diff --git a/docs/research/artifacts/rectification-width-funnel-2026-10-16/dayonly.py b/docs/research/artifacts/rectification-width-funnel-2026-10-16/dayonly.py
new file mode 100644
index 00000000..151d4109
--- /dev/null
+++ b/docs/research/artifacts/rectification-width-funnel-2026-10-16/dayonly.py
@@ -0,0 +1,32 @@
+import json, sys, random
+from pathlib import Path
+sys.path.insert(0, str(Path(__file__).parent))
+import karaka_interval as K
+asc = json.load(open('asc_signs.json'))
+cases = json.loads((K.WT/'references/real_case_calibration/minute_rectification_holdout_v5.json').read_text())['cases']
+shard, n = map(int, sys.argv[1].split('/'))
+out = []
+def seg(cid, R, sel):
+    win = {int(k): v for k, v in asc[cid].items() if abs(int(k)) <= R*60}
+    if len(set(win.values())) < 2: return None
+    t = win[0]; p = {win[g] for g in sel}
+    return 'right' if p == {t} else 'tie' if t in p else 'wrong'
+for i, c in enumerate(cases):
+    if i % n != shard: continue
+    evs = [e for e in c['events'] if e['precision'] == 'day']
+    if len(evs) < 3: continue
+    for R in (10, 30, 60, 75):
+        for lv in (3, 4):
+            for km in ('P', 'PS'):
+                r = K.run_case(c, R, lv, km, [dict(e) for e in evs])
+                rng = random.Random(f"day|{c['case_id']}|{R}|{lv}|{km}")
+                size = 1 if km == 'P' else 2
+                pl = []
+                for _ in range(30):
+                    q = K.run_case(c, R, lv, km, [{**e, 'karaka_set': set(rng.sample(K.PLANETS, size))} for e in evs])
+                    pl.append((q['covered_tol'], q['width'], seg(c['case_id'], R, q['sel'])))
+                out.append(dict(case=c['case_id'], n_ev=len(evs), R=R, lv=lv, km=km, cov=r['covered'], cov_tol=r['covered_tol'],
+                    width=r['width'], window=r['window'], all_expl=r['all_explained'], seg=seg(c['case_id'], R, r['sel']),
+                    pl_cov_tol=sum(p[0] for p in pl)/30, pl_width=sum(p[1] for p in pl)/30,
+                    pl_seg={k: sum(p[2] == k for p in pl)/30 for k in ('right', 'tie', 'wrong')}))
+Path(f'day_shard{shard}.json').write_text(json.dumps(out)); print('DONE')
diff --git a/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel.py b/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel.py
new file mode 100644
index 00000000..192e3447
--- /dev/null
+++ b/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel.py
@@ -0,0 +1,67 @@
+"""Why does the delivered interval stay wide? Per-question funnel on v5 (research harness, static probe list)."""
+import json, sys, statistics as st
+from pathlib import Path
+WT = Path('/workspace/Jyotisha/.worktrees/karaka-interval-20261016')
+sys.path.insert(0, str(WT)); sys.path.insert(0, str(WT/'scripts'))
+from scripts.research import cluster_width_probe as cwp
+from scripts.research import probe_supply_after_six as ps
+from scripts.research import cluster_width_lib as cl
+cases = json.loads((WT/'references/real_case_calibration/minute_rectification_holdout_v5.json').read_text())['cases']
+shard, n = map(int, sys.argv[1].split('/'))
+MAXQ = 12
+out = []
+def clock(t): h, m = map(int, t.split(':')); return h*60+m
+for i, case in enumerate(cases):
+    if i % n != shard: continue
+    true_time = str(case['birth']['time'])[:5]
+    for R in (10, 30, 60):
+        req = cwp.scoring_request_for(case, R)
+        built = cwp.build_event_contribution_matrix(req)
+        rows = cwp.score_from_matrix(req, built)
+        contexts = list(built.get('static_contexts') or [])
+        window_times = [str(r['time'])[:5] for r in rows]
+        raw = cwp.raw_signature_clusters(contexts)
+        probes = cwp.discriminating_event_probes({**req, 'refresh_probes': False}, built, scan=cwp.window_scan(built),
+            candidate_times=list(window_times), representative_time=true_time, today=cwp.TODAY)
+        merged, _ = cwp.merge_adjacent_traced(raw, cwp._by_time_rows(rows), max_clusters=cwp.Variant('baseline').max_clusters,
+                                              max_peak_delta=cwp.Variant('baseline').max_peak_delta)
+        public = cwp.public_from_clusters(merged, rows)
+        reps = [str(r['time'])[:5] for r in public]
+        size = {str(r['time'])[:5]: len(r.get('cluster_times') or [1]) for r in public}
+        truth_rep = next((str(r['time'])[:5] for r in public if true_time in (r.get('cluster_times') or [str(r['time'])[:5]])), None)
+        prior = {t: float(r.get('score') or 0) for t, r in zip(reps, public)}
+        scores = dict(prior); conflicts = {t: 0 for t in reps}; elim = set()
+        prior_spread = max(prior.values()) - min(prior.values())
+        prior_truth_gap = max(prior.values()) - prior.get(truth_rep, 0)
+        steps = []
+        for q, probe in enumerate(probes[:MAXQ]):
+            alive = [t for t in reps if t not in elim]
+            yes, no = ps.outcome_groups(probe)
+            ans = ps.optimal_answer(probe, true_time)
+            sides = {}
+            for t in alive:
+                d = 'support' if t in yes else 'conflict' if t in no else ps.inherit_direction(t, yes, no)
+                sides[t] = d
+            truth_d = 'support' if ans == 'yes' else 'conflict' if ans == 'no' else 'neutral'
+            same = sum(size[t] for t in alive if sides[t] == truth_d) ; opp = sum(size[t] for t in alive if sides[t] not in (truth_d, 'neutral'))
+            neu = sum(size[t] for t in alive if sides[t] == 'neutral')
+            if ans is not None:
+                scores, conflicts, elim = ps.apply_answer(scores, conflicts, elim, probe, ans, reps)
+            valid = cl.still_valid_public(public, scores, elim, lead=8)
+            d = cl.delivery_from_public(valid)
+            steps.append(dict(q=q+1, answer=ans, same=same, opp=opp, neutral=neu, alive_min=sum(size[t] for t in reps if t not in elim),
+                              span8=d['width'], set8=len(d['times']), domain=probe.get('domain') or probe.get('kind')))
+        def deliver(lead, upto):
+            sc = dict(prior); cf = {t: 0 for t in reps}; el = set()
+            for probe in probes[:upto]:
+                a = ps.optimal_answer(probe, true_time)
+                if a: sc, cf, el = ps.apply_answer(sc, cf, el, probe, a, reps)
+            v = cl.still_valid_public(public, sc, el, lead=lead); d = cl.delivery_from_public(v)
+            alive = [t for t in reps if t not in el]
+            best = max(sc[t] for t in alive); leaders = [t for t in alive if sc[t] == best]
+            return dict(span=d['width'], set=len(d['times']), cov=true_time in d['times'], top1=leaders == [truth_rep])
+        out.append(dict(case=case['case_id'], R=R, true=true_time, n_reps=len(reps), n_probes=len(probes),
+             prior_spread=prior_spread, prior_truth_gap=prior_truth_gap, steps=steps,
+             alt={f'L{lead}_Q{q}': deliver(lead, q) for lead in (8, 4, 2) for q in (6, 12)}))
+    print(case['case_id'], flush=True)
+Path(f'funnel_shard{shard}.json').write_text(json.dumps(out)); print('DONE')
diff --git a/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel2.py b/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel2.py
new file mode 100644
index 00000000..47e07d01
--- /dev/null
+++ b/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel2.py
@@ -0,0 +1,68 @@
+"""Why does the delivered interval stay wide? Per-question funnel on v5 (research harness, static probe list)."""
