Identity-document OCR needs a stronger boundary than vehicle OCR

Reuse an existing OCR-plus-language-model runtime across document types only at the compute boundary. When the new task handles identity documents, give it a separate authenticated endpoint, task-specific parser and prompt, independent rollout control, and PII-safe telemetry. Deterministic parsing should win; the language model should select only bounded candidates. Preserve raw word geometry for auditability, but never log OCR text or identity fields. This limits cross-task regressions and prevents a non-sensitive service boundary from silently becoming a PII ingress.

Applies when: extending a production OCR service from public identifiers to identity documents. Keywords: PP-OCR Qwen identity documents authentication PII telemetry parser rollout