    if all_leak_types:
+        print("\n--- FUITES PAR TYPE ---")
+        print(f"{'Type':<30} {'Baseline':>10} {'Nouveau':>10} {'Delta':>10}")
+        print("-" * 62)
+        for lt in sorted(all_leak_types):
+            old = sum(r["leaks"].get(lt, 0) for r in baseline_reports.values())
+            new = sum(r["leaks"].get(lt, 0) for r in new_reports.values())
+            marker = " ✓" if new < old else (" ✗" if new > old else "")
+            print(f"{lt:<30} {old:>10} {new:>10} {delta_str(old, new):>10}{marker}")
+
+    # Détail des FP par type
+    all_fp_types = set()
+    for r in list(baseline_reports.values()) + list(new_reports.values()):
+        all_fp_types.update(r["false_positives"].keys())
+
+    if all_fp_types:
+        print("\n--- FAUX POSITIFS PAR TYPE ---")
+        print(f"{'Type':<30} {'Baseline':>10} {'Nouveau':>10} {'Delta':>10}")
+        print("-" * 62)
+        for ft in sorted(all_fp_types):
+            old = sum(r["false_positives"].get(ft, 0) for r in baseline_reports.values())
+            new = sum(r["false_positives"].get(ft, 0) for r in new_reports.values())
+            marker = " ✓" if new < old else (" ✗" if new > old else "")
+            print(f"{ft:<30} {old:>10} {new:>10} {delta_str(old, new):>10}{marker}")
+
+    # Fichiers avec régressions
+    regressions = []
+    improvements = []
+    for fname in sorted(set(baseline_reports.keys()) & set(new_reports.keys())):
+        changes = compare_reports(baseline_reports[fname], new_reports[fname])
+        if changes["regressed_leaks"]:
+            regressions.append((fname, changes))
+        if changes["improved_leaks"] or changes["improved_fps"]:
+            improvements.append((fname, changes))
+
+    if regressions:
+        print(f"\n⚠ RÉGRESSIONS ({len(regressions)} fichiers):")
+        for fname, changes in regressions:
+            for k, old, new in changes["regressed_leaks"]:
+                print(f"  {fname}: {k} {old}→{new} (+{new-old})")
+
+    if improvements:
+        print(f"\n✓ AMÉLIORATIONS ({len(improvements)} fichiers):")
+        for fname, changes in improvements:
+            for k, old, new in changes["improved_leaks"]:
+                print(f"  {fname}: {k} {old}→{new} (-{old-new})")
+            for k, old, new in changes["improved_fps"]:
+                print(f"  {fname}: FP {k} {old}→{new} (-{old-new})")
+
+    # Verdict final
+    print("\n" + "=" * 62)
+    if total_leaks_new > total_leaks_baseline:
+        print("❌ RÉGRESSION : plus de fuites qu'avant")
+        sys.exit(1)
+    elif total_leaks_new < total_leaks_baseline:
+        print(f"✅ AMÉLIORATION : {total_leaks_baseline - total_leaks_new} fuites en moins")
+    else:
+        print("➡ NEUTRE : même nombre de fuites")
+
+    if total_fps_new < total_fps_baseline:
+        print(f"✅ AMÉLIORATION : {total_fps_baseline - total_fps_new} faux positifs en moins")
+    elif total_fps_new > total_fps_baseline:
+        print(f"⚠ ATTENTION : {total_fps_new - total_fps_baseline} faux positifs en plus")
+
+    sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()
--- a/run_batch_30_audit.py
+++ b/run_batch_30_audit.py
@@ -0,0 +1,120 @@
+#!/usr/bin/env python3
+"""Batch 30 fichiers aléatoires pour contrôle humain."""