+import json, sys, statistics as st
+from pathlib import Path
+WT = Path('/workspace/Jyotisha/.worktrees/karaka-interval-20261016')
+sys.path.insert(0, str(WT)); sys.path.insert(0, str(WT/'scripts'))
+from scripts.research import cluster_width_probe as cwp
+from scripts.research import probe_supply_after_six as ps
+from scripts.research import cluster_width_lib as cl
+cases = json.loads((WT/'references/real_case_calibration/minute_rectification_holdout_v5.json').read_text())['cases']
+shard, n = map(int, sys.argv[1].split('/'))
+MAXQ = 12
+out = []
+def clock(t): h, m = map(int, t.split(':')); return h*60+m
+for i, case in enumerate(cases):
+    if i % n != shard: continue
+    true_time = str(case['birth']['time'])[:5]
+    for R in (10, 30, 60):
+        req = cwp.scoring_request_for(case, R)
+        built = cwp.build_event_contribution_matrix(req)
+        rows = cwp.score_from_matrix(req, built)
+        contexts = list(built.get('static_contexts') or [])
+        window_times = [str(r['time'])[:5] for r in rows]
+        raw = cwp.raw_signature_clusters(contexts)
+        probes = cwp.discriminating_event_probes({**req, 'refresh_probes': False}, built, scan=cwp.window_scan(built),
+            candidate_times=list(window_times), representative_time=true_time, today=cwp.TODAY)
+        merged, _ = cwp.merge_adjacent_traced(raw, cwp._by_time_rows(rows), max_clusters=cwp.Variant('baseline').max_clusters,
+                                              max_peak_delta=cwp.Variant('baseline').max_peak_delta)
+        public = cwp.public_from_clusters(merged, rows)
+        reps = [str(r['time'])[:5] for r in public]
+        size = {str(r['time'])[:5]: len(r.get('cluster_times') or [1]) for r in public}
+        truth_rep = next((str(r['time'])[:5] for r in public if true_time in (r.get('cluster_times') or [str(r['time'])[:5]])), None)
+        prior = {t: float(r.get('score') or 0) for t, r in zip(reps, public)}
+        scores = dict(prior); conflicts = {t: 0 for t in reps}; elim = set()
+        prior_spread = max(prior.values()) - min(prior.values())
+        prior_truth_gap = max(prior.values()) - prior.get(truth_rep, 0)
+        steps = []
+        for q, probe in enumerate(probes[:MAXQ]):
+            alive = [t for t in reps if t not in elim]
+            yes, no = ps.outcome_groups(probe)
+            ans = ps.optimal_answer(probe, true_time)
+            sides = {}
+            for t in alive:
+                d = 'support' if t in yes else 'conflict' if t in no else ps.inherit_direction(t, yes, no)
+                sides[t] = d
+            truth_d = 'support' if ans == 'yes' else 'conflict' if ans == 'no' else 'neutral'
+            same = sum(size[t] for t in alive if sides[t] == truth_d) ; opp = sum(size[t] for t in alive if sides[t] not in (truth_d, 'neutral'))
+            neu = sum(size[t] for t in alive if sides[t] == 'neutral')
+            if ans is not None:
+                scores, conflicts, elim = ps.apply_answer(scores, conflicts, elim, probe, ans, reps)
+            valid = cl.still_valid_public(public, scores, elim, lead=8)
+            d = cl.delivery_from_public(valid)
+            steps.append(dict(q=q+1, answer=ans, same=same, opp=opp, neutral=neu, alive_min=sum(size[t] for t in reps if t not in elim),
+                              span8=d['width'], set8=len(d['times']), domain=probe.get('domain') or probe.get('kind')))
+        def deliver(lead, upto, flip=None):
+            sc = dict(prior); cf = {t: 0 for t in reps}; el = set()
+            for k, probe in enumerate(probes[:upto]):
+                a = ps.optimal_answer(probe, true_time)
+                if a and flip is not None and k == flip: a = 'no' if a == 'yes' else 'yes'
+                if a: sc, cf, el = ps.apply_answer(sc, cf, el, probe, a, reps)
+            v = cl.still_valid_public(public, sc, el, lead=lead); d = cl.delivery_from_public(v)
+            alive = [t for t in reps if t not in el]
+            best = max(sc[t] for t in alive); leaders = [t for t in alive if sc[t] == best]
+            return dict(span=d['width'], set=len(d['times']), cov=true_time in d['times'], top1=leaders == [truth_rep])
+        out.append(dict(case=case['case_id'], R=R, true=true_time, n_reps=len(reps), n_probes=len(probes),
+             prior_spread=prior_spread, prior_truth_gap=prior_truth_gap, steps=steps,
+             alt={f'L{lead}_Q{q}_F{f}': deliver(lead, q, None if f < 0 else f) for lead in (8, 6, 5, 4) for q in (6, 12) for f in (-1, 0, 2, 4)}))
+    print(case['case_id'], flush=True)
+Path(f'funnel2_shard{shard}.json').write_text(json.dumps(out)); print('DONE')
diff --git a/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel_summary.json b/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel_summary.json
new file mode 100644
index 00000000..584c3797
--- /dev/null
+++ b/docs/research/artifacts/rectification-width-funnel-2026-10-16/funnel_summary.json
@@ -0,0 +1,199 @@
+{
+ "dataset": "minute_rectification_holdout_v5.json",
+ "staging_head": "9b7509155",
+ "harness": "cluster_width_probe static probe list, simulated correct answers; one-wrong = flip answer #1/#3/#5",
+ "cells": {
+  "pm10_lead8_q6": {
+   "n": 77,
+   "live_minutes_median": 14,
+   "span_median": 13,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.983,
+   "top1": 51
+  },
+  "pm10_lead8_q12": {
+   "n": 77,
+   "live_minutes_median": 12,
+   "span_median": 11,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.987,
+   "top1": 53
+  },
+  "pm10_lead6_q6": {
+   "n": 77,
+   "live_minutes_median": 10,
