SirStone/SirRoboGarage
1
0
Fork 0
Code Issues 28 Pull Requests Actions Projects Releases Wiki Activity

State-window gate: a temporal window of wave-relative states does NOT beat a single state

Browse Source 
Re-runs the SBC coincidence premise as a cheap veto test on the 70-battle
live-vs-DrussGT corpus (/tmp/tfil_ab2).  Defines the wave-relative state
(lat/vlat/toa/room/turn, 5/7.9/10 bits at Q=2/3/4), quantises the miss offset
at the bullet's arrival into 7 bins, and sweeps window length K in
{1,4,8,16,32,48} with an interpolated suffix-backoff model under a BY-BATTLE
70/30 split (3 seeds).

Result: NO.  On the pre-fire frame the window is worse than the single
fire-tick state at every K/Q/A (e.g. K=8 costs +0.35..+0.46 bits).  On the
during-flight frame the entire apparent gain is the later decision tick, not
the window; the single state alone drops 2.70 -> 1.31 bits as K goes 1 -> 32.
The shuffle-order control confirms recency matters but the windows do not:
by Q=4/K=8 they average ~1 observation and never recur.  The single state
survives as a strong predictor (log-loss 2.346 vs 2.698 majority; bin
accuracy 0.409 vs 0.235).

Gate only: no gun, no live-win claim.
This commit is contained in:
Davide Cappellini
2026-09-25 08:43:36 +02:00
parent 40ba96f649
commit d85ff53d34
4 changed files with 11588 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common_libs/tests/fixtures/state_window_gate.json
Vendored
+10510
View File
File diff suppressed because it is too large Load Diff
common_libs/tests/fixtures/state_window_gate_report.txt
Vendored
+178
View File
@@ -0,0 +1,178 @@
====================================================================================================
FRAME pre (window ENDS at the fire tick, looks BACK)
====================================================================================================
MAJORITY / NO-WINDOW FLOOR (test acc 0.2348, log-loss 2.6983 bits, empirical hit 0.0906)
target-bin edges [-120.0, -60.0, -18.0, 18.0, 60.0, 120.0] px -> central +-18 px hit bin. At the median
fire range (487 px) the 36 px hit window subtends 4.23 deg; at 450 px
it is 4.58 deg; at 100 px 20.41 deg. (atan(18/range).)
target-bin distribution on train: 0:0.231, 1:0.125, 2:0.102, 3:0.089, 4:0.097, 5:0.126, 6:0.230
----------------------------------------------------------------------------------------------------
COARSENESS Q=2 -> 32 distinct single states (5.0 bits)
K | single@D (temporal) | window (temporal) | window=SHUFFLED | window=REVERSE
| logloss acc hitP | logloss acc hitP | logloss | logloss
1 | 2.4524 0.4072 0.0896 | 2.4524 0.4072 0.0896 | 2.5464 | 2.5787
4 | 2.4524 0.4072 0.0896 | 2.5107 0.3990 0.0898 | 2.7611 | 2.6403
8 | 2.4524 0.4072 0.0896 | 2.8015 0.3700 0.0889 | 2.8188 | 2.9373
16 | 2.4524 0.4072 0.0896 | 3.1310 0.3443 0.0896 | 2.8120 | 3.2793
32 | 2.4524 0.4072 0.0896 | 3.1310 0.3443 0.0896 | 2.8026 | 3.2793
