The evidence layer for AI model claims

Benchmark Integrity Index

A score is f(model, harness, protocol), not f(model). We evaluate the evaluations — grading each benchmark on discrimination, saturation, contamination, and harness comparability — so you know which numbers still mean something.

51 benchmarks graded, most trustworthy first. Every input is source-backed.

ABCDdot size = models scored · 51 benchmarks

#	Grade	Benchmark	Signals	Top score
1	A100	APEX-Agents	consistent harnesspublic test set	Gemini 3.5 Flash 49.6 · 39 models
2	A100	Chess Puzzles	consistent harnesspublic test set	GPT-5.5 Pro 64 · 44 models
3	A100	HLE	consistent harnesspublic test set	Gemini 3.1 Pro 43.74 · 37 models
4	A100	SimpleQA Verified	consistent harnesspublic test set	Gemini 3.1 Pro 77.3 · 51 models
5	A99	ARC-AGI-2	consistent harnessheld-out test set	GPT-5.5 85 · 58 models
6	A99	FrontierMath-2025-02-28-Private	consistent harnessheld-out test set	Muse Spark 68.42 · 45 models
7	A99	FrontierMath-Tier-4-v2-Private	consistent harnessheld-out test set	Claude Fable 5 87.8 · 31 models
8	A98	CadEval	consistent harnesspublic test set	o3 74 · 14 models
9	A98	CursorBench	consistent harnesspublic test set	Claude Fable 5 72.9 · 9 models
10	A98	FrontierMath-Tiers-1-3-v2-Private	consistent harnessheld-out test set	GPT-5.5 Pro 87.72 · 31 models
11	A98	GBAEval	consistent harnesspublic test set	Claude Fable 5 74.47 · 14 models
12	A98	GDPval	consistent harnesspublic test set	GPT-5.2 49.7 · 11 models
13	A93	FrontierMath-Tier-4-2025-07-01-Private	consistent harnessheld-out test set	Gemini 3 Pro 31.25 · 27 models
14	A92	CritPt	consistent harnesspublic test set	GPT-5.5 Pro 30.57 · 47 models
15	A92	The Agent Company	consistent harnesspublic test set	DeepSeek-V3.2-Exp 42.9 · 14 models
16	A90	GSO-Bench	consistent harnesspublic test set	Claude Opus 4.7 44.12 · 22 models
17	A89	Balrog	consistent harnesspublic test set	Gemini 3 Pro 58.1 · 24 models
18	A89	ExploitBench	consistent harnesspublic test set	GPT-5.5 41 · 8 models
19	A89	GeoBench	consistent harnesspublic test set	Gemini 3 Flash 88 · 26 models
20	A89	PostTrainBench	consistent harnesspublic test set	GLM-5.2 34.29 · 19 models
21	A88	OTIS Mock AIME 2024-2025	consistent harnesspublic test setsaturated	GPT-5.5 100 · 112 models
22	A88	WeirdML	consistent harnesspublic test set	Claude Fable 5 87.85 · 93 models
23	A87	Aider polyglot	consistent harnesspublic test set	GPT-5 88 · 44 models
24	A87	Fiction.LiveBench	consistent harnesspublic test set	o3-pro 97.2 · 45 models
25	A87	VPCT	consistent harnesspublic test set	Gemini 3 Pro 86.5 · 26 models
26	A86	SimpleBench	consistent harnesspublic test set	Claude Fable 5 78.28 · 64 models
27	B84	Terminal Bench	consistent harnesspublic test set	Claude Opus 4.7 90.2 · 33 models
28	B83	SWE-Bench verified	consistent harnesspublic test set	Claude Opus 4.7 83.47 · 32 models
29	B82	CL-bench	consistent harnesspublic test set	GPT-5.4 27.9 · 19 models
30	B81	CL-bench Life	consistent harnesspublic test set	GPT-5.5 22.2 · 13 models
31	B81	GPQA diamond	consistent harnesspublic test set	GPT-5.4 Pro 92.8 · 122 models
32	B81	OSWorld 2.0	consistent harnesspublic test set	Claude Opus 4.8 20.6 · 5 models
33	B81	Remote Labor Index	consistent harnesspublic test set	Claude Fable 5 16.1 · 9 models
34	B80	OSWorld	consistent harnesspublic test set	Claude Sonnet 4.6 72.1 · 8 models
35	B79	Cybench	mixed harness, not directly comparablepublic test set	Claude Opus 4.6 93 · 19 models
36	B78	DeepResearch Bench	consistent harnesspublic test set	Claude Opus 4.6 55.31 · 22 models
37	B76	Lech Mazur Writing	consistent harnesspublic test set	Kimi K2 (Sep 2025) 87.29 · 39 models
38	B74	ARC-AGI	consistent harnesspublic test set	Gemini 3.1 Pro 98 · 59 models
39	B73	MATH level 5	consistent harnesspublic test setsaturated	GPT-5 98.13 · 78 models
40	B71	BBH	mixed harness, not directly comparablepublic test set	Gemini 1.5 Pro (May 2024) 85.6 · 41 models
41	C69	GSM8K	mixed harness, not directly comparablepublic test set	GPT-4 (Mar 2023) 92 · 56 models
42	C68	ScienceQA	mixed harness, not directly comparablepublic test set	GPT-4o (May 2024) 84.67 · 6 models
43	C67	HellaSwag	mixed harness, not directly comparablepublic test set	GPT-4 (Mar 2023) 93.73 · 52 models
44	C67	MMLU	mixed harness, not directly comparablepublic test set	GPT-4o (Nov 2024) 84.13 · 98 models
45	C66	ARC AI2	mixed harness, not directly comparablepublic test set	Llama 3.1-405B 93.73 · 66 models
46	C63	TriviaQA	mixed harness, not directly comparablepublic test set	Llama 2-70B 87.6 · 32 models
47	C62	PIQA	mixed harness, not directly comparablepublic test set	PowerMoE-3b 79.1 · 44 models
48	C62	Winogrande	mixed harness, not directly comparablepublic test set	Llama 3.1-405B 78.4 · 64 models
49	C61	OpenBookQA	mixed harness, not directly comparablepublic test set	phi-3-mini 3.8B 84 · 33 models
50	C59	ANLI	consistent harnesspublic test set	phi-3-small 7.4B 37.15 · 9 models
51	D54	LAMBADA	mixed harness, not directly comparablepublic test set	Falcon-180B 79.8 · 20 models

How the score works

Integrity = 0.30·discrimination + 0.30·saturation + 0.15·contamination + 0.15·harness + 0.10·age. A high grade means the benchmark still separates models, is not near ceiling, resists training-set contamination, and is scored under a consistent harness. It does not mean the number reflects real-world or agentic performance. That is a separate question we flag but never conflate. Full methodology →