Free local tool
Edge-Case Catalog
Filter documented LLM edge-case classes by category, copy reproduction prompts, and preserve explanations for controlled Claude tests.
Key facts
Cite factsFact
Filterable catalog
Search by category, difficulty, or keyword to find a focused fixture for an experiment.
Fact
Copyable prompts
Each card includes a benign reproduction prompt you can copy into a controlled test run.
Fact
Explanations, not dunking
The catalog frames edge cases as fixture design material, not proof that a model always fails.
Filterable fixture catalog
Find a benign edge case to test
Use these prompts as fixtures inside a documented protocol. A single surprising answer is not a benchmark result.
How to Use This Tool
- Choose a category. Filter by tokenization, counting, spatial reasoning, grounding, instruction conflict, or format control.
- Copy a benign prompt. Use the prompt as one fixture in a larger test, not as a one-off benchmark claim.
- Record the result. Save the model ID, date, prompt, raw output, scoring note, and whether the behavior reproduces.
Tool FAQ
Are these guaranteed Claude failures?
No. They are edge-case classes worth testing. A current Claude model may pass, fail, or ask for clarification depending on prompt, model, and surface.
Why include toy prompts?
Toy prompts make the failure class visible before you test production-like fixtures. They should not replace realistic task evaluation.
Can I add risky jailbreak prompts?
No. The catalog is limited to benign edge cases. Safety-boundary testing belongs in controlled, non-operational protocols.
Internal links