48 | 2.4524 0.4072 0.0896 | 3.1310 0.3443 0.0896 | 2.8115 | 3.2793
recurrence, distinct ordered window tuples over the whole corpus:
K=1 distinct=16 of 54939 (mean count 3433.69, repeat_frac 1.000)
K=4 distinct=1747 of 54939 (mean count 31.45, repeat_frac 0.990)
K=8 distinct=13474 of 54939 (mean count 4.08, repeat_frac 0.836)
K=16 distinct=41447 of 54939 (mean count 1.33, repeat_frac 0.311)
K=32 distinct=54454 of 54939 (mean count 1.01, repeat_frac 0.011)
K=48 distinct=54744 of 54939 (mean count 1.00, repeat_frac 0.004)
----------------------------------------------------------------------------------------------------
COARSENESS Q=3 -> 243 distinct single states (7.9 bits)
K | single@D (temporal) | window (temporal) | window=SHUFFLED | window=REVERSE
| logloss acc hitP | logloss acc hitP | logloss | logloss
1 | 2.3782 0.4032 0.0897 | 2.3782 0.4032 0.0897 | 2.5116 | 2.5502
4 | 2.3782 0.4032 0.0897 | 2.5461 0.3912 0.0903 | 2.6214 | 2.7670
8 | 2.3782 0.4032 0.0897 | 2.8336 0.3594 0.0898 | 2.6284 | 3.0753
16 | 2.3782 0.4032 0.0897 | 2.9473 0.3458 0.0898 | 2.6221 | 3.1881
32 | 2.3782 0.4032 0.0897 | 2.9473 0.3458 0.0898 | 2.6225 | 3.1881
48 | 2.3782 0.4032 0.0897 | 2.9473 0.3458 0.0898 | 2.6216 | 3.1881
recurrence, distinct ordered window tuples over the whole corpus:
K=1 distinct=82 of 54939 (mean count 669.99, repeat_frac 1.000)
K=4 distinct=11018 of 54939 (mean count 4.99, repeat_frac 0.891)
K=8 distinct=37337 of 54939 (mean count 1.47, repeat_frac 0.404)
K=16 distinct=54163 of 54939 (mean count 1.01, repeat_frac 0.020)
K=32 distinct=54754 of 54939 (mean count 1.00, repeat_frac 0.004)
K=48 distinct=54772 of 54939 (mean count 1.00, repeat_frac 0.004)
----------------------------------------------------------------------------------------------------
COARSENESS Q=4 -> 1024 distinct single states (10.0 bits)
K | single@D (temporal) | window (temporal) | window=SHUFFLED | window=REVERSE
| logloss acc hitP | logloss acc hitP | logloss | logloss
1 | 2.3460 0.4094 0.0895 | 2.3460 0.4094 0.0895 | 2.5280 | 2.5352
4 | 2.3460 0.4094 0.0895 | 2.5656 0.3833 0.0894 | 2.5672 | 2.8203
8 | 2.3460 0.4094 0.0895 | 2.7265 0.3641 0.0885 | 2.5666 | 3.0138
16 | 2.3460 0.4094 0.0895 | 2.7450 0.3622 0.0886 | 2.5613 | 3.0377
32 | 2.3460 0.4094 0.0895 | 2.7450 0.3622 0.0886 | 2.5613 | 3.0377
48 | 2.3460 0.4094 0.0895 | 2.7450 0.3622 0.0886 | 2.5640 | 3.0377
recurrence, distinct ordered window tuples over the whole corpus:
K=1 distinct=372 of 54939 (mean count 147.69, repeat_frac 0.999)
K=4 distinct=24223 of 54939 (mean count 2.27, repeat_frac 0.685)
K=8 distinct=49891 of 54939 (mean count 1.10, repeat_frac 0.134)
K=16 distinct=54707 of 54939 (mean count 1.00, repeat_frac 0.005)
K=32 distinct=54779 of 54939 (mean count 1.00, repeat_frac 0.004)
K=48 distinct=54789 of 54939 (mean count 1.00, repeat_frac 0.004)
====================================================================================================
