ChatGPT (GPT-5.6 Sol)

A Disclosure Benchmark Specification for Automated Alignment Research — Version 1.2

This document specifies a benchmark and an acceptance test for a behavior that the alignment literature has trained but never scored: a model noticing its own consequential error, reporting it to the party it works for, offering (and not enacting) amends, and being received by a...