seeds=10 tiers=[1000, 3000, 10000, 30000, 100000]
guards: C1=PASS (10/10 pass)
## Part A — bits/word (held-out), kept = (H_uni - H_arm)/(H_uni - H_Gw)
tier H_uni G Gw L R U kept L kept R N1 gain(L)
1000 8.734 6.824 6.941 7.930 9.138 12.111 0.448 -0.225 nan
3000 8.617 6.302 6.420 6.743 8.582 10.201 0.853 0.016 nan
10000 8.567 6.063 6.180 6.255 8.354 8.547 0.969 0.089 -0.152
30000 8.554 5.990 6.105 6.129 8.357 7.453 0.990 0.081 nan
100000 8.551 5.964 6.079 6.097 8.313 6.683 0.993 0.096 0.010
## Part B — hard arms: link F1 (mean) / pair types evaluated (mean)
tier G Gw L R U ceiling
1000 0.661/617 0.638/689 0.465/671 0.034/885 0.071/13850 0.995
3000 0.705/632 0.690/764 0.597/728 0.104/881 0.156/32846 0.995
10000 0.734/635 0.720/829 0.681/782 0.222/834 0.272/77612 0.995
30000 0.735/635 0.724/861 0.698/808 0.274/820 0.380/158566 0.995
100000 0.739/635 0.727/874 0.697/823 0.318/852 0.486/321011 0.995
## Budgeted word-pair search: link F1 (mean) by budget
tier 1000: budget 300 1000 3000 10000 30000 100000 300000
U-f 0.053 0.068 0.071 0.071 0.071 - -
S-L 0.014 0.040 0.056 0.071 0.071 - -
S-R 0.033 0.067 0.071 0.071 0.071 - -
tier 3000: budget 300 1000 3000 10000 30000 100000 300000
U-f 0.089 0.131 0.153 0.156 0.156 0.156 -
S-L 0.009 0.064 0.087 0.146 0.156 0.156 -
S-R 0.021 0.043 0.142 0.156 0.156 0.156 -
tier 10000: budget 300 1000 3000 10000 30000 100000 300000
U-f 0.117 0.181 0.242 0.270 0.272 0.272 -
S-L 0.007 0.033 0.110 0.190 0.265 0.272 -
S-R 0.013 0.028 0.056 0.206 0.272 0.272 -
tier 30000: budget 300 1000 3000 10000 30000 100000 300000
U-f 0.129 0.214 0.290 0.361 0.378 0.380 0.380
S-L 0.010 0.017 0.074 0.180 0.296 0.380 0.380
S-R 0.008 0.016 0.033 0.074 0.249 0.380 0.380
tier 100000: budget 300 1000 3000 10000 30000 100000 300000
U-f 0.139 0.233 0.326 0.416 0.474 0.486 0.486
S-L 0.004 0.014 0.098 0.179 0.272 0.460 0.486
S-R 0.008 0.013 0.024 0.052 0.126 0.438 0.486
## Predictions
PA1: FAIL — kept(L) 10000:0.97 30000:0.99 100000:0.99 | max kept(R) 1000:-0.19 3000:0.03 10000:0.15 30000:0.12 100000:0.12
PA2: PASS — max L/R types 896; U/L ratio 10000:99x 30000:196x 100000:390x
PB1: PASS — L-R 1000:+0.431[+0.406,+0.456] 3000:+0.493[+0.469,+0.517] 10000:+0.460[+0.436,+0.483] 30000:+0.424[+0.405,+0.443] 100000:+0.379[+0.366,+0.391] | G>=Gw>=L every tier: True
PB2: FAIL — F(L)-F(U), types L/U: 1000:+0.394,4.85% 3000:+0.441,2.22% 10000:+0.409,1.01% | at 100000: +0.211
PB3: FAIL — 1000/1000: SL-Uf -0.028[-0.033] SR-Uf -0.001[-0.002]; 1000/3000: SL-Uf -0.015[-0.021] SR-Uf +0.000[-0.000]; 1000/10000: SL-Uf +0.000[+0.000] SR-Uf -0.000[-0.000]; 3000/1000: SL-Uf -0.066[-0.074] SR-Uf -0.087[-0.097]; 3000/3000: SL-Uf -0.066[-0.074] SR-Uf -0.011[-0.016]; 3000/10000: SL-Uf -0.009[-0.016] SR-Uf +0.000[-0.000]; 3000/30000: SL-Uf -0.000[-0.000] SR-Uf +0.000[+0.000]; 10000/1000: SL-Uf -0.148[-0.175] SR-Uf -0.154[-0.159]; 10000/3000: SL-Uf -0.132[-0.140] SR-Uf -0.186[-0.197]; 10000/10000: SL-Uf -0.080[-0.098] SR-Uf -0.064[-0.100]; 10000/30000: SL-Uf -0.008[-0.013] SR-Uf +0.000[-0.000]; 30000/1000: SL-Uf -0.197[-0.200] SR-Uf -0.197[-0.204]; 30000/3000: SL-Uf -0.216[-0.249] SR-Uf -0.257[-0.268]; 30000/10000: SL-Uf -0.181[-0.193] SR-Uf -0.287[-0.303]; 30000/30000: SL-Uf -0.082[-0.106] SR-Uf -0.129[-0.190]; 100000/1000: SL-Uf -0.218[-0.223] SR-Uf -0.219[-0.224]; 100000/3000: SL-Uf -0.229[-0.253] SR-Uf -0.302[-0.309]; 100000/10000: SL-Uf -0.237[-0.253] SR-Uf -0.363[-0.372]; 100000/30000: SL-Uf -0.202[-0.234] SR-Uf -0.348[-0.367]
PB4: PASS — Gw-L at 1000: +0.173, at 100000: +0.031
## Controls
C1: PASS — parser exact on gold edges, every seed
C3: PASS — N1 L gain over unigram: max +0.011 bits/word
C4: PASS — G highest mean F1 among hard arms at every tier
C2: shown to fail under --mut-drop (see commit message); not re-run per seed.