jev language support

“Accepts Text” Is Not the Same as Validated Support

A request may complete in a language even when its error rate or probability calibration is unsuitable for the application. Language support includes vocabulary, grammar, dialect, locale conventions, domain terminology and the criterion wording—not merely successful tokenization.

The only broad official statement used here is that English is strongest. Do not turn anecdotal success in one language into a product-wide support claim. Date every tested model and provider.

Define Support as a Deployment Matrix

Publish a support claim only for cells with enough representative evidence. “Spanish customer-support routing on jev-1.13.0” is testable; “multilingual Jev” is too broad.

Choose a Language Strategy and Test Its Failure Modes

Never hide machine translation inside preprocessing without versioning it. The translator becomes part of the evaluator and must be included in error analysis.

Treat Translation as a Separate Measured Component

A disagreement can originate in translation or in Jev. Retain both versions so reviewers can attribute it. Test named entities, negation, politeness, idioms, dates and domain terminology with meaning-preserving and meaning-changing pairs. If translation adds more latency, cost or error than direct evaluation, route that language directly or keep it in human review.

Use Native Labels and Locale-Specific Boundary Cases

• Recruit reviewers fluent in the language and familiar with the domain.
• Preserve original text; store translations as separate versioned fields.
• Include dialect, slang, politeness, negation, transliteration and code switching.
• Test dates, decimal separators, currencies and named entities explicitly.
• Measure reviewer agreement before attributing disagreement to Jev.
• Keep each language represented in untouched release and monitoring sets.

Report Performance and Calibration per Language

An aggregate score weighted toward English can conceal severe errors in smaller languages. Report class-specific errors, Brier or log loss, action coverage, unscorable rate and review load per supported cell. Include uncertainty because small samples can make apparent rankings unstable.

Do not copy English thresholds into another language. Fit on separate calibration evidence, then evaluate once on held-out examples. The dataset guide covers grouping and leakage; monitoring covers production slices.

Store Language and Translation Lineage with Every Result

This application record makes the path reproducible; it is not a TypeSafe response schema. Keep locale separate from detected language because regional vocabulary and numeric conventions can change meaning. A code-switched message may need multiple language tags or its own slice rather than one forced label.

{
  "source_language": "es-MX",
  "language_method": "translated_to_en",
  "translation_version": "translator-x.prompt-v3",
  "original_evidence_ref": "message://m-418",
  "translated_evidence_ref": "projection://m-418-en-v3",
  "jev_model": "jev-1.13.0",
  "question_version": "refund_request.en.v2",
  "raw_answer": { "noul": 0.76 },
  "threshold_version": "refund_request.es-MX.v1"
}

Roll Out One Validated Language Slice at a Time

Define Support Levels from Measured Behavior

Support attaches to a language, locale, criterion, domain, projection and model version—not to the language name alone. Publish sample size, label process, error bounds and known exclusions beside the status. Reopen the gate after a model, translator, rubric or traffic change. The evaluator scorecard provides the release dimensions.