1,234.50 and 1.234,50 can describe the same amount. Lowercasing I gives different results in English and Turkish. A fluent response can still violate the structured output your application expects.
LocaleCrash is a ufodia regression workbench for making those details explicit. Its library contains 36 English/Turkish pairs across numbers, calendar dates, dotted and dotless I, and negative commands.
Inspect the contract
Open the case library and inspect a deliberately naive parser failure. Compare the prompt, expected JSON, actual output, and field-level mismatch.
Choose Run reference parser to see how a finite locale-aware implementation handles the included fixtures.
Export the prompt pack, run it through your own external model or agent harness, and import the documented response file. Review failures and missing answers, then export a JSON or CSV report.
The included workbench and parsers run locally without an account or API key. LocaleCrash does not automatically call a model or upload your responses.
Read the score narrowly
The naive and reference parsers are local demonstrations, not LLMs. Their scores are not model rankings. Scoring uses strict JSON equality: types and extra fields matter. These small synthetic fixtures do not measure general language ability, safety, or real-application reliability. Imported labels are user supplied. Save exports before refreshing.
Open LocaleCrash · Inspect the MIT-licensed source
If this helps you catch a locale regression, one optional coffee supports ufodia's fixtures, examples, and adapters.
Real LocaleCrash workspace showing synthetic English/Turkish fixtures and the deliberately naive local parser. Displayed pass/fail counts are demonstration results, not an LLM ranking.
