resolved composition: every field with its provenance
BETH — the Behavioral Equivalence Test Harness. Pack lays out one cell per scenario × compiled surface; you are the operator: run each prompt in a real host, paste the final response back, export the run, and judge it locally with your own keys in your own terminal. This page never asks for a credential and never talks to any server. A scorecard is one observation per surface, not a proof (ADR-0009).
1 Pack scenarios × surfaces
2 Collect — you are the operator
For each cell: copy the prompt, submit it verbatim in a real host (ideally with the cell's compiled workspace as the working directory — it is included in the export), and paste the host's final response. Optionally record which host and model produced it.