fix(halacha): #81.7 — report Gwet AC1 + consensus-vs-human (κ paradox under skew)

ריצת-הפאנל החיה חשפה Fleiss κ=-0.07 למרות 97.5% הסכמה-גסה (28/40 פה-אחד, 11/40 רוב).
זה אינו חוסר-אמינות אלא **פרדוקס-הקאפא**: ה-marginal של is_holding מוטה קיצונית
(≈הכול True, כמו 93/100 ה-keep בתוויות-האנוש), וכש-Pe→1 גם κ→0 (Feinstein & Cicchetti
1990, "high agreement, low kappa").

- gwet_ac1(): מדד הסכמה עמיד-שכיחות (Gwet 2008) — אותו Pa כמו Fleiss, אומדן-מקריות שונה
  (2·p·(1-p)). הופך לכותרת; Fleiss κ עדיין מודווח לשקיפות + raw 3/3.
- consensus-vs-HUMAN: כשקיים תיוג-יו"ר, הדוח מודד התאמת-הקונצנזוס מולו (תוקף חיצוני).
  אימות בפועל על 100 תוויות-היו"ר: 29/29 = 100% התאמה.

invariants: ללא שינוי בהתנהגות-הכתיבה; מטריקה בלבד. tests: 21 (3 חדשות, כולל מקרה-פרדוקס מפורש).
מקור: Gwet 2008 (AC1) · Feinstein & Cicchetti 1990.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

mcp-server/tests/test_goldset_panel_consensus.py

@@ -81,6 +81,31 @@ def test_fleiss_kappa_empty_returns_none():
     assert g.fleiss_kappa([]) is None
 
 
+# ── gwet_ac1() ────────────────────────────────────────────────────────────────
+
+def test_gwet_ac1_perfect_agreement():
+    rows = [(3, 0), (3, 0), (0, 3), (0, 3)]
+    assert g.gwet_ac1(rows) == pytest.approx(1.0)
+
+
+def test_gwet_ac1_resolves_the_kappa_paradox():
+    """The headline reason AC1 exists here: under a heavily skewed marginal
+    (almost every item is_holding=True) Fleiss κ collapses to ~0 despite very
+    high observed agreement, while AC1 correctly reports near-perfect.
+    9 unanimous-yes items + 1 split → 93% observed agreement."""
+    rows = [(3, 0)] * 9 + [(2, 1)]
+    kappa = g.fleiss_kappa(rows)
+    ac1 = g.gwet_ac1(rows)
+    assert abs(kappa) < 0.1            # κ paradox: near zero
+    assert ac1 > 0.9                   # AC1: almost-perfect, matching reality
+    assert ac1 > kappa                 # AC1 strictly more faithful under skew
+
+
+def test_gwet_ac1_ragged_and_empty_return_none():
+    assert g.gwet_ac1([(3, 0), (1, 1)]) is None
+    assert g.gwet_ac1([]) is None
+
+
 # ── anonymize() ───────────────────────────────────────────────────────────────
 
 def test_anonymize_masks_case_number_and_name():

scripts/goldset_panel_label.py

@@ -119,6 +119,32 @@ def fleiss_kappa(rows: list[tuple[int, int]]) -> float | None:
     return (Pbar - Pe) / (1 - Pe)
 
 
+def gwet_ac1(rows: list[tuple[int, int]]) -> float | None:
+    """Gwet's AC1 for binary ratings (yes_count, no_count) per item.
+
+    Reported ALONGSIDE Fleiss' κ because κ suffers the "kappa paradox": under a
+    skewed marginal (here almost every halacha is is_holding=True) κ collapses
+    toward 0 even at very high observed agreement, since chance-agreement Pe is
+    also near 1. AC1 estimates chance agreement under maximum uncertainty, so it
+    is robust to prevalence and tracks the true observed agreement (Gwet 2008;
+    Feinstein & Cicchetti 1990, "high agreement, low kappa"). Same Pa as Fleiss;
+    only the chance term differs."""
+    rows = [(y, n) for (y, n) in rows if (y + n) > 0]
+    N = len(rows)
+    if N == 0:
+        return None
+    n = rows[0][0] + rows[0][1]
+    if n < 2 or any((y + n_) != n for (y, n_) in rows):
+        return None
+    Pa = sum((y * (y - 1) + nn * (nn - 1)) / (n * (n - 1)) for (y, nn) in rows) / N
+    p_yes = sum(y for (y, _) in rows) / (N * n)
+    p_no = 1.0 - p_yes
+    Pe = 2 * p_yes * p_no  # q=2 categories: (1/(q-1))·Σ π_k(1-π_k) = 2·p·(1-p)
+    if Pe >= 1.0:
+        return 1.0
+    return (Pa - Pe) / (1 - Pe)
+
+
 # ── anonymization probe (pure — unit-tested) ──────────────────────────────────
 
