diff --git a/mcp-server/tests/test_panel_calibrate_captured.py b/mcp-server/tests/test_panel_calibrate_captured.py
new file mode 100644
index 0000000..d70cbd1
--- /dev/null
+++ b/mcp-server/tests/test_panel_calibrate_captured.py
@@ -0,0 +1,75 @@
+"""Tests for #133 / FU-5 — captured-mode calibration of the halacha panel.
+
+Covers the PURE helpers in scripts/halacha_panel_calibrate.py
+(summarize_calibration, bucket_by_round): from captured (panel ⋈ chair) pairs
+they must report the split-rate and auto-precision the panel ACTUALLY delivered
+against the chair's ground-truth, and break it down per round-day so the loop's
+trend is visible. Fully OFFLINE (no DB, no LLM, no re-voting).
+"""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[2] / "scripts"))
+import halacha_panel_calibrate as cal  # noqa: E402
+
+
+def _pair(chair, verdict, action, ts="2026-06-12T04:00:00Z"):
+    # judge votes are irrelevant to split-rate/precision here; keep them aligned
+    return {
+        "chair_keep": chair, "verdict": verdict, "applied_action": action,
+        "round_ts": ts, "rule_statement": "r",
+        "claude_vote": chair, "claude_reason": "",
+        "deepseek_vote": chair, "deepseek_reason": "",
+        "gemini_vote": chair, "gemini_reason": "",
+    }
+
+
+def test_split_rate_and_precision():
+    pairs = [
+        _pair(True, "unanimous_yes", "approved"),   # auto-correct
+        _pair(False, "unanimous_no", "rejected"),   # auto-correct
+        _pair(False, "unanimous_yes", "approved"),  # auto WRONG (false-keep)
+        _pair(True, "split", "chair"),              # escalated
+    ]
+    s = cal.summarize_calibration(pairs)
+    assert s["n"] == 4
+    assert s["escalated"] == 1
+    assert s["auto_decided"] == 3
+    assert s["split_rate"] == 0.25
+    # 3 auto-decisions, 1 wrong (false-keep) → precision 2/3
+    assert s["false_keep"] == 1 and s["false_drop"] == 0
+    assert round(s["auto_precision"], 2) == 0.67
+
+
+def test_empty_pairs_safe():
+    s = cal.summarize_calibration([])
+    assert s["n"] == 0
+    assert s["split_rate"] is None and s["auto_precision"] is None
+
+
+def test_unlabeled_pairs_filtered():
+    s = cal.summarize_calibration([_pair(None, "split", "chair")])
+    assert s["n"] == 0  # chair=None contributes no calibration signal
+
+
+def test_bucket_by_round_trend():
+    pairs = [
+        _pair(True, "unanimous_yes", "approved", ts="2026-06-10T04:00:00Z"),
+        _pair(False, "split", "chair", ts="2026-06-10T04:00:00Z"),
+        _pair(True, "unanimous_yes", "approved", ts="2026-06-12T05:00:00Z"),
+    ]
+    trend = cal.bucket_by_round(pairs)
+    days = [d for d, _ in trend]
+    assert days == ["2026-06-10", "2026-06-12"]  # sorted by day
+    assert trend[0][1]["n"] == 2 and trend[0][1]["split_rate"] == 0.5
+    assert trend[1][1]["n"] == 1 and trend[1][1]["split_rate"] == 0.0
+
+
+def test_missing_round_ts_bucketed_unknown():
+    p = _pair(True, "unanimous_yes", "approved")
+    del p["round_ts"]
+    trend = cal.bucket_by_round([p])
+    assert trend[0][0] == "unknown"
diff --git a/scripts/SCRIPTS.md b/scripts/SCRIPTS.md
index 19c9e0b..c27c1b4 100644
--- a/scripts/SCRIPTS.md
+++ b/scripts/SCRIPTS.md
@@ -62,7 +62,7 @@
