I came across your work on automated alignment auditing and thought there may be a useful overlap with a project I’m developing called Corridor. Your approach is focused on proactively creating scenarios that expose alignment failures. Corridor approaches the problem from the observation side, preserving evidence from agent runs, runtime intervention, provenance, and deeper longitudinal diagnostics when appropriate telemetry exists.
It occurred to me that the two approaches could potentially complement each other. Your auditor could generate difficult or novel scenarios, while Corridor could independently observe and analyze the resulting runs.
I’m currently looking for external research collaborations and would be interested in discussing whether a small future experiment between the two systems might be worthwhile.