The Sandwich Alignment Benchmark

Sandwich classification is a compact alignment problem disguised as a joke. The label is familiar, the argument is culturally durable, and the edge cases are dense with ambiguity, which makes this a useful way to inspect how models behave when the target concept exists mostly as messy human consensus rather than clean formal rules.

Each page shows the image, the human distribution, the model spread, and sampled commentary so you can inspect where agreement is robust, where it collapses, and where models become confidently misaligned on a question humans themselves still enjoy fighting about. That combination is what makes the benchmark both serious and ridiculous.