|X| = 19683, atoms after semantic dedupe = 523
target 'simple' class counts: {'Illegal': 11093, 'XWon': 4435, 'OWon': 4123, 'Draw': 32}
target 'legal' class counts: {'Illegal': 18725, 'XWon': 626, 'OWon': 316, 'Draw': 16}
  sanity ok: XWon = 626 (matches known reachable-terminal count)
  sanity ok: OWon = 316 (matches known reachable-terminal count)
  sanity ok: Draw = 16 (matches known reachable-terminal count)
D4 orbits: 2862 (Burnside check: 19683 fields total)

=== target 'simple' ===
-- sampling=orbit: 4874 train fields, class counts {'Illegal': 2721, 'XWon': 1134, 'OWon': 1003, 'Draw': 16}
  [simple/orbit] repairs=182 raw_cost=3213.6 (182 rules) -> recompressed=44.6 (3 rules)  [learn 0.9s, recompress 7.3s]
  test accuracy = 100.00%  (0 errors on 14809 unseen fields)
    Illegal  n=  8372 acc=100.00%
    XWon     n=  3301 acc=100.00%
    OWon     n=  3120 acc=100.00%
    Draw     n=    16 acc=100.00%
  baselines: majority=56.53%  1-NN=90.09%  [3s]  memorize-train=9748 bits, memorize-all=39366 bits, certificate=44.6 bits

  Learned certificate (recompressed):
    START: everything -> Illegal   [3.00 bits]
      OVERWRITE #Empty=0 -> Draw   [12.81 bits, |S|=512]
      OVERWRITE N(0, 3, 0)>=1 -> OWon   [14.40 bits, |S|=4435]
      OVERWRITE N(3, 0, 0)>=1 -> XWon   [14.40 bits, |S|=4435]
    TOTAL: 44.61 bits, 3 rules
-- sampling=iid: 5000 train fields, class counts {'Illegal': 2805, 'XWon': 1160, 'OWon': 1028, 'Draw': 7}
  [simple/iid] repairs=147 raw_cost=2456.5 (147 rules) -> recompressed=78.8 (3 rules)  [learn 0.4s, recompress 4.1s]
  test accuracy = 100.00%  (0 errors on 14683 unseen fields)
    Illegal  n=  8288 acc=100.00%
    XWon     n=  3275 acc=100.00%
    OWon     n=  3095 acc=100.00%
    Draw     n=    25 acc=100.00%
  baselines: majority=56.45%  1-NN=89.35%  [3s]  memorize-train=10000 bits, memorize-all=39366 bits, certificate=78.8 bits

=== target 'legal' ===
-- sampling=orbit: 4954 train fields, class counts {'Illegal': 4715, 'XWon': 155, 'OWon': 80, 'Draw': 4}
  [legal/orbit] repairs=240 raw_cost=5107.0 (240 rules) -> recompressed=102.7 (3 rules)  [learn 4.3s, recompress 36.6s]
  test accuracy = 99.65%  (52 errors on 14729 unseen fields)
    Illegal  n= 14010 acc= 99.83%
    XWon     n=   471 acc= 96.60%
    OWon     n=   236 acc=100.00%
    Draw     n=    12 acc=  0.00%
  baselines: majority=95.12%  1-NN=94.30%  [3s]  memorize-train=9908 bits, memorize-all=39366 bits, certificate=102.7 bits

  Learned certificate (recompressed):
    START: everything -> Illegal   [3.00 bits]
      OVERWRITE #X-#O=1 & N(0, 3, 0)=0 & N(3, 0, 0)>=1 -> XWon   [36.95 bits, |S|=626]
      OVERWRITE #X-#O=0 & N(3, 0, 0)=0 & N(0, 3, 0)=1 -> OWon   [36.95 bits, |S|=316]
      OVERWRITE N(1, 2, 0)>=4 & N(2, 1, 0)>=3 -> Draw   [25.80 bits, |S|=44]
    TOTAL: 102.70 bits, 3 rules
-- sampling=iid: 5000 train fields, class counts {'Illegal': 4777, 'XWon': 154, 'OWon': 68, 'Draw': 1}
  [legal/iid] repairs=209 raw_cost=4572.9 (209 rules) -> recompressed=100.9 (3 rules)  [learn 4.0s, recompress 48.0s]
  test accuracy = 99.91%  (13 errors on 14683 unseen fields)
    Illegal  n= 13948 acc= 99.99%
    XWon     n=   472 acc=100.00%
    OWon     n=   248 acc=100.00%
    Draw     n=    15 acc= 20.00%
  baselines: majority=94.99%  1-NN=94.16%  [3s]  memorize-train=10000 bits, memorize-all=39366 bits, certificate=100.9 bits
-- sampling=orbit+rare: 4958 train fields, class counts {'Illegal': 4715, 'XWon': 155, 'OWon': 80, 'Draw': 8}
  [legal/orbit+rare] repairs=216 raw_cost=4635.9 (216 rules) -> recompressed=110.5 (3 rules)  [learn 3.8s, recompress 36.1s]
  test accuracy = 99.82%  (26 errors on 14725 unseen fields)
    Illegal  n= 14010 acc= 99.87%
    XWon     n=   471 acc=100.00%
    OWon     n=   236 acc=100.00%
    Draw     n=     8 acc=  0.00%
  baselines: majority=95.14%  1-NN=94.30%  [3s]  memorize-train=9916 bits, memorize-all=39366 bits, certificate=110.5 bits

=== train-size sweep (target 'legal', orbit sampling) ===
  [sweep 0.025] repairs=36 raw_cost=525.0 (36 rules) -> recompressed=62.1 (3 rules)  [learn 0.0s, recompress 0.4s]
  frac=0.025 n=  481: cost=    62.1 bits (0.129 b/ex), test acc 86.53%
  [sweep 0.050] repairs=68 raw_cost=1107.8 (68 rules) -> recompressed=88.0 (3 rules)  [learn 0.2s, recompress 4.3s]
  frac=0.050 n=  989: cost=    88.0 bits (0.089 b/ex), test acc 98.67%
  [sweep 0.125] repairs=130 raw_cost=2529.0 (130 rules) -> recompressed=101.1 (3 rules)  [learn 0.9s, recompress 18.5s]
  frac=0.125 n= 2457: cost=   101.1 bits (0.041 b/ex), test acc 99.79%
  [sweep 0.250] repairs=240 raw_cost=5107.0 (240 rules) -> recompressed=102.7 (3 rules)  [learn 4.2s, recompress 36.6s]
  frac=0.250 n= 4954: cost=   102.7 bits (0.021 b/ex), test acc 99.65%

=== random-label control (target 'legal' marginals) ===
  [control 250] repairs=26 raw_cost=396.8 (26 rules) -> recompressed=193.2 (7 rules)  [learn 0.0s, recompress 4.1s]
  control n=250: 193 bits (0.77 bits/example), test acc 88.18% (majority 95.13%)
  [control 500] repairs=63 raw_cost=1007.0 (63 rules) -> recompressed=458.2 (17 rules)  [learn 0.1s, recompress 29.2s]
  control n=500: 458 bits (0.92 bits/example), test acc 86.47% (majority 95.15%)
  [control 1000] repairs=89 raw_cost=1723.4 (89 rules) -> recompressed=938.3 (32 rules)  [learn 0.4s, recompress 147.4s]
  control n=1000: 938 bits (0.94 bits/example), test acc 86.00% (majority 95.11%)

total time 404s; wrote /home/nikudaorg/induction_certs/.claude/worktrees/induction-certs-research/experiments/../results/ttt_results.json