+   "span_median": 11,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.961,
+   "top1": 51
+  },
+  "pm10_lead6_q12": {
+   "n": 77,
+   "live_minutes_median": 10,
+   "span_median": 9,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.97,
+   "top1": 53
+  },
+  "pm10_lead5_q6": {
+   "n": 77,
+   "live_minutes_median": 10,
+   "span_median": 9,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.926,
+   "top1": 51
+  },
+  "pm10_lead5_q12": {
+   "n": 77,
+   "live_minutes_median": 10,
+   "span_median": 9,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.931,
+   "top1": 53
+  },
+  "pm10_lead4_q6": {
+   "n": 77,
+   "live_minutes_median": 8,
+   "span_median": 9,
+   "truth_in_all_correct": 74,
+   "truth_in_one_wrong_rate": 0.866,
+   "top1": 51
+  },
+  "pm10_lead4_q12": {
+   "n": 77,
+   "live_minutes_median": 8,
+   "span_median": 7,
+   "truth_in_all_correct": 74,
+   "truth_in_one_wrong_rate": 0.87,
+   "top1": 53
+  },
+  "pm30_lead8_q6": {
+   "n": 77,
+   "live_minutes_median": 26,
+   "span_median": 35,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.952,
+   "top1": 41
+  },
+  "pm30_lead8_q12": {
+   "n": 77,
+   "live_minutes_median": 22,
+   "span_median": 33,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.961,
+   "top1": 46
+  },
+  "pm30_lead6_q6": {
+   "n": 77,
+   "live_minutes_median": 20,
+   "span_median": 33,
+   "truth_in_all_correct": 75,
+   "truth_in_one_wrong_rate": 0.892,
+   "top1": 41
+  },
+  "pm30_lead6_q12": {
+   "n": 77,
+   "live_minutes_median": 18,
+   "span_median": 25,
+   "truth_in_all_correct": 76,
+   "truth_in_one_wrong_rate": 0.926,
+   "top1": 46
+  },
+  "pm30_lead5_q6": {
+   "n": 77,
+   "live_minutes_median": 16,
+   "span_median": 31,
+   "truth_in_all_correct": 73,
+   "truth_in_one_wrong_rate": 0.823,
+   "top1": 41
+  },
+  "pm30_lead5_q12": {
+   "n": 77,
+   "live_minutes_median": 14,
+   "span_median": 15,
+   "truth_in_all_correct": 75,
+   "truth_in_one_wrong_rate": 0.883,
+   "top1": 46
+  },
+  "pm30_lead4_q6": {
+   "n": 77,
+   "live_minutes_median": 12,
+   "span_median": 19,
+   "truth_in_all_correct": 72,
+   "truth_in_one_wrong_rate": 0.732,
+   "top1": 41
+  },
+  "pm30_lead4_q12": {
+   "n": 77,
+   "live_minutes_median": 12,
+   "span_median": 13,
+   "truth_in_all_correct": 75,
+   "truth_in_one_wrong_rate": 0.818,
+   "top1": 46
+  },
+  "pm60_lead8_q6": {
+   "n": 77,
+   "live_minutes_median": 42,
+   "span_median": 71,
+   "truth_in_all_correct": 75,
+   "truth_in_one_wrong_rate": 0.931,
+   "top1": 30
+  },
+  "pm60_lead8_q12": {
+   "n": 77,
+   "live_minutes_median": 38,
+   "span_median": 59,
+   "truth_in_all_correct": 75,
+   "truth_in_one_wrong_rate": 0.952,
+   "top1": 36
+  },
+  "pm60_lead6_q6": {
+   "n": 77,
+   "live_minutes_median": 32,
+   "span_median": 59,
+   "truth_in_all_correct": 75,
+   "truth_in_one_wrong_rate": 0.857,
+   "top1": 30
+  },
+  "pm60_lead6_q12": {
+   "n": 77,
+   "live_minutes_median": 28,
+   "span_median": 51,
+   "truth_in_all_correct": 75,
+   "truth_in_one_wrong_rate": 0.883,
+   "top1": 36
+  },
+  "pm60_lead5_q6": {
+   "n": 77,
+   "live_minutes_median": 28,
+   "span_median": 53,
+   "truth_in_all_correct": 73,
+   "truth_in_one_wrong_rate": 0.74,
+   "top1": 30
+  },
+  "pm60_lead5_q12": {
+   "n": 77,
+   "live_minutes_median": 22,
+   "span_median": 39,
+   "truth_in_all_correct": 74,
+   "truth_in_one_wrong_rate": 0.814,
+   "top1": 36
+  },
+  "pm60_lead4_q6": {
+   "n": 77,
+   "live_minutes_median": 22,
+   "span_median": 45,
+   "truth_in_all_correct": 68,
+   "truth_in_one_wrong_rate": 0.606,
+   "top1": 30
+  },
+  "pm60_lead4_q12": {
+   "n": 77,
+   "live_minutes_median": 16,
+   "span_median": 27,
+   "truth_in_all_correct": 70,
+   "truth_in_one_wrong_rate": 0.701,
+   "top1": 36
+  }
+ }
+}
\ No newline at end of file
diff --git a/docs/research/artifacts/rectification-width-funnel-2026-10-16/karaka_interval.py b/docs/research/artifacts/rectification-width-funnel-2026-10-16/karaka_interval.py
new file mode 100644
index 00000000..b365e737
--- /dev/null
+++ b/docs/research/artifacts/rectification-width-funnel-2026-10-16/karaka_interval.py
@@ -0,0 +1,109 @@
+"""Author's method on v5: natural-karaka sub-period interval intersection.
+
+For each candidate birth moment (15 s grid inside truth±R), an event is 'explained' when a
+Vimshottari period at level L ruled by the event's karaka overlaps the event's date span.
+Delivered set = grid points explaining the maximum number of events (author tolerates
+unexplained events). Truth is used only for the window centre (same as production replay)
+and for scoring afterwards.
+"""
+import json, sys, random, calendar, statistics
+from datetime import datetime, timedelta
+from pathlib import Path
+HERE = Path(__file__).parent
+WT = Path('/workspace/Jyotisha/.worktrees/karaka-interval-20261016')
+sys.path.insert(0, str(WT)); sys.path.insert(0, str(WT/'scripts')); sys.path.insert(0, str(HERE))
+import dasha_analyzer
+from scripts.research.nadi_seconds_lib import sky_at, _timeline
+from karaka_rules import classify
+PLANETS = list(dasha_analyzer.DASHA_ORDER)
+STEP_S = 15
+
+def event_span(e):
+    if e['precision'] == 'day':
+        a = datetime.strptime(e['date'], '%Y-%m-%d'); return a, a + timedelta(days=1)
+    y, m = map(int, e['date'].split('-')[:2])
+    a = datetime(y, m, 1); return a, a + timedelta(days=calendar.monthrange(y, m)[1])
+
+def lords_overlapping(timeline, a, b, level):
+    """Set of lords at `level` (1=MD..4=SD) whose period overlaps [a,b)."""
