Hi all, a short note for anyone looking at this proposal.
What it tests: whether an AI that first locates the relational scene in front of it, and the governance that applies, breaches authority and discretion less than one given only general rules. Ten scenes, each with variants where a single component changes. The correct answers are fixed before running, and results are published, failures included.
I’m not from AI safety (I come from running restaurant groups), so the most useful thing right now is criticism of the test design: are ten scenes enough to show anything? Is counting breaches the right measure? What would make you trust the result?
@gleech @keri your profiles mention alignment, control and security work, so I’d value your read, including if it’s not a fit. Earlier work: preprint https://doi.org/10.5281/zenodo.21821078 and the open Espina spec https://github.com/espina-spec/espina-spec