S3: Agentic RAG

Claude Sonnet 4.6 with autonomous tool use, self-reflection, and critic scoring via OpenRouter.

1Reason → 2Tools → 3Translate → 4Validate → 5Reflect → 6Critic → ↻Revise?

lookup_glossary lookup_translation_memory lookup_grammar_pattern validate_locale

COMET 0.9284

Term Acc 0.6092

Coverage 0.83

Source — English

Output — Japanese

Coverage

critic score

Tool Calls

glossary + tm + grammar + locale

Latency

end to end

Model

Sonnet

claude 4.6