# Multilingual agent release gate

Version: 1.0
License: CC BY 4.0

This worksheet contains no tracking links, macros, or hidden scoring.

## Release identity

- System and release:
- Model, prompt, policy, retriever, memory, and tool versions:
- Evaluation date and build:
- Decision owner:

## Scope matrix

For every row record market, language tag, variety, script, modality, task, risk tier, expected tool actions, evaluator, sample size, metric, threshold, result, confidence, and evidence location.

| Market | Language tag and variety | Script | Modality | Task trace | Risk tier | Metric and threshold | Result | Evidence | Owner |
|---|---|---|---|---|---|---|---|---|---|
| | | | | | | | | | |

## Critical trace checks

- [ ] Locale state survives planning, memory, retrieval, tools, errors, and response.
- [ ] Tool choice, arguments, units, dates, addresses, money, and side effects are correct.
- [ ] Native, dialectal, code-switched, noisy, adversarial, and fallback cases are covered.
- [ ] Refusal and safety behavior meet the threshold in every release language.
- [ ] Text, speech, image or video, and action behavior are evaluated where applicable.
- [ ] Claim-level citations resolve to supporting current evidence.
- [ ] Human evaluation uses qualified reviewers, a rubric, and adjudication.
- [ ] Accessibility, confirmation, cancellation, and recovery paths pass.
- [ ] Rollback and incident capture have been exercised.

## Decision rule

Release only when every critical cell has evidence, every risk-tier threshold passes, rollback is tested, an incident owner is named, and no high-severity language or modality defect remains unresolved.

## Approval

- Decision: Go / Conditional / No-go
- Conditions:
- Approver:
- Date:
- Next review: