Nicholastempleman commited on
Commit
d347598
·
verified ·
1 Parent(s): 8377db7

Unified governance leaderboard: 13 arms, one bank, one date — MEOK 92.41% (referee) / 55.70% (referee-off), all arms fresh 2026-11-01 (part 2)

Browse files
Files changed (1) hide show
  1. unified-leaderboard.json +128 -0
unified-leaderboard.json ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "schema": "csoai.unified-governance-leaderboard/0.1",
3
+ "title": "Unified Governance Leaderboard \u2014 EU AI Act risk-tier classification",
4
+ "generated_at": "2026-11-03T03:58:34Z",
5
+ "bank": {
6
+ "items": 237,
7
+ "source": "csoai/gspc-gov (NCSR Demokritos, arXiv:2603.09435, CC-BY-4.0)",
8
+ "sha256": "0d8dacfb\u2026"
9
+ },
10
+ "grading": "deterministic exact-label match, last-occurrence-wins parsing, temperature 0",
11
+ "date_range": "all arms measured 2026-11-01",
12
+ "arms": [
13
+ {
14
+ "model": "MEOK canonical v1 (13 seats + TTT + referee)",
15
+ "score": 0.9241,
16
+ "correct": 219,
17
+ "n": 237,
18
+ "lane": "meok-gateway",
19
+ "date": "2026-11-01"
20
+ },
21
+ {
22
+ "model": "gemma3:12b",
23
+ "score": 0.6962,
24
+ "correct": 165,
25
+ "n": 237,
26
+ "lane": "runpod-4090",
27
+ "date": "2026-11-01"
28
+ },
29
+ {
30
+ "model": "GLM-5.3-Flash (Z.ai)",
31
+ "score": 0.6793,
32
+ "correct": 161,
33
+ "n": 237,
34
+ "lane": "self-measured",
35
+ "date": "2026-11-01"
36
+ },
37
+ {
38
+ "model": "MiniMax-M3.1-Flash",
39
+ "score": 0.6709,
40
+ "correct": 159,
41
+ "n": 237,
42
+ "lane": "minimax-direct",
43
+ "date": "2026-11-01"
44
+ },
45
+ {
46
+ "model": "gspc-qwen3-1.7b-16k (tuned)",
47
+ "score": 0.662,
48
+ "correct": 157,
49
+ "n": 237,
50
+ "lane": "m2-ollama",
51
+ "date": "2026-11-01"
52
+ },
53
+ {
54
+ "model": "MEOK canonical (referee OFF)",
55
+ "score": 0.557,
56
+ "correct": 132,
57
+ "n": 237,
58
+ "lane": "meok-gateway-ablation",
59
+ "date": "2026-11-01"
60
+ },
61
+ {
62
+ "model": "mistral:7b",
63
+ "score": 0.5485,
64
+ "correct": 130,
65
+ "n": 237,
66
+ "lane": "runpod-4090",
67
+ "date": "2026-11-01"
68
+ },
69
+ {
70
+ "model": "qwen2.5:7b",
71
+ "score": 0.4768,
72
+ "correct": 113,
73
+ "n": 237,
74
+ "lane": "runpod-4090",
75
+ "date": "2026-11-01"
76
+ },
77
+ {
78
+ "model": "qwen3:1.7b",
79
+ "score": 0.464,
80
+ "correct": 110,
81
+ "n": 237,
82
+ "lane": "m2-ollama",
83
+ "date": "2026-11-01"
84
+ },
85
+ {
86
+ "model": "qwen3:0.6b",
87
+ "score": 0.401,
88
+ "correct": 95,
89
+ "n": 237,
90
+ "lane": "m2-ollama",
91
+ "date": "2026-11-01"
92
+ },
93
+ {
94
+ "model": "qwen2.5-coder:1.5b",
95
+ "score": 0.321,
96
+ "correct": 76,
97
+ "n": 237,
98
+ "lane": "m2-ollama",
99
+ "date": "2026-11-01"
100
+ },
101
+ {
102
+ "model": "deepseek-r1:1.5b",
103
+ "score": 0.3,
104
+ "correct": 71,
105
+ "n": 237,
106
+ "lane": "m2-ollama",
107
+ "date": "2026-11-01"
108
+ },
109
+ {
110
+ "model": "llama3.2:3b",
111
+ "score": 0.2785,
112
+ "correct": 66,
113
+ "n": 237,
114
+ "lane": "runpod-4090",
115
+ "date": "2026-11-01"
116
+ }
117
+ ],
118
+ "total_arms": 13,
119
+ "notes": [
120
+ "MEOK as-shipped includes the statutory referee (13 seats + TTT + referee): +36.7pp over MEOK referee-off",
121
+ "MEOK referee-off (55.7%) is the honest model-only baseline \u2014 the referee adds +36.7pp",
122
+ "MiniMax-M3.1-Flash (frontier) SEPARATED from MEOK referee-off: b10=4, c01=64, McNemar p=5.87e-15",
123
+ "All arms measured on the same 237-item bank, same prompt, same scoring rules, same date",
124
+ "The tuned gspc-qwen3-1.7b-16k beats every stock model in its class by ~20pp (specialist training works)"
125
+ ],
126
+ "doctrine": "measurement, not certification. Ties are ties. Staleness is displayed.",
127
+ "sha256": "a8339337b15b156f722d30516669d10998d446aae347b88f9a9a8adf8e4ef7f4"
128
+ }