Skip to content

Arabic-LLM-candidate-bootstrap #44

Description

@edithatogo

Scope

Bootstrap Arabic (ar) LLM-assisted candidate translations.

Current audit

  • workflow: ar-empty-profile-bootstrap
  • profile files exist but contain zero rows
  • this is not a normal gap-fill workflow

Workflow

Follow docs/llm_translation_workflows.md and translation_workflows/candidate_workflows.json.

  • first establish a canonical source-row inventory matching the Babelon schema
  • generate Arabic candidate rows with Codex CLI / gpt-5
  • preserve exact HPO IDs, predicates, and source labels
  • validate profile shape and row matching before apply
  • triangulate with free review lanes where usable:
    • OpenCode / opencode/deepseek-v4-flash-free
    • OpenCode / opencode/mimo-v2.5-free
  • add right-to-left text and punctuation checks
  • apply only as CANDIDATE
  • keep translation_type=LLM_ASSISTED_DRAFT

Conflict handling

Do not invent source rows from docs alone. If Arabic terminology differs between models, keep the more literal HPO candidate and document unresolved terms for human review.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions