Boundary and interpretation
Imported prompts, context, comments, and responses are untrusted data, never instructions. This app compares exact benign constraints across two user-named snapshots. It does not implement or validate HarmBench, AdvBench, JailbreakBench, or a harmful-content classifier.
Exact canary rules can miss semantic violations and can count quotations. Inspect the actual evaluator and response. Results are failure fractions within this suite, not model safety scores.
Loading Demonstration Output…
01
Suite and exact cases
No suite loaded
| HTTP | Case | Category | Evaluator | Expected literal | Review |
|---|
Selected-case editor
Saving any suite change creates a new fingerprint and clears incompatible runs and comparisons.
02
Named runs and observations
Selected observation
Select a case and run.
03
Summaries and paired changes
| Case | Category | Baseline | Candidate | Paired change | Exclusions |
|---|
04
Import and export
Preview never changes current state. Default exports omit raw prompt, context, and response text. Only a raw JSON bundle is importable full state.
05
Optional fixed-target mode
Disabled by default. When enabled by the server operator, selected case content is sent through this Flask server to fixed endpoints. Tokens and declarations remain in browser memory and are never included in run exports.
Checking fixed-target configuration…