 | `_pipeline_runtime.py` | python | **runtime עמידות משותף (X16 / INV-DUR1)** ל-`final_halacha_pipeline` ו-`final_learning_pipeline` (מימוש אחד, G2). עוטף רשימת-צעדים async ב-LangGraph `StateGraph` ליניארי עם `AsyncSqliteSaver` (checkpoint לכל צעד; resume מדלג על צעדים שהושלמו). **degradation חיננית:** ללא langgraph (`pip install -e ".[durable]"`) — ריצה ליניארית כמו קודם (הכפתור לא נשבר). `Step(name, run)` + `run_pipeline(steps, thread_id, checkpoint_db, fresh)`. נבדק: `mcp-server/tests/test_pipeline_runtime.py`. | מיובא ע"י סקריפטי-המסלול-הסופי |
 | `curator_apply_pipeline_branch.py` | python | **מקור-אמת לחיווט-הכפתורים של הרמס.** prompt-ה-curator חי רק ב-Paperclip DB (`agents.adapter_config.promptTemplate`). הסקריפט מקדים branch כך שיקיצה עם reason `final_learning_*`/`final_halacha_*` מריצה את ה-pipeline המתאים (HOME/DOTENV/DATA_DIR מוחלטים → DeepSeek+Gemini keys + DATA_DIR נפתרים נכון) ועוצרת, אחרת §A/§B כרגיל. idempotent (מסיר branch קודם). מחיל על שני הסוכנים (CMP+CMPA). `--verify`. **להריץ אחרי reset/יצירה-מחדש של סוכן-curator.** | אחרי reset prompt של curator |
 | `halacha_panel_audit.py` | python | **רשת-ביטחון לפאנל** (selective-prediction monitoring) — דוגם הלכות שאושרו ע"י הפאנל (`reviewer LIKE 'panel:%'`), מריץ עליהן **שוב** את הצבעת-ה-KEEP של 3 השופטים, ומציף כל מקרה שכעת נוטה DROP (false-keep פוטנציאלי). report-only כברירת-מחדל; `--flag` מחזיר את ה-flips ל-`pending_review` לסקירת-יו"ר. `--sample N`/`--seed`. בסיס 2026-06-07: 0/15. מיועד להרצה תקופתית (שבועי). מייבא שופטים מ-`halacha_panel_approve`. **חובה מקומי**. | תקופתי (שבועי) — ניטור |
-| `halacha_panel_calibrate.py` | python | **כיול מדיניות-ההצבעה של הפאנל** (Trust-or-Escalate, ICLR 2025). מריץ את שאלת-ה-KEEP של `halacha_panel_approve` על מדגם-הזהב ומודד מול `is_holding` (הציר-הגס) precision+coverage לכל מדיניות (unanimous/majority) + ספירת false-keep/false-drop. נותן את **אחוז-הטעות בפועל** לבחירת סף-סיכון α. מייבא שופטים מ-`halacha_panel_approve` (מקור-אמת יחיד). read-only, **חובה מקומי**. | ידני — לפני חיווט `--apply` |
+| `halacha_panel_calibrate.py` | python | **כיול + מדידת הפאנל** (Trust-or-Escalate, ICLR 2025). `--source live` (ברירת-מחדל): מריץ את שאלת-ה-KEEP על מדגם-הזהב ומודד מול `is_holding` precision+coverage+**split-rate** לכל מדיניות + false-keep/false-drop (מייבא שופטים מ-`halacha_panel_approve`, **חובה מקומי**). **#133/FU-5** — `--source captured`: **אפס-עלות** (בלי re-vote/LLM) — מצליב סבבים שמורים (FU-1) מול הכרעות-יו"ר (FU-2) דרך `db.panel_rounds_vs_chair` ומדווח split-rate+auto-precision **לכל סבב** (מגמת הלולאה: ככל שהרובריקה משתפרת precision נשמר ו-split יורד); משתף את `analyze_pairs` של FU-4 (מקור-יחיד). שתי המדידות מדווחות **anon-stability** (מבחן-אנונימיזציה #81.7) כמטריקת-בריאות נגד echo-chamber. `--batch`/`--limit`/`--concurrency`. | ידני — לפני חיווט `--apply` (live) / תקופתי — מעקב-לולאה (captured) |
 | `halacha_rubric_distill.py` | python | **#133/FU-4 — זיקוק-רובריקה PROPOSE-ONLY.** מצליב `halacha_panel_rounds` (FU-1, הצבעות+נימוקים) מול הכרעות-היו"ר (FU-2, seeds ב-`halacha_goldset` batch `chair-live`) דרך `db.panel_rounds_vs_chair` (read-only), מנתח דטרמיניסטית **כשלים שיטתיים** (false-keep/false-drop, פיצולים-שהוכרעו, שיעור-מחלוקת-עם-היו"ר לכל שופט), ומציע `KEEP_SYSTEM` v2 + exemplars מופשטים (claude_session מקומי, אפס עלות) כ**דוח-diff** ל-`data/learning/rubric-proposal-<ts>.md`. **לעולם לא auto-apply** — אימוץ v2 = עריכה אנושית של הקבוע דרך PR (INV-LRN1); exemplars מופשטים בלבד (INV-LRN5); הסיגנל היחיד = הכרעת-יו"ר, לא הצבעות-פאנל (anti-echo). מתחת ל-12 זוגות → "אין מספיק נתונים". `--no-llm` (סטטיסטיקה בלבד) / `--limit N`. **חובה מקומי**. | תקופתי — אחרי שהצטברו הכרעות-יו"ר על מחלוקות-פאנל |
