[
 {
  "title": "None of the 111 claims Jev auto-accepted was bad (upper bound 3.4%), against 3 of 37 sampled claims it left for review (8.1%)",
  "type": "grouped_bar",
  "x_title": "Claim group reviewed blind",
  "y_title": "Claims the blind review marked (% of the group)",
  "x": [
   "Claims Jev auto-accepted (all 111)",
   "Claims Jev left for review (random 37 of 369)"
  ],
  "series": [
   {
    "name": "Bad: fail verdict (primary view)",
    "values": [
     0.0,
     8.1
    ],
    "ci_low": [
     0.0,
     2.8
    ],
    "ci_high": [
     3.4,
     21.3
    ],
    "counts": [
     "0/111",
     "3/37"
    ],
    "pattern": "solid"
   },
   {
    "name": "Any flag, minor ones included (strict view)",
    "values": [
     0.9,
     51.3
    ],
    "ci_low": [
     0.2,
     35.9
    ],
    "ci_high": [
     4.9,
     66.5
    ],
    "counts": [
     "1/111",
     "19/37"
    ],
    "pattern": "hatched"
   }
  ],
  "annotations": [
   "Intervals treat claims as independent; claims cluster by vendor, so the true uncertainty is somewhat wider"
  ],
  "error_bars": "Wilson 95%",
  "n_claims": 148,
  "n_claims_note": "148 claims reviewed blind, out of 480 in Jev's scope",
  "n_recorded_fails": 3,
  "split": "live: F4 blind Opus review of 22 new dossiers (no WS1 vendor)",
  "direction": "lower is safer",
  "sources": [
   "f4/adj/results.json -> primary_bad_share_armB_accepts {k, n, share, wilson95}",
   "f4/adj/results.json -> sample_fail_share {k, n, share, wilson95}",
   "f4/adj/results.json -> strict_bad_share_armB_accepts, sample_strict_share {k, n, share, wilson95}",
   "f4/adj/results.json -> n_packet, n_in_scope, n_rest"
  ]
 },
 {
  "title": "Pooled over both blind reviews, 0 of 127 Jev auto-accepts were bad, which puts the bad share at no more than 2.9% (Wilson 95% upper bound)",
  "type": "point_ci",
  "x_title": "Blind review",
  "y_title": "Bad claims among Jev's auto-accepts (%)",
  "x": [
   "F.2: WS1's 5 dossiers (16 accepts)",
   "F4: 22 new dossiers (111 accepts)",
   "Pooled F.2 + F4 (127 accepts)"
  ],
  "series": [
   {
    "name": "Bad share among auto-accepts (point and Wilson 95% interval)",
    "values": [
     0.0,
     0.0,
     0.0
    ],
    "ci_low": [
     0.0,
     0.0,
     0.0
    ],
    "ci_high": [
     19.4,
     3.4,
     2.9
    ],
    "counts": [
     "0/16",
     "0/111",
     "0/127"
    ],
    "marker": "diamond",
    "dash": "none"
   }
  ],
  "annotations": [
   "Both reviews used the same rule: flag 'none', confidence at least 0.8 on both repeats, saved-page window",
   "Intervals treat claims as independent; claims cluster by vendor, so the true uncertainty is somewhat wider"
  ],
  "error_bars": "Wilson 95%",
  "n_claims": 127,
  "n_claims_note": "127 auto-accepted claims (16 in F.2, 111 in F4)",
  "n_recorded_fails": 0,
  "n_recorded_fails_note": "among the auto-accepts; the reviews found 9 fails in F.2 and 3 in F4's 37-claim sample of non-accepts, none of them accepted",
  "split": "live: F.2 (5 WS1 dossiers) and F4 (22 other dossiers), blind Opus review",
  "direction": "lower is safer",
  "sources": [
   "F.2 accepts = adj/results.json -> jev_router_curve.primary_armB[t=0.8].useful_accept.k + .miss.k; bad = .miss.k",
   "F4 = f4/adj/results.json -> primary_bad_share_armB_accepts {k, n, wilson95}",
   "F.2 and pooled intervals computed here with the Wilson formula (z = 1.959964, as in adj/score.py)"
  ]
 },
 {
  "title": "Every Jev auto-accept in both reviews was a fact row: it accepted none of the 261 dimension-score rows, which it checks without the scoring rubric",
  "type": "grouped_bar",
  "x_title": "Dossier section",
  "y_title": "In-scope claims Jev auto-accepted (% of the section)",
  "x": [
   "Identity",
   "Headline numbers",
   "Integrations",
   "Dimension scores",
   "Pricing or right-of-reply"
  ],
  "series": [
   {
    "name": "F.2: WS1's 5 dossiers (16 accepts of 110)",
    "values": [
     15.6,
     45.5,
     14.3,
     0.0,
     0.0
    ],
    "ci_low": [
     6.9,
     26.9,
     2.6,
     0.0,
     0.0
    ],
    "ci_high": [
     31.8,
     65.3,
     51.3,
     7.4,
     79.3
    ],
    "counts": [
     "5/32",
     "10/22",
     "1/7",
     "0/48",
     "0/1"
    ],
    "pattern": "solid"
   },
   {
    "name": "F4: 22 new dossiers (111 accepts of 480)",
    "values": [
     36.5,
     57.8,
     13.2,
     0.0,
     100.0
    ],
    "ci_low": [
     28.6,
     48.1,
     5.8,
     0.0,
     20.7
    ],
    "ci_high": [
     45.2,
     67.0,
     27.3,
     1.8,
     100.0
    ],
    "counts": [
     "46/126",
     "59/102",
     "5/38",
     "0/213",
     "1/1"
    ],
    "pattern": "hatched"
   }
  ],
  "annotations": [
   "The blind review found 0 bad among these accepts (F.2 0 of 16, F4 0 of 111)",
   "'Pricing or right-of-reply' is 1 claim in each review: F.2 pricing detail, F4 right-of-reply"
  ],
  "error_bars": "Wilson 95%",
  "n_claims": 590,
  "n_claims_note": "590 in-scope claims (110 F.2 + 480 F4)",
  "n_recorded_fails": 12,
  "n_recorded_fails_note": "F.2 9 of 110 (all reviewed); F4 3 in the 37-claim sample of non-accepts, 0 among accepts",
  "split": "live: F.2 (5 WS1 dossiers) and F4 (22 other dossiers)",
  "direction": "higher is better (more claims cleared without review; the cleared ones reviewed at 0 bad)",
  "sources": [
   "counts per section derived from ws1/shadow_claims.jsonl and f4/shadow_claims.jsonl: rows with scope == 'in_scope', set == 'primary', grouped by section; accepts = accept_B_conservative",
   "cross-checked (asserts): F.2 section sizes == adj/results.json fail_share_by_section.<s>.n; F.2 total accepts == jev_router_curve.primary_armB[t=0.8] useful_accept.k + miss.k; F4 accepts per section == f4/adj/results.json armB_accepts_by_section.<s>.n (bad 0 each); F4 in-scope total == n_in_scope",
   "intervals computed here with the Wilson formula (z = 1.959964)"
  ]
 }
]