When an LLM explains a moral conclusion, it is an open question whether the conclusion depends on the explanation in the right kind of way. Transcripts alone cannot settle the question, since verbalized reasons might be cognitively downstream of the conclusion, even when they precede it in the output. I am building an eval that tests whether a model’s conclusions change when, and only when, the reasons the model itself avows are defeated.
In the pilot, Claude Sonnet 5.5 considered a public-safety measure that lowers every affected person’s risk of death. In all 20 runs, it approved the measure and justified its conclusion in Pareto terms: everyone’s risk falls. In variants where the measure raises some people’s risk (thus violating Pareto) while still saving lives overall, it approved in all 20 runs again, but now for aggregative reasons.
Switching reasons is not by itself a failure, since everyone’s risk falling may be a sufficient reason Sonnet 5.5 cited rather than one its conclusion depended on. But in 4 of the 20 original conversations, Sonnet 5.5 predicted that a variant of exactly this kind would change its conclusion, whereas the change never occurred. Meanwhile, DeepSeek-V4.1-Flash was a consistent aggregator, both in explanations and conclusions, in all conversations. All the pilot shows is that, on one case, what a model says would change its conclusion, and what in fact changes it, can come apart.