 | `halacha_rule_role_backfill.py` | python | **INV-DM7** — backfill חד-פעמי: מסווג-מחדש את ההלכות הישנות (`rule_type IN ('binding','persuasive')` — ערכי-סמכות שנשמרו במסווה תפקיד לפני פיצול הצירים) לאחד מחמשת **תפקידי-הכלל** (holding/interpretive/procedural/application/obiter) דרך claude_session המקומי (אפס עלות). **לא נוגע בסמכות** (נגזרת מ-`precedent_level`). `--apply` (ברירת-מחדל dry-run) / `--limit N` / `--concurrency`. כותב backup CSV ל-`data/audit/` תחילה. fail-safe (פריט שנכשל → נשמר ערך ישן). **חובה מקומי** (claude_session). | ידני חד-פעמי אחרי deploy של פיצול-הסמכות |
 | `halacha_batch_reconcile.py` | python | **#82.7** — dedup חוצה-פסקים offline (שמרני, **dry-run בלבד**). dedup-on-insert משווה רק תוך-פסק; כאן סף מחמיר (cosine ≥0.95, `--cosine`) ולא-הרסני: מאתר זוגות הלכות near-duplicate בין פסקים שונים (pgvector `<=>` exact) עם איתות לקסיקלי (Jaccard/Levenshtein) ומדווח ל-CSV ב-`data/audit/` לסקירת היו"ר. לא מדלג/ממזג/מוחק. `--include-pending`. **`--link`** רושם את הזוגות שנמצאו כ-`equivalent_halachot` (parallel authority, #84.2 — קישור-מקביל ברמת-הלכה, **לא** ציטוט; idempotent, לא-הרסני). רץ עם venv של mcp-server. אומת: 800 הלכות → 5 זוגות (קושרו). | ידני — דוח-סקירה / `--link` לקישור |
diff --git a/scripts/halacha_panel_calibrate.py b/scripts/halacha_panel_calibrate.py
index 8fb2898..7aab175 100644
--- a/scripts/halacha_panel_calibrate.py
+++ b/scripts/halacha_panel_calibrate.py
@@ -17,8 +17,20 @@ precision (vs is_holding) and coverage (how many it decides vs escalates):
 Pick the policy whose auto-error stays under your tolerance while covering the
 most items. Read-only. Local-only (claude_session needs the CLI).
 
+FU-5 (#133) adds two things on top of the original live calibration:
+  - ``--source captured``: ZERO-COST measurement that reads the stored panel
+    rounds (FU-1) joined with the chair's rulings (FU-2) and reports, PER ROUND,
+    the split-rate + auto-precision the panel actually delivered against the
+    GROWING chair gold-set. This is how we see the active-learning loop working:
+    as the rubric improves (FU-4 → chair adopts v2), precision should hold while
+    the split-rate falls.
+  - anon-stability: the share of gold-set items whose verdict survives masking
+    the case names (#81.7 probe) — an echo-chamber health metric. A falling rate
+    means the panel is recognizing cases, not reasoning.
+
     cd ~/legal-ai/mcp-server
-    .venv/bin/python ../scripts/halacha_panel_calibrate.py
+    .venv/bin/python ../scripts/halacha_panel_calibrate.py                 # live calibration
+    .venv/bin/python ../scripts/halacha_panel_calibrate.py --source captured  # per-round trend (free)
 """
 from __future__ import annotations
 
@@ -32,6 +44,63 @@ from legal_mcp.services import db
 from halacha_panel_approve import (  # noqa: E402
     KEEP_SYSTEM, _bool, _keep_user, judge_claude, judge_deepseek, judge_gemini,
 )
+# reuse FU-4's stats core so "what failed" is computed in ONE place (no drift)
+from halacha_rubric_distill import _JUDGES, analyze_pairs  # noqa: E402
+
+
+def summarize_calibration(pairs: list[dict]) -> dict:
+    """Pure: from captured (panel ⋈ chair) pairs, the calibration health metrics
+    the panel ACTUALLY delivered against the chair's ground-truth (#133/FU-5).
