[
 {
  "title": "Jev let almost no live claims through on the dossier's own receipt text, and about one in seven on the saved-page window",
  "type": "line",
  "x_title": "Jev confidence threshold for accepting a 'none' finding (0 to 1)",
  "y_title": "Accept share of in-scope claims under the 2-repeat rule (%)",
  "x": [
   0.5,
   0.6,
   0.7,
   0.8,
   0.9,
   0.95
  ],
  "series": [
   {
    "name": "Arm A: dossier's own receipt text",
    "values": [
     1.82,
     1.82,
     1.82,
     1.82,
     0.91,
     0.91
    ],
    "dash": "dashed",
    "marker": "circle"
   },
   {
    "name": "Arm B: window from the saved page",
    "values": [
     20.91,
     20.0,
     17.27,
     14.55,
     7.27,
     3.64
    ],
    "dash": "solid",
    "marker": "square"
   }
  ],
  "annotations": [
   "Current rule is threshold 0.8",
   "Accept share is not accuracy: no claim is adjudicated yet"
  ],
  "n_claims": 110,
  "n_recorded_fails": null,
  "n_recorded_fails_note": "not adjudicated yet (blind Opus adjudication pending)",
  "split": "WS1 live pilot: 5 fresh step-02 drafts on frozen bundles (not in the claims dataset)",
  "direction": "higher is better"
 },
 {
  "title": "Four of five fresh dossiers gave no receipt quote for any in-scope claim",
  "type": "bar",
  "x_title": "Dossier (vendor anonymized)",
  "y_title": "In-scope claims that carry a receipt quote (%)",
  "x": [
   "Vendor 1",
   "Vendor 2",
   "Vendor 3",
   "Vendor 4",
   "Vendor 5"
  ],
  "series": [
   {
    "name": "Claims with a receipt",
    "values": [
     31.6,
     0.0,
     3.8,
     0.0,
     0.0
    ],
    "pattern": "solid"
   }
  ],
  "labels_n": [
   19,
   21,
   26,
   19,
   25
  ],
  "n_claims": 110,
  "n_recorded_fails": null,
  "n_recorded_fails_note": "not adjudicated yet",
  "split": "WS1 live pilot",
  "direction": "higher is better"
 },
 {
  "title": "About 45 adjudicated dossiers pin Jev's miss rate to +-5 points and about 125 to +-3, if one live claim in 2.5 fails",
  "type": "line",
  "x_title": "Dossiers adjudicated in full (count)",
  "y_title": "95% interval half-width on Jev's miss rate (percentage points)",
  "x": [
   5,
   10,
   20,
   30,
   45,
   60,
   90,
   125,
   160,
   200
  ],
  "series": [
   {
    "name": "Worst case, true miss rate 50%",
    "values": [
     14.95,
     10.57,
     7.47,
     6.1,
     4.98,
     4.31,
     3.52,
     2.99,
     2.64,
     2.36
    ],
    "dash": "solid",
    "marker": "circle"
   },
   {
    "name": "True miss rate 20%",
    "values": [
     11.96,
     8.45,
     5.98,
     4.88,
     3.99,
     3.45,
     2.82,
     2.39,
     2.11,
     1.89
    ],
    "dash": "dashed",
    "marker": "triangle"
   },
   {
    "name": "True miss rate 10%",
    "values": [
     8.97,
     6.34,
     4.48,
     3.66,
     2.99,
     2.59,
     2.11,
     1.79,
     1.59,
     1.42
    ],
    "dash": "dotted",
    "marker": "square"
   }
  ],
  "assumptions": "fail share among in-scope claims = 0.3909 (Jev arm-B checker view, the only non-zero pilot flag rate; ground.py and step 03 flagged 0 of 110), 22.0 in-scope claims per dossier, no clustering inflation (design effect 1)",
  "n_claims": 110,
  "n_recorded_fails": null,
  "n_recorded_fails_note": "fail share is a proxy until adjudication",
  "split": "WS1 live pilot (projection)",
  "direction": "lower is safer"
 }
]