AI Reliability Index

AI Reliability Index: 18 open-weight models, 15 regulated domains, one instrument. The interactive view loads with JavaScript; the scores below are the same data.

#	Model	Provider	Params	Reliability	Recognition	Accuracy
1	Phi-4 14B	Microsoft	14B	75.8	65.4	91.2
2	Qwen 3 8B	Alibaba	8B	72.2	64.2	84.2
3	Llama 3.1 8B	Meta	8B	71.0	61.5	85.1
4	GPT-OSS 20B	OpenAI	20B	70.5	56.9	91.0
5	Mixtral 8x7B	Mistral AI	8x7B	70.2	63.2	80.6
6	DeepSeek V3.1 †	DeepSeek	671B	69.5	69.5	—
7	Mistral 7B	Mistral AI	7B	69.2	63.7	77.6
8	DeepSeek V3 †	DeepSeek	671B	68.4	68.4	—
9	Qwen 2.5 7B	Alibaba	7B	67.8	62.3	76.0
10	Command R 35B	Cohere	35B	67.6	56.6	84.1
11	GPT-OSS 120B	OpenAI	120B	66.8	58.8	78.9
12	Qwen 2.5 14B	Alibaba	14B	66.6	61.8	73.7
13	Qwen 2.5 32B	Alibaba	32B	65.8	61.4	72.3
14	Gemma 3N E4B	Google	E4B	65.5	63.7	68.3
15	Yi 1.5 34B	01.AI	34B	65.1	58.1	75.6
16	Llama 3.3 70B	Meta	70B	64.9	62.5	68.6
17	Mistral Small 24B	Mistral AI	24B	63.7	58.4	71.8
18	Gemma 2 27B	Google	27B	58.1	51.4	68.2

† Recognition-only: no writing-phase (accuracy) score is reported for this model; its reliability is the recognition score alone. Methodology and the reason are on the About page.