[
 {
  "title": "Jev auto-accepted none of the 9 live fails at any threshold, and with a saved-page window it cleared 16 of 101 good claims at the current 0.8 rule",
  "type": "line",
  "x_title": "Jev confidence threshold for accepting a 'none' finding (0 to 1)",
  "y_title": "Claims auto-accepted without review (% of the claim group)",
  "x": [
   0.5,
   0.6,
   0.7,
   0.8,
   0.9,
   0.95
  ],
  "series": [
   {
    "name": "Good claims accepted, saved-page window (arm B, of 101)",
    "values": [
     22.8,
     21.8,
     18.8,
     15.8,
     7.9,
     4.0
    ],
    "ci_low": [
     15.7,
     14.8,
     12.4,
     10.0,
     4.1,
     1.6
    ],
    "ci_high": [
     31.9,
     30.8,
     27.5,
     24.2,
     14.9,
     9.7
    ],
    "counts": [
     "23/101",
     "22/101",
     "19/101",
     "16/101",
     "8/101",
     "4/101"
    ],
    "dash": "solid",
    "marker": "circle"
   },
   {
    "name": "Good claims accepted, draft's own quote (arm A, of 101)",
    "values": [
     2.0,
     2.0,
     2.0,
     2.0,
     1.0,
     1.0
    ],
    "ci_low": [
     0.5,
     0.5,
     0.5,
     0.5,
     0.2,
     0.2
    ],
    "ci_high": [
     6.9,
     6.9,
     6.9,
     6.9,
     5.4,
     5.4
    ],
    "counts": [
     "2/101",
     "2/101",
     "2/101",
     "2/101",
     "1/101",
     "1/101"
    ],
    "dash": "dashed",
    "marker": "square"
   },
   {
    "name": "Live fails accepted, either arm (of 9; the two arms are identical)",
    "values": [
     0.0,
     0.0,
     0.0,
     0.0,
     0.0,
     0.0
    ],
    "ci_low": [
     0,
     0,
     0,
     0,
     0,
     0
    ],
    "ci_high": [
     29.9,
     29.9,
     29.9,
     29.9,
     29.9,
     29.9
    ],
    "counts": [
     "0/9",
     "0/9",
     "0/9",
     "0/9",
     "0/9",
     "0/9"
    ],
    "dash": "dotted",
    "marker": "triangle"
   }
  ],
  "annotations": [
   "Current rule is threshold 0.8 on both repeats",
   "Points are nudged sideways a little so the error bars do not overlap; the x values are the thresholds shown"
  ],
  "error_bars": "Wilson 95%",
  "n_claims": 110,
  "n_recorded_fails": 9,
  "split": "live: F.2 blind Opus review of WS1's 5 fresh dossiers (not in the claims dataset)",
  "direction": "lower is safer for live fails accepted; higher is better for good claims accepted",
  "sources": [
   "adj/results.json -> jev_router_curve.primary_armB[*].useful_accept {p, lo, hi, k, n} (series 1)",
   "adj/results.json -> jev_router_curve.primary_armA[*].useful_accept {p, lo, hi, k, n} (series 2)",
   "adj/results.json -> jev_router_curve.primary_armB[*].miss {p, lo, hi, k, n}; asserted equal to primary_armA[*].miss (series 3)",
   "adj/results.json -> jev_router_curve.primary_armB[*].t (x)",
   "adj/results.json -> n_labelled (n claims), fail_share.k (n fails)"
  ]
 },
 {
  "title": "Today's two publish gates caught none of the 9 live fails; Jev used as a checker caught 5 to 8 of them but also flagged 38% to 94% of good claims",
  "type": "grouped_bar",
  "x_title": "Check",
  "y_title": "Claims flagged by the check (% of the claim group)",
  "x": [
   "ground.py (production gate)",
   "Step 03, Mistral on the whole draft (production gate)",
   "Jev as checker, draft's own quote (arm A)",
   "Jev as checker, saved-page window (arm B)"
  ],
  "series": [
   {
    "name": "Live fails caught (of 9)",
    "values": [
     0.0,
     0.0,
     88.9,
     55.6
    ],
    "ci_low": [
     0,
     0,
     56.5,
     26.7
    ],
    "ci_high": [
     29.9,
     29.9,
     98.0,
     81.1
    ],
    "counts": [
     "0/9",
     "0/9",
     "8/9",
     "5/9"
    ],
    "pattern": "solid"
   },
   {
    "name": "Good claims flagged (of 101; each one is review work)",
    "values": [
     null,
     null,
     94.1,
     37.6
    ],
    "ci_low": [
     null,
     null,
     87.6,
     28.8
    ],
    "ci_high": [
     null,
     null,
     97.2,
     47.4
    ],
    "counts": [
     "not scored",
     "not scored",
     "95/101",
     "38/101"
    ],
    "pattern": "hatched"
   }
  ],
  "annotations": [
   "'not scored': adj/results.json scores the production gates on the 9 fails only",
   "Jev checker view: the claim counts as flagged if either repeat reports an issue"
  ],
  "error_bars": "Wilson 95%",
  "n_claims": 110,
  "n_recorded_fails": 9,
  "split": "live: F.2 blind Opus review of WS1's 5 fresh dossiers",
  "direction": "higher is better for live fails caught; lower is better for good claims flagged",
  "sources": [
   "adj/results.json -> production_gates.ground_py_caught / step03_caught {p, lo, hi, k, n}",
   "adj/results.json -> jev_checker_view.A|B.fails_caught {p, lo, hi, k, n}",
   "adj/results.json -> jev_checker_view.A|B.passes_flagged {p, lo, hi, k, n}"
  ]
 },
 {
  "title": "Live fails turned up in four of five dossier sections, but with 9 fails in all the section rates overlap and cannot be ranked",
  "type": "bar",
  "x_title": "Dossier section (claims in scope)",
  "y_title": "Claims the blind review failed (% of the section's claims)",
  "x": [
   "Dimension scores (48)",
   "Identity (32)",
   "Headline numbers (22)",
   "Integrations (7)",
   "Pricing detail (1)"
  ],
  "series": [
   {
    "name": "Fail share (primary verdict)",
    "values": [
     4.2,
     12.5,
     9.1,
     14.3,
     0.0
    ],
    "ci_low": [
     1.2,
     5.0,
     2.5,
     2.6,
     0
    ],
    "ci_high": [
     14.0,
     28.1,
     27.8,
     51.3,
     79.3
    ],
    "counts": [
     "2/48",
     "4/32",
     "2/22",
     "1/7",
     "0/1"
    ],
    "pattern": "solid"
   }
  ],
  "annotations": [
   "Descriptive only: the review report says rates by section are shown for description, not comparison"
  ],
  "error_bars": "Wilson 95%",
  "n_claims": 110,
  "n_recorded_fails": 9,
  "split": "live: F.2 blind Opus review of WS1's 5 fresh dossiers",
  "direction": "lower is safer",
  "sources": [
   "adj/results.json -> fail_share_by_section.<section> {k, n, p, lo, hi}"
  ]
 }
]