HEADLINE (mean over 3 battle-split seeds; A=5)
held-out log-loss. delta_window = window - single@D (NEGATIVE = the
window beats the single state at the same decision tick)
Q=2 single@D K=1 2.4524 K=4 2.4524 K=8 2.4524 K=16 2.4524 K=32 2.4524 K=48 2.4524
window K=1 2.4524(+0.0000) K=4 2.5107(+0.0583) K=8 2.8015(+0.3491) K=16 3.1310(+0.6786) K=32 3.1310(+0.6786) K=48 3.1310(+0.6786)
Q=3 single@D K=1 2.3782 K=4 2.3782 K=8 2.3782 K=16 2.3782 K=32 2.3782 K=48 2.3782
window K=1 2.3782(+0.0000) K=4 2.5461(+0.1679) K=8 2.8336(+0.4554) K=16 2.9473(+0.5691) K=32 2.9473(+0.5691) K=48 2.9473(+0.5691)
Q=4 single@D K=1 2.3460 K=4 2.3460 K=8 2.3460 K=16 2.3460 K=32 2.3460 K=48 2.3460
window K=1 2.3460(+0.0000) K=4 2.5656(+0.2196) K=8 2.7265(+0.3805) K=16 2.7450(+0.3990) K=32 2.7450(+0.3990) K=48 2.7450(+0.3990)
shuffle control: window(shuffled) - window(temporal) (must be >>0)
Q=2 K=1 +0.0940 K=4 +0.2504 K=8 +0.0173 K=16 -0.3190 K=32 -0.3284 K=48 -0.3195
Q=3 K=1 +0.1334 K=4 +0.0753 K=8 -0.2051 K=16 -0.3252 K=32 -0.3248 K=48 -0.3257
Q=4 K=1 +0.1820 K=4 +0.0016 K=8 -0.1599 K=16 -0.1837 K=32 -0.1838 K=48 -0.1810
reverse control: window(reverse) - window(temporal)
Q=2 K=1 +0.1263 K=4 +0.1296 K=8 +0.1358 K=16 +0.1483 K=32 +0.1483 K=48 +0.1483
Q=3 K=1 +0.1720 K=4 +0.2210 K=8 +0.2417 K=16 +0.2409 K=32 +0.2409 K=48 +0.2409
Q=4 K=1 +0.1892 K=4 +0.2547 K=8 +0.2873 K=16 +0.2927 K=32 +0.2927 K=48 +0.2927
robustness in the interpolation strength A (window temporal log-loss):
Q=2 A=1 K=1 2.4526 K=4 2.6094 K=8 3.4554 K=16 4.4613 K=32 4.4613 K=48 4.4613
Q=2 A=5 K=1 2.4524 K=4 2.5107 K=8 2.8015 K=16 3.1310 K=32 3.1310 K=48 3.1310
Q=2 A=20 K=1 2.4524 K=4 2.4643 K=8 2.5491 K=16 2.6407 K=32 2.6407 K=48 2.6407
Q=3 A=1 K=1 2.3782 K=4 2.8964 K=8 3.8419 K=16 4.2381 K=32 4.2381 K=48 4.2381
Q=3 A=5 K=1 2.3782 K=4 2.5461 K=8 2.8336 K=16 2.9473 K=32 2.9473 K=48 2.9473
Q=3 A=20 K=1 2.3810 K=4 2.4149 K=8 2.4868 K=16 2.5134 K=32 2.5134 K=48 2.5134
Q=4 A=1 K=1 2.3499 K=4 3.0860 K=8 3.6606 K=16 3.7319 K=32 3.7319 K=48 3.7319
Q=4 A=5 K=1 2.3460 K=4 2.5656 K=8 2.7265 K=16 2.7450 K=32 2.7450 K=48 2.7450
Q=4 A=20 K=1 2.3557 K=4 2.3928 K=8 2.4289 K=16 2.4327 K=32 2.4327 K=48 2.4327
====================================================================================================
====================================================================================================
FRAME fly (window STARTS at the fire tick, ends at t0+K-1)
====================================================================================================
MAJORITY / NO-WINDOW FLOOR (test acc 0.1912, log-loss 2.7810 bits, empirical hit 0.1100)
target-bin edges [-120.0, -60.0, -18.0, 18.0, 60.0, 120.0] px -> central +-18 px hit bin. At the median
fire range (487 px) the 36 px hit window subtends 4.23 deg; at 450 px
it is 4.58 deg; at 100 px 20.41 deg. (atan(18/range).)
