Structured code-switched synthetic dialogues¶

generate_code_switch_dialogue generates authored semantic turn sequences with patient, caregiver, interpreter and clinician roles. Each DialogueTurnSpec selects explicit English/Spanish/Hindi segments, an authored fictional clinical concept, assertion status, experiencer role and optional correction reference. The generator does not concatenate unrelated random strings: every switched segment preserves the same declared clinical concept/status and role context.

Existing locale-PHI generators provide seeded synthetic PERSON spans, consistent for the same experiencer role and locale. Every span is checked against shared extended-grapheme boundaries and exact generated text. Language-tagged ranges retain within-turn switch points. Existing Journey ClinicalFact, ClinicalArtifact and EvidenceLocator records bind clinical gold annotations to actual statement offsets; no alternate clinical truth graph is introduced.

Contradictory status declarations for the same concept/experiencer flag all involved turns for adjudication. Uncertainty also requires review. Corrections must point backward to the same clinical identity and retain parent fact/evidence links; they never erase the earlier assertion or auto-resolve a conflict. Reports expose counts, controlled roles/languages and artifact hashes without names or transcript content. Training exports contain only generated synthetic turns, spans and declared facts.

The three language templates are limited authored demonstrations, not native review or clinical linguistic qualification. Concepts are fictional software stimuli, not medication advice or patient records. No translation model, remote service, real dataset or autonomous clinical decision is involved. PHI generation is reused rather than replacing deterministic privacy checks. Synthetic origin and permissive template rights do not establish model quality or privacy.