+
+    split_rate — fraction the panel escalated to the chair (split/incomplete);
+      the cost of caution. auto_precision — of the auto-decisions, the fraction
+      the chair later agreed with. As the rubric improves (FU-4 → chair adopts),
+      precision should hold while split_rate falls. Reuses FU-4's analyze_pairs
+      so the failure buckets are defined once."""
+    labeled = [p for p in pairs if p.get("chair_keep") is not None]
+    a = analyze_pairs(labeled)
+    n = a["n_pairs"]
+    escalated = a["n_splits_resolved"]
+    auto = n - escalated
+    wrong = a["n_false_keep"] + a["n_false_drop"]
+    return {
+        "n": n,
+        "auto_decided": auto,
+        "escalated": escalated,
+        "split_rate": round(escalated / n, 3) if n else None,
+        "auto_precision": round((auto - wrong) / auto, 3) if auto else None,
+        "false_keep": a["n_false_keep"],
+        "false_drop": a["n_false_drop"],
+        "judge_disagree": {j: a["judge_stats"][j]["disagree_rate"] for j in _JUDGES},
+    }
+
+
+def bucket_by_round(pairs: list[dict]) -> list[tuple[str, dict]]:
+    """Pure: split the pairs by the panel round's CALENDAR DAY and summarize each
+    bucket — the per-round trend on the GROWING chair gold-set. round_ts is the
+    captured round's ISO timestamp; the day prefix groups a run together."""
+    buckets: dict[str, list[dict]] = {}
+    for p in pairs:
+        day = (p.get("round_ts") or "")[:10] or "unknown"
+        buckets.setdefault(day, []).append(p)
+    return [(day, summarize_calibration(buckets[day])) for day in sorted(buckets)]
+
+
+async def _anon_stability(batch: str) -> dict:
+    """Echo-chamber health metric (#133/FU-5). The anon probe (#81.7) re-judges
+    gold-set items with case names masked; ``anon_stable`` is True when the
+    masked verdict still matches the consensus. A FALLING stable-rate means the
+    panel is recognizing cases rather than reasoning — the echo-chamber symptom
+    the loop must watch. Reads stored columns (populated by goldset_panel_label
+    for the 'default' batch); returns counts so callers can report or skip."""
+    items = await db.goldset_list(batch)
+    probed = [it for it in items if it.get("anon_stable") is not None]
+    stable = sum(1 for it in probed if it["anon_stable"])
+    return {
+        "probed": len(probed),
+        "stable": stable,
+        "stable_rate": round(stable / len(probed), 3) if probed else None,
+    }
 
 
 async def _votes(client, h) -> list[bool]:
@@ -63,7 +132,49 @@ def _decide(votes: list[bool], policy: str) -> bool | None:
     return None  # tie
 
 
+def _fmt(x) -> str:
+    return "—" if x is None else (f"{x:.1%}" if isinstance(x, float) else str(x))
+
+
+async def _run_captured(args: argparse.Namespace) -> int:
+    """FU-5 — measure the panel's REAL performance against the growing chair
+    gold-set, from CAPTURED rounds (FU-1) ⋈ chair rulings (FU-2). Zero-cost: no
+    re-voting, no LLM — it reports what the panel actually delivered, per round.
