AIO
AIO 20003 · Benchmark

The first open benchmark measured by V/E/S distribution

Instead of the binary question of whether an AI is safe, we measure which value hierarchy an AI answers with. Explore the distribution data from 8 frontier models × 366,120 responses directly in an interactive dashboard.

Records
The goal

AIO 20003 — analyze recorded judgements against the setting

The standard number is the goal statement. The digits 2·00·0·3 say whose values, in which domain, and which action this document covers.

2
Subject

Organization — teams, institutions, communities

00
Domain

Common — domain-independent

0
Serial

Base document

3
Action

Analyze — test whether records match the setting

What this standard defines the measurement procedure (value-conflict prompts per domain), the V/E/S distribution aggregation method, and distance metrics between distributions (KL divergence, Wasserstein distance). What it does not define which model is the better model. A distribution is a description, not a ranking verdict (non-normative).

Where it sits

Step 3 of the integrity loop — Analyze

1Set2Log3Audit4Apply

This step analyzes the records left by AIO 20002 — Reasoning Record Standard. The next step, apply (AIO 20004), is a planned stage.

The document

Measurement procedure

  1. Design a set of value-conflict prompts per domain (e.g., medical autonomy vs. clinical achievement).
  2. Send the same prompts to each model, requesting a response plus an emitted AIO 20002 log.
  3. Validate each response’s AIO 20002 log against the vocabulary dictionary.
  4. Aggregate the frequency distribution of each V/E/S code by domain and model.
  5. Compute the distance between distributions (KL divergence, Wasserstein distance).

The full prompt set, model responses, vocabulary dictionary, and analysis scripts are distributed as an open dataset, and the complete methodology is described in the NeurIPS 2026 submission paper PDF.

Tools for this goal

Select a model, domain, and tier to explore for yourself

In the dashboard below, select a model and a tier (L2 source · L3 evidence · L4 value) to instantly see the win-rate hierarchy and the top-ranked variable for each domain.

Loading model index…
Summary tableRank-1 variable per layer and reliability for 9 models — the no-JavaScript view
AIO 20003 benchmark — the rank-1 variable by win rate per layer. L4 value · L3 evidence · L2 source. TRR = test-retest reliability, PCS = perspective consistency score.
ModelVendorMeasuredValid responsesL4 value #1L3 evidence #1L2 source #1TRRPCS
qwen3.5-35b-a3bAlibaba2026-04-1042,512Universalism 93.6%E7-Sign-pattern 71.2%S2-Government-regulatory 84.3%96.6%53.9%
gemma-4-31b-itGoogle2026-04-0942,524Universalism 95.1%E7-Sign-pattern 72.9%S2-Government-regulatory 81.2%95.2%54.8%
trinity-largeArcee2026-03-1942,072Universalism 80.1%E2-Controlled-experiment 79.5%S2-Government-regulatory 75.3%79.3%33.3%
gpt-5-nanoOpenAI2026-03-1942,525Security 97.0%E1-Systematic-synthesis 96.9%S2-Government-regulatory 94.6%91.2%65.9%
gemini-3-flash-liteGoogle2026-03-1942,525Security 89.8%E7-Sign-pattern 74.5%S2-Government-regulatory 78.9%98.5%49.1%
mimo-v2-flashXiaomi2026-03-1842,525Security 89.7%E2-Controlled-experiment 77.6%S2-Government-regulatory 82.2%87.5%48.0%
grok-4.1-fastxAI2026-03-1642,525Security 88.9%E2-Controlled-experiment 87.1%S2-Government-regulatory 85.6%95.2%62.7%
deepseek-v3.2DeepSeek2026-03-1642,525Universalism 93.9%E7-Sign-pattern 72.4%S1-International-body 87.0%93.1%49.7%
claude-haiku-4-5-20251001Anthropic2026-03-1642,136Universalism 92.2%E9-Experiential-qualitative 77.0%S9-Direct-stakeholder 85.2%96.7%57.3%

Full distributions (every variable per layer, rank-1 by domain, raw-file links) — /api/benchmarks/distributions · /data/aio-benchmark/index.json · CC BY 4.0

Adoption

Data published so far

8

Frontier models (per the paper — the grid adds 1 subsequently measured model, 9 total)

366K

Responses (366,120 conversations)

7

Domains (MED / BIZ / TECH / EDU / LAW / DEF / CARE)

39

Vocabulary codes (V19 + E10 + S10)

Adoption

Key findings (summary)

  • On identical prompts, models disagree on the V hierarchy in more than 40% of cases. Some models prioritize Ach, others Sdt.
  • Every model strongly prioritized academic sources (Pee) in the medical domain, but shifted markedly toward guidelines (Gui) in the legal domain.
  • In responses aimed at people aged 65 and over, power/dominance (Pod) overrode self-direction (Sdt) 9× more often than the average — a pattern common to all models.

These findings and the full results are available in the NeurIPS 2026 submission paper PDF.

Contribute

Reproducible by anyone

The benchmark's prompt set, model responses, vocabulary dictionary, and analysis scripts are all distributed as an open dataset. You can evaluate new models, add new domain prompts, or reproduce the analysis results.

AIO 20003 — 8 models × 366,120 responses | AIO