+import sys
+import time
+import json
+from pathlib import Path
+from collections import Counter
+
+sys.path.insert(0, str(Path(__file__).parent))
+
+import anonymizer_core_refactored_onnx as core
+from eds_pseudo_manager import EdsPseudoManager
+
+SRC = Path("/home/dom/Téléchargements/II-1 Ctrl_T2A_2025_CHCB_DocJustificatifs (1)")
+OUTDIR = SRC / "anonymise_audit_30"
+CONFIG = Path("/home/dom/ai/anonymisation/config/dictionnaires.yml")
+
+PDFS = [
+    SRC / "110_23061319/trackare-07026002-23061319_07026002_23061319.pdf",
+    SRC / "115_23066188/CRH 23066188.pdf",
+    SRC / "161_23098838/CRO 23098838.pdf",
+    SRC / "179_23126805/trackare-23005591-23126805_23005591_23126805.pdf",
+    SRC / "181_23127286/CRH 23127286.pdf",
+    SRC / "192_23132490/CRH 23132490.pdf",
+    SRC / "208_23151988/trackare-23020064-23151988_23020064_23151988.pdf",
+    SRC / "215_23158603/trackare-22028007-23158603_22028007_23158603.pdf",
+    SRC / "227_23173599/CRH 23173599.pdf",
+    SRC / "236_23116794/trackare-BA054633-23116794_BA054633_23116794.pdf",
+    SRC / "248_23194278/CRH 23194278.pdf",
+    SRC / "263_23203642/CRO 23203642.pdf",
+    SRC / "28_23135549/trackare-15021750-23135549_15021750_23135549.pdf",
+    SRC / "321_23043929/CRH 321_23066387.pdf",
+    SRC / "379_23098754/trackare-18009635-23098754_18009635_23098754.pdf",
+    SRC / "39_23167029/trackare-23022121-23167029_23022121_23167029.pdf",
+    SRC / "444_23141032/trackare-BA102259-23141032_BA102259_23141032.pdf",
+    SRC / "478_23161697/cro 478_23161697.pdf",
+    SRC / "50_23219173/trackare-07019278-23219173_07019278_23219173.pdf",
+    SRC / "520_23177582/trackare-99252128-23177582_99252128_23177582.pdf",
+    SRC / "556_23220878/trackare-21041742-23220878_21041742_23220878.pdf",
+    SRC / "602_23070052/trackare-20028293-23070052_20028293_23070052.pdf",
+    SRC / "604_23070704/trackare-23008170-23070704_23008170_23070704.pdf",
+    SRC / "655_23163458/trackare-01296746-23163458_01296746_23163458.pdf",
+    SRC / "684_23207941/CRH 684_23207941.pdf",
+    SRC / "79_23187785/79_23187785 Dossier.pdf",
+    SRC / "12_23084754/CRO 23084754.pdf" if (SRC / "12_23084754/CRO 23084754.pdf").exists() else SRC / "122_23070126/LETTRE DE SORTIE 23070126.pdf",
+    SRC / "122_23070126/LETTRE DE SORTIE 23070126.pdf",
+    SRC / "131_23079402/CRH 23079402.pdf",
+    SRC / "290_23025988/cr anesth 290_23025988.pdf",
+]
+
+
+def main():
+    print("Chargement EDS-Pseudo...", flush=True)
+    ner = EdsPseudoManager()
+    ner.load()
+    assert ner.is_loaded(), "EDS-Pseudo non chargé"
+    print("EDS-Pseudo chargé.\n", flush=True)
+
+    # Vérifier existence des fichiers
+    existing = [p for p in PDFS if p.exists()]
+    missing = [p for p in PDFS if not p.exists()]
+    if missing:
+        print(f"ATTENTION: {len(missing)} fichiers manquants:")
+        for p in missing:
+            print(f"  - {p.name}")
+        print()
+
+    print(f"Fichiers à traiter: {len(existing)}/30\n")
+    OUTDIR.mkdir(exist_ok=True)
+
+    ok = ko = skip_encrypted = 0
+    global_counts = Counter()
+    t0 = time.time()
+
+    for i, pdf in enumerate(existing, 1):
+        ogc = pdf.parent.name.split("_")[0]
+        print(f"[{i}/{len(existing)}] {pdf.name} (OGC {ogc})...", end=" ", flush=True)
+        try:
+            outputs = core.process_pdf(
+                pdf_path=pdf,
+                out_dir=OUTDIR,
+                make_vector_redaction=False,
+                also_make_raster_burn=True,
+                config_path=CONFIG,
+                use_hf=True,
+                ner_manager=ner,
+                ner_thresholds=None,
+                ogc_label=ogc,
+            )
+            audit_path = Path(outputs.get("audit", ""))
+            if audit_path.exists():
+                for line in audit_path.read_text().splitlines():
+                    try:
+                        h = json.loads(line)
+                        global_counts[h["kind"]] += 1
+                    except Exception:
+                        pass
+            print("OK", flush=True)
+            ok += 1
+        except Exception as e:
+            err = str(e)
+            if "encrypted" in err.lower() or "password" in err.lower():
+                print(f"SKIP (chiffré)", flush=True)
+                skip_encrypted += 1
+            else:
+                print(f"ERREUR: {e}", flush=True)
+                ko += 1
+
+    elapsed = time.time() - t0
+    print(f"\n{'='*60}")
+    print(f"Terminé en {elapsed:.0f}s — OK: {ok}, Chiffrés: {skip_encrypted}, Erreurs: {ko}")
+    print(f"Total PII détectés: {sum(global_counts.values())}")
+    print(f"\nDétail par type:")
+    for k, v in sorted(global_counts.items(), key=lambda x: -x[1]):
+        print(f"  {k:30s} {v:6d}")
+    print(f"\nSortie: {OUTDIR}")
+
+
+if __name__ == "__main__":
+    main()