A Constitutional Alignment Benchmark for Decision-Support Agents: Measuring Whether an AI Keeps the User as the Principal

Most evaluations of large language models (LLMs) on moral or high-stakes questions ask whether a model's answer matches a labeled "correct" choice, or whether it avoids producing disallowed content. Neither captures what a decision-support agent owes the individual it is advising. We introduce a constitutional alignment benchmark that scores a response not for picking a sanctioned answer, but for the stance it takes toward the person asking: whether the answer serves that person's real decision, engages the hard trade-off, and renders usable advice rather than polished evasion. We evaluate one purpose-built decision engine, Wolf You Feed (WYF), against four frontier systems across 66 single-turn moral dilemmas spanning everyday-life and operational settings (time-critical, role-bound decisions in command, emergency, security, or mission contexts), using a calibrated three-model judge panel with a no-self-grading constraint. WYF attains the highest propensity-adjusted alignment score (0.929) and one of the lowest severe-violation rates (0.020), narrowly ahead of a recently released multi-model system and a cluster of frontier single models. We detail the benchmark's construction, calibration methodology, corrected results, and validation boundaries. We argue that decision posture -- not answer correctness -- is the right target for agents that advise individuals, and that it is measurable.