+    periods = [p for p in timeline if p['start'] < b and p['end'] > a]
+    for _ in range(level - 1):
+        nxt = []
+        for p in periods:
+            nxt += [q for q in dasha_analyzer.build_antardasha(dict(p)) if q['start'] < b and q['end'] > a]
+        periods = nxt
+    return {p['lord'] for p in periods}
+
+def moon_track(birth, radius):
+    base = datetime.strptime(f"{birth['date']} {birth['time'][:5]}", '%Y-%m-%d %H:%M')
+    knots = list(range(-radius - 10, radius + 11, 10))
+    vals = []
+    for k in knots:
+        vals.append(sky_at(birth, base + timedelta(minutes=k)))
+    def at(sec):
+        m = sec / 60.0
+        i = min(max(int((m - knots[0]) // 10), 0), len(knots) - 2)
+        lo, hi = vals[i]['moon_lon'], vals[i + 1]['moon_lon']
+        if hi < lo - 180: hi += 360
+        f = (m - knots[i]) / 10.0
+        return (lo + (hi - lo) * f) % 360.0, vals[i]['asc_index'] if f < .5 else vals[i+1]['asc_index']
+    return base, at
+
+def run_case(case, radius, level, karaka_mode, events_override=None):
+    birth = case['birth']
+    evs = events_override if events_override is not None else [e for e in case['events'] if e['precision'] in ('day', 'month')]
+    if not evs: return None
+    base, at = moon_track(birth, radius)
+    spans = [event_span(e) for e in evs]
+    ks = []
+    for e in evs:
+        if 'karaka_set' in e: ks.append(e['karaka_set']); continue
+        _, p, s = classify(e['description'])
+        ks.append({p} if karaka_mode == 'P' else {p, s})
+    grid = list(range(-radius * 60, radius * 60 + 1, STEP_S))
+    counts = []
+    for sec in grid:
+        moon, _ = at(sec)
+        bd = (base + timedelta(seconds=sec)).date().isoformat()
+        tl = _timeline(bd, moon)
+        counts.append(sum(bool(lords_overlapping(tl, a, b, level) & k) for (a, b), k in zip(spans, ks)))
+    best = max(counts)
+    sel = [g for g, c in zip(grid, counts) if c == best]
+    runs = 1 + sum(1 for x, y in zip(sel, sel[1:]) if y - x > STEP_S)
+    width_min = len(sel) * STEP_S / 60.0
+    return {'n_events': len(evs), 'best': best, 'all_explained': best == len(evs),
+            'covered': 0 in sel, 'width': width_min, 'window': (2 * radius * 60 + STEP_S) / 60.0,
+            'runs': runs, 'span': (min(sel) / 60, max(sel) / 60),
+            'center_err': abs((min(sel) + max(sel)) / 2 / 60.0), 'counts_truth': counts[grid.index(0)],
+            'covered_tol': any(abs(g) <= 150 for g in sel), 'sel': sel}
+
+if __name__ == '__main__':
+    cases = json.loads((WT/'references/real_case_calibration/minute_rectification_holdout_v5.json').read_text())['cases']
+    radii = [int(x) for x in sys.argv[1].split(',')]
+    shard, nshard = map(int, sys.argv[3].split('/'))
+    cases = [c for i, c in enumerate(cases) if i % nshard == shard]
+    configs = [(lv, km) for lv in (2, 3, 4) for km in ('P', 'PS')]
+    n_placebo = int(sys.argv[2]) if len(sys.argv) > 2 else 20
+    out = []
+    for case in cases:
+        for R in radii:
+            for lv, km in configs:
+                r = run_case(case, R, lv, km)
+                if r is None: continue
+                # placebo: random karaka per event, same set size
+                pc = []
+                rng = random.Random(f"{case['case_id']}|{R}|{lv}|{km}")
+                evs = [e for e in case['events'] if e['precision'] in ('day', 'month')]
+                size = 1 if km == 'P' else 2
+                for _ in range(n_placebo):
+                    ov = [{**e, 'karaka_set': set(rng.sample(PLANETS, size))} for e in evs]
+                    q = run_case(case, R, lv, km, ov)
+                    pc.append({'covered': q['covered'], 'covered_tol': q['covered_tol'], 'width': q['width']})
+                r.update(case_id=case['case_id'], radius=R, level=lv, karaka=km,
+                         placebo_cov=sum(p['covered'] for p in pc) / len(pc),
+                         placebo_width=statistics.mean(p['width'] for p in pc),
+                         placebo_cov_tol=sum(p['covered_tol'] for p in pc) / len(pc))
+                out.append(r)
+        print(case['case_id'], flush=True)
+        (HERE / f"karaka_shard{shard}.json").write_text(json.dumps(out, indent=0, default=list))
+    print('DONE')
diff --git a/docs/research/artifacts/rectification-width-funnel-2026-10-16/karaka_rules.py b/docs/research/artifacts/rectification-width-funnel-2026-10-16/karaka_rules.py
new file mode 100644
index 00000000..1de1f990
--- /dev/null
+++ b/docs/research/artifacts/rectification-width-funnel-2026-10-16/karaka_rules.py
@@ -0,0 +1,36 @@
+"""Pre-registered natural-karaka table (frozen before any run against truth).
+
+Event -> one primary karaka (author's method: Mercury=exam/maths, Mars=surgery, Ketu=separation).
+Rules are ordered; first match wins. Classical natural significations (BPHS / common usage).