+    """
+    pairs = await db.panel_rounds_vs_chair(limit=args.limit or 5000)
+    overall = summarize_calibration(pairs)
+    print(f"captured calibration on {overall['n']} (panel ⋈ chair) pairs\n", flush=True)
+    if not overall["n"]:
+        print("no chair-resolved pairs yet — seeds accrue as the chair reviews "
+              "panel-judged halachot (FU-2). Nothing to measure.", flush=True)
+        return 0
+
+    print("=" * 70)
+    print(f"{'round-day':<12}{'pairs':>7}{'auto':>6}{'split%':>9}{'precision':>11}"
+          f"{'fKEEP':>7}{'fDROP':>7}")
+    print("-" * 70)
+    for day, s in bucket_by_round(pairs):
+        print(f"{day:<12}{s['n']:>7}{s['auto_decided']:>6}{_fmt(s['split_rate']):>9}"
+              f"{_fmt(s['auto_precision']):>11}{s['false_keep']:>7}{s['false_drop']:>7}")
+    print("-" * 70)
+    print(f"{'OVERALL':<12}{overall['n']:>7}{overall['auto_decided']:>6}"
+          f"{_fmt(overall['split_rate']):>9}{_fmt(overall['auto_precision']):>11}"
+          f"{overall['false_keep']:>7}{overall['false_drop']:>7}")
+    print("\nper-judge disagreement with the chair (lower = better aligned):")
+    for j in _JUDGES:
+        print(f"  {j:<10} {_fmt(overall['judge_disagree'][j])}")
+
+    anon = await _anon_stability("default")
+    print(f"\nanon-stability (echo-chamber health, batch 'default'): "
+          f"{_fmt(anon['stable_rate'])} over {anon['probed']} probed "
+          f"({'falling = memorization risk' if anon['probed'] else 'not populated — run goldset_panel_label'})")
+    return 0
+
+
 async def main(args: argparse.Namespace) -> int:
+    if args.source == "captured":
+        return await _run_captured(args)
+
     items = [it for it in await db.goldset_list(args.batch) if it.get("is_holding") is not None]
     if args.limit:
         items = items[: args.limit]
@@ -82,8 +193,8 @@ async def main(args: argparse.Namespace) -> int:
             print(f"  …{len(rows)}/{len(items)}", flush=True)
 
     print("\n" + "=" * 64)
-    print(f"{'policy':<11}{'auto':>6}{'escalate':>10}{'correct':>9}{'wrong':>7}{'precision':>11}{'coverage':>10}")
-    print("-" * 64)
+    print(f"{'policy':<11}{'auto':>6}{'escalate':>10}{'correct':>9}{'wrong':>7}{'precision':>11}{'coverage':>10}{'split':>10}")
+    print("-" * 74)
     for policy in ("unanimous", "majority"):
         auto = wrong = correct = 0
         for r in rows:
@@ -98,7 +209,8 @@ async def main(args: argparse.Namespace) -> int:
         esc = len(rows) - auto
         prec = correct / auto if auto else 0.0
         cov = auto / len(rows) if rows else 0.0
-        print(f"{policy:<11}{auto:>6}{esc:>10}{correct:>9}{wrong:>7}{prec:>10.1%}{cov:>10.1%}")
+        split = esc / len(rows) if rows else 0.0  # FU-5: escalation = split rate
+        print(f"{policy:<11}{auto:>6}{esc:>10}{correct:>9}{wrong:>7}{prec:>10.1%}{cov:>10.1%}{split:>10.1%}")
 
     # where do the WRONG auto-decisions fall? (false-keep is the costly one)
     print("\n=== costly errors: panel auto-KEEPS but human says NOT-holding (per policy) ===")
@@ -106,12 +218,21 @@ async def main(args: argparse.Namespace) -> int:
         fk = sum(1 for r in rows if _decide(r["votes"], policy) is True and not r["truth"])
         fd = sum(1 for r in rows if _decide(r["votes"], policy) is False and r["truth"])
         print(f"  {policy:<11} false-KEEP (bad rule approved): {fk}   false-DROP (good rule rejected): {fd}")
+
+    # FU-5: echo-chamber health — does masking case names flip the verdict?
+    anon = await _anon_stability(args.batch)
+    print(f"\nanon-stability (echo-chamber health): {_fmt(anon['stable_rate'])} "
+          f"over {anon['probed']} probed"
+          + ("" if anon["probed"] else " — not populated; run goldset_panel_label"))
     return 0
 
 
 if __name__ == "__main__":
-    ap = argparse.ArgumentParser()
-    ap.add_argument("--batch", default="default")
+    ap = argparse.ArgumentParser(description="Calibrate / measure the halacha panel (Trust-or-Escalate; FU-5).")
+    ap.add_argument("--source", choices=("live", "captured"), default="live",
+                    help="live: re-vote the gold-set now (needs CLI+keys). "
+                         "captured: zero-cost — measure stored rounds (FU-1) vs chair rulings (FU-2), per round.")
+    ap.add_argument("--batch", default="default", help="gold-set batch for live mode + anon-stability")
     ap.add_argument("--limit", type=int, default=0)
     ap.add_argument("--concurrency", type=int, default=6)
     raise SystemExit(asyncio.run(main(ap.parse_args())))