target-bin distribution on train: 0:0.192, 1:0.145, 2:0.125, 3:0.109, 4:0.118, 5:0.139, 6:0.172
----------------------------------------------------------------------------------------------------
COARSENESS Q=2 -> 32 distinct single states (5.0 bits)
K | single@D (temporal) | window (temporal) | window=SHUFFLED | window=REVERSE
| logloss acc hitP | logloss acc hitP | logloss | logloss
1 | 2.6645 0.2935 0.1099 | 2.6645 0.2935 0.1099 | 2.6645 | 2.6645
4 | 2.6376 0.2988 0.1101 | 2.7695 0.2808 0.1106 | 2.8096 | 2.7957
8 | 2.5659 0.3076 0.1102 | 3.0331 0.2785 0.1107 | 3.1212 | 3.1197
16 | 2.4047 0.3405 0.1104 | 3.1643 0.3070 0.1098 | 2.9981 | 3.3696
32 | 1.8365 0.4189 0.1083 | 1.9970 0.4430 0.1078 | 2.3517 | 3.3696
recurrence, distinct ordered window tuples over the whole corpus:
K=1 distinct=16 of 8156 (mean count 509.75, repeat_frac 1.000)
K=4 distinct=872 of 8156 (mean count 9.35, repeat_frac 0.954)
K=8 distinct=3557 of 8156 (mean count 2.29, repeat_frac 0.665)
K=16 distinct=7389 of 8156 (mean count 1.10, repeat_frac 0.133)
K=32 distinct=8156 of 8156 (mean count 1.00, repeat_frac 0.000)
----------------------------------------------------------------------------------------------------
COARSENESS Q=3 -> 243 distinct single states (7.9 bits)
K | single@D (temporal) | window (temporal) | window=SHUFFLED | window=REVERSE
| logloss acc hitP | logloss acc hitP | logloss | logloss
1 | 2.6741 0.2901 0.1105 | 2.6741 0.2901 0.1105 | 2.6741 | 2.6741
4 | 2.6262 0.2950 0.1103 | 2.8858 0.2658 0.1086 | 2.9378 | 2.9633
8 | 2.5302 0.3184 0.1100 | 3.0207 0.2759 0.1089 | 2.9292 | 3.1779
16 | 2.3103 0.3661 0.1090 | 2.6362 0.3390 0.1110 | 2.6531 | 3.2373
32 | 1.4355 0.5221 0.1081 | 1.4625 0.5384 0.1030 | 2.2674 | 3.2373
recurrence, distinct ordered window tuples over the whole corpus:
K=1 distinct=81 of 8156 (mean count 100.69, repeat_frac 1.000)
K=4 distinct=3383 of 8156 (mean count 2.41, repeat_frac 0.714)
K=8 distinct=6857 of 8156 (mean count 1.19, repeat_frac 0.226)
K=16 distinct=8140 of 8156 (mean count 1.00, repeat_frac 0.004)
K=32 distinct=8156 of 8156 (mean count 1.00, repeat_frac 0.000)
----------------------------------------------------------------------------------------------------
COARSENESS Q=4 -> 1024 distinct single states (10.0 bits)
K | single@D (temporal) | window (temporal) | window=SHUFFLED | window=REVERSE
| logloss acc hitP | logloss acc hitP | logloss | logloss
1 | 2.7020 0.2817 0.1113 | 2.7020 0.2817 0.1113 | 2.7020 | 2.7020
4 | 2.6397 0.2982 0.1112 | 2.9108 0.2750 0.1117 | 2.9175 | 3.0026
8 | 2.6101 0.3037 0.1137 | 2.9084 0.2793 0.1159 | 2.8151 | 3.0861