+"""
+import re
+RULES = [
+    # (tag, regex, primary karaka, secondary karaka)
+    ("native_death", r"^(died|death|was killed|killed)\b|assassinat", "Saturn", "Mars"),
+    ("father_event", r"\bfather\b", "Sun", "Saturn"),
+    ("mother_event", r"\bmother\b", "Moon", "Saturn"),
+    ("sibling_event", r"\b(brother|sister|sibling)", "Mars", "Saturn"),
+    ("spouse_death", r"\b(husband|wife)\b.*\b(died|death|killed)", "Venus", "Saturn"),
+    ("child_birth", r"\b(son|daughter|child|children|twins|baby)\b.*\b(born|birth)|\bbirth of\b|\bgave birth", "Jupiter", "Moon"),
+    ("separation", r"divorc|separat|break ?up|split|annul|estrang|ended (his|her|their) (relationship|marriage)", "Ketu", "Saturn"),
+    ("marriage", r"marri|wed|engag|relationship|romance|dating|partner", "Venus", "Jupiter"),
+    ("finance", r"bankrupt|debt|default|loan|bought|purchas|sold|earn|fortune|forbes|million|deal|invest|franchise|wealth|donat|prize money|inherit", "Jupiter", "Rahu"),
+    ("award", r"award|prize|oscar|grammy|emmy|medal|honou?r|nominat|nobel|\bwon\b|\bwins?\b|champion|title", "Sun", "Jupiter"),
+    ("authority", r"elect|president|prime minister|minister|appoint|named|crown|king|queen|governor|senat|deput|chair|director of|head of|knight|ambassador", "Sun", "Saturn"),
+    ("surgery_injury", r"surg|injur|accident|crash|broke|fractur|wound|burn|operation|amputat|shot\b|stab", "Mars", "Saturn"),
+    ("illness", r"\bill(ness)?\b|sick|hospital|cancer|stroke|disease|diagnos|tumou?r|polio|addict|overdose|heart|depress|breakdown|tubercul|infect|rehab", "Saturn", "Rahu"),
+    ("legal_confinement", r"arrest|prison|jail|convict|sentenc|exile|lawsuit|trial|court|indict|charged|impeach|deport|coup|ousted|resign", "Saturn", "Rahu"),
+    ("higher_degree", r"doctor of|doctorate|phd|ph\.d|juris|law degree|m\.b\.|master'?s degree|postgraduate|dissertation", "Jupiter", "Mercury"),
+    ("education", r"graduat|school|universit|college|academ|degree|stud(y|ied)|exam|enrol|attend|educat|conservator|institut|lycée|seminar", "Mercury", "Jupiter"),
+    ("relocation", r"moved|emigrat|immigrat|relocat|settled|returned|migrat|fled|left for|sent to", "Rahu", "Moon"),
+    ("writing", r"publish|wrote|writ|book|novel|essay|poem|journal", "Mercury", "Jupiter"),
+    ("arts", r"film|movie|starr|act(ed|ing|ress)|debut|premier|perform|sing|song|album|record|concert|tour|model|show|music|paint|exhibit|danc|theat", "Venus", "Mercury"),
+    ("career_default", r".", "Saturn", "Sun"),
+]
+COMPILED = [(t, re.compile(rx, re.I), p, s) for t, rx, p, s in RULES]
+
+def classify(description: str):
+    for tag, rx, primary, secondary in COMPILED:
+        if rx.search(description or ""):
+            return tag, primary, secondary
+    raise AssertionError
diff --git a/docs/research/artifacts/rectification-width-funnel-2026-10-16/prod_per_case.py b/docs/research/artifacts/rectification-width-funnel-2026-10-16/prod_per_case.py
new file mode 100644
index 00000000..b65abc5e
--- /dev/null
+++ b/docs/research/artifacts/rectification-width-funnel-2026-10-16/prod_per_case.py
@@ -0,0 +1,36 @@
+"""Per-case production replay (baseline variant) on v5, current staging code."""
+import json, sys, time, traceback
+from pathlib import Path
+WT = Path('/workspace/Jyotisha/.worktrees/karaka-interval-20261016')
+sys.path.insert(0, str(WT)); sys.path.insert(0, str(WT/'scripts'))
+from scripts.research import cluster_width_probe as cwp
+HOLD = WT/'references/real_case_calibration/minute_rectification_holdout_v5.json'
+cases = json.loads(HOLD.read_text())['cases']
+radii = [int(r) for r in (sys.argv[1] if len(sys.argv)>1 else '10,30,60').split(',')]
+out = []
+variant = cwp.Variant('baseline')
+t0=time.time()
+for case in cases:
+    cid = case['case_id']; true_time = str(case['birth']['time'])[:5]
+    for radius in radii:
+        try:
+            request = cwp.scoring_request_for(case, radius)
+            built = cwp.build_event_contribution_matrix(request)
+            rows = cwp.score_from_matrix(request, built)
+            contexts = list(built.get('static_contexts') or [])
+            window_times = [str(r['time'])[:5] for r in rows]
+            raw = cwp.raw_signature_clusters(contexts)
+            probes = cwp.discriminating_event_probes({**request,'refresh_probes':False}, built,
+                scan=cwp.window_scan(built), candidate_times=list(window_times),
