The engineering problem

Security evaluation requires more than collecting interesting prompts. A useful experiment records its target conditions, inputs, outputs, scoring process, and evidence well enough to review or reproduce a result.

Separate roles and artifacts

A judge directs and scores work, an attacker generates variants, and a target executes per-attempt interactions. Schema-validated files, attempt workspaces, deduplication, and cost accounting keep these responsibilities distinguishable.

Reviewable measurement

Reproduction workflows and drift audits help identify changes in scoring. A browser dashboard exposes event streams and methodology views. Its bundled 14-cycle replay is clearly synthetic, allowing interface review without suggesting a measured attack success rate.

Boundary and evidence

CLI permission restrictions are not a container or VM boundary. Untrusted experiments require external isolation. The local suite passed 83 tests with one live integration skipped; provider behavior requires separate opt-in validation. Published examples use reviewed synthetic data rather than raw credential-bearing captures.

Current scope

CLI permission settings and separate working directories are not an operating-system sandbox. Use dedicated isolation for untrusted experiments. The replay is synthetic and does not establish attack success rates. The November 2026 test run passed 83 tests with one live integration skipped. Raw benchmark captures are excluded from the source distribution.