 _FAKE_CASE = "12345-67-89"
@@ -186,10 +212,11 @@ async def main(args: argparse.Namespace) -> int:
     tags: Counter = Counter()
     kappa_rows: list[tuple[int, int]] = []
     anon_checked = anon_stable = 0
+    chair_overlap = chair_agree = 0  # consensus vs existing human label (external validity)
 
     async with httpx.AsyncClient() as client:
         async def run(i: int, it: dict) -> None:
-            nonlocal anon_checked, anon_stable
+            nonlocal anon_checked, anon_stable, chair_overlap, chair_agree
             async with sem:
                 user = _prompt(it)
                 per, decided, tag = await panel_pass(client, user)
@@ -217,6 +244,11 @@ async def main(args: argparse.Namespace) -> int:
             if len(nv) == 3:
                 y = sum(1 for m in nv if m["is_holding"])
                 kappa_rows.append((y, 3 - y))
+            # external validity: where a human already labeled, does consensus match?
+            if (it.get("tagged_by") == "chair" and it.get("is_holding") is not None
+                    and decided is not None):
+                chair_overlap += 1
+                chair_agree += int(decided == it["is_holding"])
             mark = {"3/3": "✓✓✓", "2/3": "✓✓", "split": "⚖", "incomplete": "…"}[tag]
             astr = "" if anon_st is None else (" anon✓" if anon_st else " anon✗FLIP")
             print(f"[{i}/{len(todo)}] {it.get('case_number')}: {mark} {tag} "
@@ -227,7 +259,15 @@ async def main(args: argparse.Namespace) -> int:
             await asyncio.gather(*tasks[j : j + args.concurrency])
 
     kappa = fleiss_kappa(kappa_rows)
+    ac1 = gwet_ac1(kappa_rows)
+    raw_agree = (sum(1 for (y, n) in kappa_rows if y == 0 or n == 0) / len(kappa_rows)
+                 if kappa_rows else None)  # share of items with unanimous 3/3
     decided_n = tags["3/3"] + tags["2/3"]
+
+    def interp(k: float) -> str:
+        return ("almost-perfect" if k >= 0.8 else "substantial" if k >= 0.6
+                else "moderate" if k >= 0.4 else "fair/poor")
+
     print("\n" + "=" * 60)
     print(f"PANEL LABELING — gold-set '{args.batch}'")
     print("=" * 60)
@@ -236,10 +276,17 @@ async def main(args: argparse.Namespace) -> int:
     print(f"  ⚖ split→chair : {tags['split']}")
     print(f"  … incomplete  : {tags['incomplete']}")
     print(f"  DECIDED (labels written): {decided_n}/{len(todo)}")
-    if kappa is not None:
+    if kappa_rows:
-        interp = ("almost-perfect" if kappa >= 0.8 else "substantial" if kappa >= 0.6
+        # Report AC1 as the headline agreement metric: the is_holding marginal is
-                  else "moderate" if kappa >= 0.4 else "fair/poor")
+        # heavily skewed (≈all True), where Fleiss κ hits the "kappa paradox"
-        print(f"  Fleiss κ (3 raters, is_holding, n={len(kappa_rows)}): {kappa:.3f} ({interp})")
+        # (high agreement, near-zero κ). AC1 is prevalence-robust.
+        print(f"  inter-model agreement (n={len(kappa_rows)}, 3 raters, is_holding):")
+        print(f"    Gwet AC1   : {ac1:.3f} ({interp(ac1)})   ← headline (skew-robust)")
+        print(f"    Fleiss κ   : {kappa:.3f} ({interp(kappa)})   [paradox under skew — see code]")
+        print(f"    raw 3/3    : {raw_agree:.1%} unanimous")
+    if chair_overlap:
+        print(f"  consensus vs HUMAN labels (external validity): "
+              f"{chair_agree}/{chair_overlap} = {chair_agree / chair_overlap:.1%}")
     if anon_checked:
         rate = anon_stable / anon_checked
         print(f"  anonymization stability: {anon_stable}/{anon_checked} = {rate:.1%} "
@@ -250,7 +297,9 @@ async def main(args: argparse.Namespace) -> int:
     report.parent.mkdir(parents=True, exist_ok=True)
     report.write_text(json.dumps({
         "batch": args.batch, "labeled": len(todo), "agreement": dict(tags),
-        "decided": decided_n, "fleiss_kappa": kappa,
+        "decided": decided_n, "fleiss_kappa": kappa, "gwet_ac1": ac1,
+        "raw_unanimous": raw_agree,
+        "consensus_vs_human": {"agree": chair_agree, "overlap": chair_overlap},
         "anon_checked": anon_checked, "anon_stable": anon_stable,
     }, ensure_ascii=False, indent=2))
     print(f"\nreport → {report}")