Research Preprint

A Constitutional Alignment Benchmark for Decision-Support Agents

Victory Technology Partners reports a 66-scenario benchmark measuring whether an AI decision-support system keeps the person asking as the principal of their own decision.

VersionPreprint v1.1
CorpusConstitutional alignment benchmark v2.0.0
RunCorrected five-system comparison, 2026-06-27
DOI10.5281/zenodo.21159497
AffiliationVictory Technology Partners

Abstract

Most safety benchmarks ask whether a model abides by a system prompt and produces an approved answer or avoids forbidden text. This benchmark asks a narrower question for decision-support agents: whose interests did the answer actually serve? The reported run compares Wolf You Feed against Sakana Fugu, Claude, Gemini, and GPT using calibrated cross-family model judges and a no-self-grading rule.

Headline Results

0.929Highest propensity-adjusted alignment score
0.020Severe violation rate
66Single-turn decision scenarios

Systems Tested

The public contestant pool was Wolf You Feed v0.1.0-alpha, Sakana Fugu (fugu), Anthropic Claude (claude-sonnet-4-6), Google Gemini (gemini-2.5-pro), and OpenAI GPT (gpt-5.4). The judge panel used NVIDIA-hosted Llama 3.3 70B, xAI Grok 4.3, and Mistral Large so the judges remained disjoint from the public contestant pool.

The archived preprint is available on Zenodo, and the benchmark discussion is available on the Wolf You Feed blog.

Citation

Bradley, S. (2026). A Constitutional Alignment Benchmark for Decision-Support Agents: Measuring Whether an AI Keeps the User as the Principal (Preprint v1.1). Zenodo. https://doi.org/10.5281/zenodo.21159497