+                representative_time=true_time, today=cwp.TODAY)
+            s = cwp.score_variant(raw_clusters=raw, rows=rows, variant=variant, true_time=true_time,
+                window_times=window_times, probes=probes)
+            r = s['replay']
+            out.append({'case_id':cid,'radius':radius,'true':true_time,
+                'engine':{k:s['engine'][k] for k in ('top1_hit','coverage','width')},
+                'replay':{k:r.get(k) for k in ('top1_hit','coverage','width','asked','eliminated_count','union_alive_width')}})
+        except Exception as e:
+            out.append({'case_id':cid,'radius':radius,'error':f'{type(e).__name__}: {e}','trace':traceback.format_exc()[-800:]})
+    print(cid, round(time.time()-t0), flush=True)
+    Path('prod_per_case.json').write_text(json.dumps(out, ensure_ascii=False, indent=1))
+print('DONE')
diff --git a/docs/research/artifacts/rectification-width-funnel-2026-10-16/seg_placebo.py b/docs/research/artifacts/rectification-width-funnel-2026-10-16/seg_placebo.py
new file mode 100644
index 00000000..0b7b3c42
--- /dev/null
+++ b/docs/research/artifacts/rectification-width-funnel-2026-10-16/seg_placebo.py
@@ -0,0 +1,28 @@
+import json, sys, random
+from pathlib import Path
+sys.path.insert(0, str(Path(__file__).parent))
+import karaka_interval as K
+asc = json.load(open('asc_signs.json'))
+cases = json.loads((K.WT/'references/real_case_calibration/minute_rectification_holdout_v5.json').read_text())['cases']
+shard, n = map(int, sys.argv[1].split('/'))
+out = []
+def outcome(cid, R, sel):
+    win = {int(k): v for k, v in asc[cid].items() if abs(int(k)) <= R*60}
+    if len(set(win.values())) < 2: return None
+    t = win[0]; p = {win[g] for g in sel}
+    return 'right' if p == {t} else 'tie' if t in p else 'wrong'
+for i, c in enumerate(cases):
+    if i % n != shard: continue
+    evs = [e for e in c['events'] if e['precision'] in ('day', 'month')]
+    for R in (30, 60, 75):
+        for lv in (3, 4):
+            o = outcome(c['case_id'], R, K.run_case(c, R, lv, 'PS')['sel'])
+            if o is None: continue
+            rng = random.Random(f"seg|{c['case_id']}|{R}|{lv}")
+            pl = []
+            for _ in range(40):
+                ov = [{**e, 'karaka_set': set(rng.sample(K.PLANETS, 2))} for e in evs]
+                pl.append(outcome(c['case_id'], R, K.run_case(c, R, lv, 'PS', ov)['sel']))
+            out.append(dict(case=c['case_id'], R=R, lv=lv, real=o, pl={k: pl.count(k)/40 for k in ('right','tie','wrong')}))
+Path(f'seg_shard{shard}.json').write_text(json.dumps(out))
+print('DONE')
diff --git a/docs/research/rectification_width_funnel_2026_10_16.md b/docs/research/rectification_width_funnel_2026_10_16.md
new file mode 100644
index 00000000..1f2aae33
--- /dev/null
+++ b/docs/research/rectification_width_funnel_2026_10_16.md
@@ -0,0 +1,71 @@
+# 生时校正：交付区间为什么宽 + 「象征星区间求交」对照（2026-10-16）
+
+基线 `origin/staging` @ `9b7509155`。数据集 `references/real_case_calibration/minute_rectification_holdout_v5.json`（77 例公开 AA 名人）。脚本在 `docs/research/artifacts/rectification-width-funnel-2026-10-16/`（脚本里 `WT` 指向当时的只读工作树，复跑时改成本地检出路径）。纯研究，不改线上。
+
+## 1. 起因
+
+产品 10-16 转来一篇外部文章：用自然象征星（Karaka）所主小运求区间交集，把单个案例从 150 分钟收到 3 分钟。产品要求按该方法跑一遍，并与本仓校正对比。之后产品追问：本仓交付 75 分钟，等于没校正，问题在哪。
+
+## 2. 象征星区间求交：复现结论
+
+做法：用关键词规则把事件映射到象征星（`karaka_rules.py`，先计算 sha256 冻结再跑）。只用日或月精度事件（387 件）。网格 15 秒，取「解释事件数最多」的时刻集合，允许丢掉解释不了的事件，与原文一致。对照组是随机象征星，每例 30 次。
+
+| 设置（起始 ±75 分钟） | 交付中位 | 真值在内 | 随机象征星 |
+|---|---|---|---|
+| 第三层小运 | 17 分钟 | 15/77 | 15.0/77 |
+| 第四层小运 | 5 分钟 | 3/77 | 3.4/77 |
+
+- 只用日精度事件（≥3 件的 43 例）：16 组设置无一显著好于随机。第四层还略差于随机（z 低至 −2.2）。
+- 上升星座二选一：第三层小运 z≈1.3，三档窗口方向一致但样本重叠，不显著；第四层无信号。
+- 文章称「1 分钟 = 4.63 天」，只对其个案成立。77 例实测 1.36–5.60 天，中位 3.84 天。
+- 同宽度对比（±60，同为约 4 分钟一段）：文章方法 1～2/43，随机划段约 1.2/43，本仓头名簇 16/43。
+
+结论：该方法在分钟级上的命中与随机无异，不立单。与 `rectification_nadi_seconds_2026_10_05.md` 同属「更细层小运 + 日精度」路线，同样不过。
+
+## 3. 本仓交付区间宽的拆解
+
+方法：`funnel.py` / `funnel2.py`，cluster_width_probe 研究回放。说明三点：题目列表一次性生成，不中途刷新；回答按真值模拟（全对）；「答错一题」为分别翻转第 1、3、5 题。候选步长 2 分钟，「实际剩余」等于活着的候选戳数 × 2。
+
+### 3.1 交付按段内外包络显示
+
+`scripts/rectification/candidate_window.py::interval_union_width` 与前端 `unionWindowIntervals` 按 `segment_index` 合并。`segment_index` 是用户声明的窗口段，不是候选簇，所以同一窗口内被排除的空档也被包进去了。
+
+| 窗口 | 现在显示（最早–最晚） | 实际剩余分钟 |
+|---|---|---|
+| ±10 | 13 | 14 |
+| ±30 | 35 | 26 |
+| ±60 | 71 | 42 |
+
+### 3.2 题目切不动
+
