Hugging Face's logo

end000
/
Checkers-WSPForge-Russian-v2

WSPForge
weight-space-programming
russian-draughts
checkers
custom-architecture
numpy
scipy
no-gradient-training
Model card Files
xet
 Community
Checkers-WSPForge-Russian-v2
File size: 2,921 Bytes
6069338
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
{
  "name": "Checkers-WSPForge-Russian-v2",
  "method": "direct_analytic_weights",
  "training": false,
  "search_in_player": false,
  "score_scale": 10,
  "composition": "Exact sum of two affine/ReLU networks with positive/negative score channels",
  "base": "c0-integer-v1",
  "base_sha256": "eaa13632e04f6312ce2192429493dc004204a770c649adf68fbe0ef5c6cbce5e",
  "patch": {
    "name": "r1-complete-safe-patch",
    "method": "direct_analytic_weights",
    "training": false,
    "search_in_player": false,
    "score_scale": 10,
    "safe": true,
    "best_only": false,
    "sources": 56,
    "static_routes": 1254,
    "mechanism": "OR of fixed forward diagonal empty-cell paths per man, optionally gated by static capture exposure after source vacancy. Direct discounted promotion potential, summed or max-pooled by side.",
    "discount": "future king-minus-man value 2000 scaled points, discounted by 1/2 per quiet step",
    "limitations": [
      "Opponent pieces are held fixed for this geometric feature",
      "Mandatory captures and future enemy quiet moves can invalidate an apparent corridor",
      "Immediate single-capture safety only; no game tree or rollout in inference"
    ],
    "sha256": "7b1a0f0bcef6caefc92b428dbb5227b8623353306b6d4ad582e6a38fce118e32",
    "layers": 5,
    "architecture": [
      128,
      160,
      5144,
      1254,
      112,
      1
    ],
    "nonzero_weights": 110416,
    "nonzero_biases": 6326,
    "dense_parameters_including_zeros": 7441327,
    "neurons_excluding_input": 6671,
    "sparse_array_bytes": 936716,
    "float_weight_and_bias_bytes": 468348
  },
  "mechanism": "Integer-scaled v1 positional/safety network plus a directly programmed static safe-promotion-route circuit. All strategic values are computed by fixed affine/ReLU layers.",
  "sha256": "e996352b06d795395c227c08438c5ae831bdd0cc3cf7bb177ee8f613a53f5286",
  "layers": 6,
  "architecture": [
    128,
    384,
    6288,
    1527,
    114,
    4,
    1
  ],
  "nonzero_weights": 117704,
  "nonzero_biases": 7447,
  "dense_parameters_including_zeros": 12248376,
  "neurons_excluding_input": 8318,
  "sparse_array_bytes": 1008200,
  "float_weight_and_bias_bytes": 504088,
  "selected_experimental_name": "r3-v1-safe",
  "input": "128 binary planes: own men, own kings, opposing men, opposing kings; 32 dark squares b8,d8,...,g1, rotated 180 degrees for black",
  "output": "One scalar per legal afterstate; pick maximum, seeded ties",
  "evidence": "Frozen independent D3-D5 cohort, separately planned D3 replication, architecture-matched route-readout ablation",
  "limitations": [
    "Still loses most games to SimpleEngine D3-D5",
    "Static corridors hold opponent pieces fixed and are not guaranteed winning plans",
    "No significant head-to-head advantage established over v1 in the compact check",
    "D4/D5 improvements not established after multiple-comparison adjustment"
  ]
}