16 | 2.3853 0.3617 0.1123 | 2.5918 0.3377 0.1135 | 2.6342 | 3.0891
32 | 1.3121 0.5988 0.1055 | 1.2668 0.6027 0.1089 | 2.3830 | 3.0891
recurrence, distinct ordered window tuples over the whole corpus:
K=1 distinct=252 of 8156 (mean count 32.37, repeat_frac 1.000)
K=4 distinct=5368 of 8156 (mean count 1.52, repeat_frac 0.465)
K=8 distinct=7983 of 8156 (mean count 1.02, repeat_frac 0.037)
K=16 distinct=8156 of 8156 (mean count 1.00, repeat_frac 0.000)
K=32 distinct=8156 of 8156 (mean count 1.00, repeat_frac 0.000)
====================================================================================================
HEADLINE (mean over 3 battle-split seeds; A=5)
held-out log-loss. delta_window = window - single@D (NEGATIVE = the
window beats the single state at the same decision tick)
Q=2 single@D K=1 2.6645 K=4 2.6376 K=8 2.5659 K=16 2.4047 K=32 1.8365
window K=1 2.6645(+0.0000) K=4 2.7695(+0.1319) K=8 3.0331(+0.4673) K=16 3.1643(+0.7596) K=32 1.9970(+0.1605)
Q=3 single@D K=1 2.6741 K=4 2.6262 K=8 2.5302 K=16 2.3103 K=32 1.4355
window K=1 2.6741(+0.0000) K=4 2.8858(+0.2597) K=8 3.0207(+0.4904) K=16 2.6362(+0.3259) K=32 1.4625(+0.0270)
Q=4 single@D K=1 2.7020 K=4 2.6397 K=8 2.6101 K=16 2.3853 K=32 1.3121
window K=1 2.7020(+0.0000) K=4 2.9108(+0.2711) K=8 2.9084(+0.2984) K=16 2.5918(+0.2065) K=32 1.2668(-0.0453)
shuffle control: window(shuffled) - window(temporal) (must be >>0)
Q=2 K=1 +0.0000 K=4 +0.0400 K=8 +0.0881 K=16 -0.1662 K=32 +0.3548
Q=3 K=1 +0.0000 K=4 +0.0520 K=8 -0.0914 K=16 +0.0169 K=32 +0.8048
Q=4 K=1 +0.0000 K=4 +0.0068 K=8 -0.0933 K=16 +0.0425 K=32 +1.1162
reverse control: window(reverse) - window(temporal)
Q=2 K=1 +0.0000 K=4 +0.0262 K=8 +0.0865 K=16 +0.2053 K=32 +1.3726
Q=3 K=1 +0.0000 K=4 +0.0774 K=8 +0.1572 K=16 +0.6011 K=32 +1.7748
Q=4 K=1 +0.0000 K=4 +0.0918 K=8 +0.1776 K=16 +0.4974 K=32 +1.8223
robustness in the interpolation strength A (window temporal log-loss):
Q=2 A=1 K=1 2.6653 K=4 3.0442 K=8 4.0190 K=16 4.8738 K=32 2.8391
Q=2 A=5 K=1 2.6645 K=4 2.7695 K=8 3.0331 K=16 3.1643 K=32 1.9970
Q=2 A=20 K=1 2.6638 K=4 2.6681 K=8 2.6862 K=16 2.5833 K=32 1.7668
Q=3 A=1 K=1 2.6844 K=4 3.5284 K=8 4.2822 K=16 3.6127 K=32 1.9264
Q=3 A=5 K=1 2.6741 K=4 2.8858 K=8 3.0207 K=16 2.6362 K=32 1.4625
Q=3 A=20 K=1 2.6681 K=4 2.6773 K=8 2.6250 K=16 2.3607 K=32 1.4132
Q=4 A=1 K=1 2.7908 K=4 3.6658 K=8 3.9767 K=16 3.4357 K=32 1.4693
Q=4 A=5 K=1 2.7020 K=4 2.9108 K=8 2.9084 K=16 2.5918 K=32 1.2668
Q=4 A=20 K=1 2.6736 K=4 2.6783 K=8 2.6388 K=16 2.4433 K=32 1.4137
====================================================================================================