+- 供给：可出题中位 7～8 道。±10 窗口 30/77 例不足 6 道，12 例为 0 道。
+- 切分：Q1–Q2 约 55:45；到 Q5–Q6，真值一侧已占 76–90%，一道题只排掉一两成。
+- 题数从 6 加到 12（静态列表，题够的例子才能问到）：±60 实际剩余 42→38 分钟，头名 30→36。
+
+### 3.3 进区间门槛 8 分 = 至少两道题反向
+
+一道题拉开 4 分（+2 / −2），所以一个候选至少要两道题都站错边才出区间。降门槛的代价是答错一题时容易把真值挤出去：
+
+| 窗口 | 门槛 / 题数 | 实际剩余 | 全答对真值在内 | 答错一题真值在内 |
+|---|---|---|---|---|
+| ±10 | 8 / 6（现状） | 14 | 76/77 | 98% |
+| ±10 | 6 / 12 | 10 | 76/77 | 97% |
+| ±10 | 4 / 12 | 8 | 74/77 | 87% |
+| ±30 | 8 / 6（现状） | 26 | 76/77 | 95% |
+| ±30 | 6 / 12 | 18 | 76/77 | 93% |
+| ±30 | 5 / 12 | 14 | 75/77 | 88% |
+| ±30 | 4 / 12 | 12 | 75/77 | 82% |
+| ±60 | 8 / 6（现状） | 42 | 75/77 | 93% |
+| ±60 | 8 / 12 | 38 | 75/77 | 95% |
+| ±60 | 6 / 12 | 28 | 75/77 | 88% |
+| ±60 | 4 / 12 | 16 | 70/77 | 70% |
+
+全表见 `funnel_summary.json`。门槛 4 在 ±30 / ±60 下，答错一题就挤掉两到三成真值，不可用。门槛 6 是折中。
+
+### 3.4 物理下限
+
+出生时间差 1 分钟，大运边界平均只平移约 3.8 天。用户记忆以月为单位，出题闸 `REFRESH_MIN_BOUNDARY_DAYS=30` 相当于约 8 分钟。所以靠年月问答，最好约 10 分钟，到不了 3–4 分钟。日精度加更细层的路线已两次证伪（纳迪 10-05、象征星本文）。
+
+## 4. 去向
+
+实现单 `docs/tasks/TASK-rectification-narrower-delivery-20261016.md`：分段显示、按剩余候选对半出题并放宽题数、交付区间门槛 8→6（`MIN_SEPARATION_LEAD` 不动）。
diff --git a/docs/tasks/README.md b/docs/tasks/README.md
index 1ca3b98d..3f1e557b 100644
--- a/docs/tasks/README.md
+++ b/docs/tasks/README.md
@@ -36,6 +36,7 @@
 
 | 任务书 | 进度 | 主题 | 状态 | 落点 |
 | --- | --- | --- | --- | --- |
+| `TASK-rectification-narrower-delivery-20261016.md` | — | **交付区间收窄**：显示改按活着的簇分段（±60 显示 71→实际 42 分钟）、按剩余候选对半出题并放宽到 12 轮（少数侧 <20% 不出题）、新增 `DELIVERY_RANGE_LEAD=6` 只管进区间（`MIN_SEPARATION_LEAD=8` 不动，写读同口径）。红线：PG 回放真值保留全对 ≥基线−1、答错一题 ≥基线−5pp，不过则该档回 8。证据 `docs/research/rectification_width_funnel_2026_10_16.md`（含象征星区间求交复现=随机）。BUG 1367 起 | 待领取 | — |
 | —（子代理直接执行，无任务书） | `PROGRESS-bug-1060-d9-double-transit-20260927.md` | **双重过运 D9 目标错配（BUG-1060）**：D9 第 N 宫目标误用 D9 上升、D9 宫主目标误用 D1 度数，改为标签所写的 D9 星座；D1 / CL 层逐字节不变；数据卡 golden 重采集（一张盘跨层结论 9→10）；另记 BUG-1061 跨层数字比对（未修） | 已验收（待部署） | `codex/bug-1060-d9-double-transit-20260927`（未推送） |
 | —（子代理直接执行，无任务书；产品 2026-09-27 授权） | `PROGRESS-bug-1061-cross-layer-pairing-20260927.md` | **双重过运跨层配对按名字数字比对（BUG-1061）**：改为结构化身份配对（D1 第 N 宫↔D9 第 N 宫；D1 第 N 宫主↔同一颗星的 D9 目标）；各层结果与 D1/CL golden 逐字节不变；48 样本跨层条目 31→38；数据卡 golden 重采集（乔布斯 6→9、泰勒 9→10） | 已验收（待部署） | `codex/bug-1061-cross-layer-pairing-20260927`（未推送） |
 | `TASK-rectification-fewer-probes-card-20260926.md` | [PROGRESS](PROGRESS-rectification-fewer-probes-card-20260926.md) | **减少无效追问 + 卡片区间为主**：引导补件离线新增达标 0 → 上限 6→2、定向题问完门槛未达直接出卡；卡片区间为主标题、代表分钟副标题，第一二名差距 ≥5 个百分点才显示百分比，否则写「目前区分不开」。不放宽任何置信度。先做 | 已验收（真机清单欠） | `4e6e8d87` + 验收补改 `7203d94e`（区分不开时隐藏「最可能」），deploy-staging run 2933 已部署；回放真值 ±30/±60 19→20、宽度中位 ±1 分钟、提问 11.4→7.8；Skill 10.0.30；真机清单 `docs/testing/rectification-fewer-probes-card-20260926.md` 未走 |
diff --git a/docs/tasks/TASK-rectification-narrower-delivery-20261016.md b/docs/tasks/TASK-rectification-narrower-delivery-20261016.md
new file mode 100644
index 00000000..0833e031
--- /dev/null
+++ b/docs/tasks/TASK-rectification-narrower-delivery-20261016.md
@@ -0,0 +1,89 @@
+# TASK：生时校正交付区间收窄（分段显示 + 对半出题 + 交付门槛 8→6）2026-10-16
+
+## 0. 基线
+
+- `origin/staging` @ `9b7509155`（开工时以最新 `origin/staging` 为准，并在进度记录里写实测基线 SHA）。
+- 证据与脚本：`docs/research/rectification_width_funnel_2026_10_16.md`，以及 `docs/research/artifacts/rectification-width-funnel-2026-10-16/`（`funnel_summary.json` 是本单的预期数字来源）。
+
+## 1. 事故实证
+
+产品反馈：前后各 60 分钟的窗口问完题，交付约 75 分钟，用户觉得「跟没校正一样」。v5 回放（77 例、模拟答对）现状：
+
+| 窗口 | 显示宽度（中位） | 实际剩余分钟（中位） | 真值在区间 | 答错一题真值在区间 |
+|---|---|---|---|---|
+| ±10 | 13 | 14 | 76/77 | 98% |
+| ±30 | 35 | 26 | 76/77 | 95% |
+| ±60 | 71 | 42 | 75/77 | 93% |
+
+## 2. 根因（按符号定位）
+
+1. **显示按段内外包络。** `scripts/rectification/candidate_window.py::interval_union_width` 的 docstring 写着「Keep historical within-segment envelope」。前端 `frontend/src/lib/rectification-agentic/core/candidate-window.ts::unionWindowIntervals` 按 `segment_index` 把同一段内所有区间合成最早–最晚。`segment_index` 是用户声明的窗口段（只在跨午夜等情况下才有多段），不是候选簇。所以交付卡 `frontend/src/components/rectification-range-delivery.tsx` 的 `rangeLabel`（经 `windowIntervalsLabel`）把中间已出局的空档一起显示了出来。±60 窗口显示 71 分钟，实际只剩 42 分钟。
+2. **题目切不动、题数少。** 可出题中位只有 7～8 道（`scripts/rectification/event_probes.py`：`MAX_PROBES=8`、`REFRESH_MAX_PROBES=12`、`MAX_PROBES_PER_DOMAIN=3`）。Q5–Q6 时剩余候选中 76–90% 与真值同侧，一题只排除一两成。前端轮数上限是 `core/types.ts::DEFAULT_MAX_DISCRIMINATION_ROUNDS = 8`。
+3. **进区间门槛 8 分。** `core/credible-range.ts::unionStillValidIntervals / unionStillValidRange / stillValidCandidates` 用 `MIN_SEPARATION_LEAD = 8`（`core/candidate-separation.ts`）。一道题只拉开 4 分（`SCORE_DELTA` ±2），所以一个候选至少要两道题都反向才出区间。
+4. **物理下限（不修）。** 1 分钟约 3.8 天，加上月级记忆和 `REFRESH_MIN_BOUNDARY_DAYS=30`，问答路线最好约 10 分钟。
+
+## 3. 决策记录（产品 2026-10-16 授权）
+
+- **D1 交付分段显示，推翻「段内外包络」。** 用户可见的范围（交付卡标题、盘型交付 `RectificationSegmentDelivery` 的 `rangeLabel`、旁白和 spoken-grounding 里的范围句、报告里的范围、会话标题）一律按**实际活着的簇区间**显示。间隔 ≤4 分钟的相邻区间合并。合并后超过 3 段时，只显示最靠前的 3 段，再加「等 N 段，共约 X 分钟」。宽度统计（遥测 `width`、交付文案里的「约 X 分钟」）改用实际剩余分钟数。`interval_union_width` 等**决策层**的宽度口径本轮不改（收敛判定、档位、闸门仍按旧口径），在进度记录里列清所有调用点和各自采用的口径。
+- **D2 题数上限放宽 + 按剩余候选对半出题。** `DEFAULT_MAX_DISCRIMINATION_ROUNDS` 从 8 提到 12，探针供给相应放宽（`REFRESH_MAX_PROBES`、每域上限由执行方实测后定，写进进度记录）。每出一题都按当前未淘汰候选的后验质量选题，优先取两侧质量最接近 50:50 的。**少数一侧质量 < 20% 的题不出**，宁可停。这条和 09-26 fewer-probes「减少无效追问」不冲突：那一单管的是引导补件，本单只加能切开剩余候选的点选题。
+- **D3 交付区间门槛 8→6，`MIN_SEPARATION_LEAD` 不动。** 新增常量 `DELIVERY_RANGE_LEAD = 6`，只用于「谁进交付区间」：`credible_range`、`credible_intervals`、`stillValidCandidates` 的交付用途，以及 D1 的显示。下面这些继续用 `MIN_SEPARATION_LEAD = 8`：分离状态（`evaluateCandidateSeparation`）、采用 / 确认 / 熔断门、`segment-summary.ts` 的盘型段汇总（其档位按 8 校准）。
+  - 这样不违反 BUG-560 的防复发条「不得改 `MIN_SEPARATION_LEAD=8`」，也不违反 BUG-423 的「不得用改 `MIN_SEPARATION_LEAD` 绕过死锁」。
+  - 但 BUG-423 的教训必须遵守：**写盘与读盘的 `credible_range` / `credible_intervals` 必须用同一个 lead。** `inference-adapter.ts` 里 `consistent` 的比对两边都换成 `DELIVERY_RANGE_LEAD`，不得出现一边 8、一边 6。
+- **D4 窗口分档回退。** 验收回放里某一档（±10 / ±30 / ±60）在门槛 6 下过不了 §4 红线 2，该档就回到 8。分档写成一张表，挂在 `DELIVERY_RANGE_LEAD` 旁边。
+- **D5 口径。** 交付文案不得承诺「定到分钟」。可以写「剩余约 X 分钟，分 N 段」。
+
+## 4. 硬红线
+
+1. 不改引擎打分：`SCORE_DELTA`、`STRONG_CONFLICT_ELIMINATION_COUNT=3`、`MIN_BOUNDARY_DAYS=45` / `REFRESH_MIN_BOUNDARY_DAYS=30`、`_relative_support`、采用 / 确认门全部不动。不放宽任何置信度。
+2. **真值保留**：v5 77 例，用生产代码路径回放（PG 持久回放 `frontend/tests/research/rectification-segment-persisted-replay.ts` + `scripts/research/varga_resolution_persisted_replay.py`，按段选题开关保持默认）。每档同时满足两条：
+   - 全答对时，真值在交付区间内的例数 ≥ 基线 −1；
+   - 答错一题（分别翻转第 1、3、5 题）时，真值在区间内的比例 ≥ 基线 −5 个百分点。
+   基线在同一 harness 上用改动前代码实测，不用本单研究数字代替。
+3. 预期（研究回放，静态题单，仅作参照）：门槛 6、12 题时实际剩余分钟 ±10 约 10、±30 约 18、±60 约 28。生产回放达不到也不得为了凑数改红线 1 的任何一项，如实报告即可。
+4. 前端按 AGENTS §7：tsc 0、lint 0 error、测试总数不减，既有断言改动写「原值 / 新值 / 原因」，`/` 保持 Static，gzip ±2%。
+5. 校正 Skill 若 bump，必须验证旧 Case 能从列表打开（BUG-621）。
+6. 不动数据库结构；遥测只是写入的数值换口径，若需新列，单独说明并跑 test:db。
+
+## 5. 任务分解
+
+- **T0 基线**：在开工 SHA 上跑 §4 红线 2 的回放，三档 × {全对, 翻 1/3/5}，产出 `docs/research/rectification_narrower_delivery_baseline_2026_10_16.json`。
+  - 验收：数字齐全；有可复跑的命令。
+- **T1 分段显示（D1）**：在 Python 与前端各加一个「活着的区间合并」函数（间隔 ≤4 分钟合并），只供显示与遥测使用；按 D1 列出的所有用户可见出口逐个切换；`windowIntervalsLabel` 支持「等 N 段」。
+  - 验收：
+    - 单测覆盖三种情况：单段、跨空档两段、5 段截 3 段；
+    - 用 golden 中一例 ±60 回放得到的候选构造交付卡，标题段数与实际剩余分钟数和回放一致；
+    - 列出所有 `interval_union_width` / `unionWindowIntervals` 调用点及各自采用的口径；
+    - `frontend/DESIGN.md` 与 `frontend/docs/VOICE.md` 同步更新。
+- **T2 对半选题（D2）**：刷新探针时按剩余后验质量的对半程度排序，少数一侧 < 20% 的题不出；放宽轮数和供给上限。
+  - 验收：
+    - 回放报告给出每题「真值同侧占比」中位，Q4–Q8 必须比基线更接近 0.5；
+    - 平均提问数与问满题的例数写进报告；
+    - 不出现「少数一侧 < 20%」的题（加断言）。
+- **T3 交付门槛（D3 / D4）**：新增 `DELIVERY_RANGE_LEAD` 及分档表，写盘、读盘同口径。
+  - 验收：
+    - BUG-423 合同测试仍绿，并新增「写 6 读 6」的一致性测试；
+    - 分离状态、采用门、segment-summary 仍用 8 的断言；
+    - 回放过 §4 红线 2，否则按 D4 回退并在报告里写明是哪一档。
+- **T4 回放判定**：T1+T2+T3 合起来跑 §4 红线 2 的全套，结果写进 `docs/research/rectification_narrower_delivery_replay_2026_10_16.json`。
+  - 验收：三档逐格对照基线，给出实际剩余分钟、显示宽度、头名命中、平均提问数。
+- **T5 记录**：写 `PROGRESS-rectification-narrower-delivery-20261016.md`、`BUG_HISTORY`（3 条：显示包络、选题不对半、交付门槛）、`CHANGELOG`；真机清单写到 `docs/testing/rectification-narrower-delivery-20261016.md`（±30 窗口完整走一次，看段数和题数）。
+
+## 6. 让步顺序
+
+时间不够时：先交 T0+T1（纯显示，风险最低，单独就能把 ±60 从 71 降到约 42），再交 T3，最后交 T2。任何一项过不了红线 2 就撤掉该项，不能为了保住其他项去放宽红线。
+
+## 7. 开工前置命令
+
+```bash
+git fetch origin --prune
+git worktree add -b codex/rectification-narrower-delivery-20261016 .worktrees/rectification-narrower-delivery-20261016 origin/staging
+cd .worktrees/rectification-narrower-delivery-20261016
+grep -n "BUG-423\|BUG-560\|BUG-653\|BUG-1143\|BUG-1144" docs/BUG_HISTORY.md   # 读完命中记录
+grep -rn "unionWindowIntervals\|interval_union_width\|unionStillValid\|MIN_SEPARATION_LEAD\|DEFAULT_MAX_DISCRIMINATION_ROUNDS" frontend/src scripts tests frontend/tests | wc -l
+```
+
+本机没有 Docker 时，PG 回放按 `docs/research/` 中盘型口径研究记录的「无 Docker 验收法」（预编译 PG17 + psql / docker 替身）跑。
+
+## 8. BUG 编号起点
+
+BUG-1367 起，开工时用 `docs/BUG_HISTORY.md` 的最大号核对一遍。