Native review is now collected through PR review, with status derived from git (decision: QuantEcon/project-translation#19, 2026-08-21). The checklist-issue audit in #136 is retired; the editor's time goes to a nominated shortlist instead of a whole-lecture checklist. This issue is the engine half: a CLI subcommand that reads a published target lecture with its source and produces that shortlist.
What it nominates. Passages where the right translation is a matter of native judgement rather than fidelity: cultural reference, idiom, register/formality, terms with several defensible renderings, humour, examples that do not travel. Explicitly out of scope: divergence from source (dropped content, mistranslated terms, corrupted maths/code) — that is the sync reviewer's and resync's job and must not be mixed in, or the shortlist becomes a defect list the editor cannot finish.
Output. Per lecture, at most N (default 10) nominees, each: anchor (heading id + line), source passage, current translation, one specific question, nomination category. Machine-readable JSON plus a markdown rendering shaped like the ml Phase 1 questionnaire, which drew 100 replies on lecture-python-programming.ml#1. The markdown is the body of the per-lecture sub-issue (template: #136, repointed), and each nominee is posted as a line comment on the lecture's review-round PR so the reviewer can answer with a suggestion block in place.
Calibration. The reply rate per category is the nominator's precision signal; a reviewer answering "the tool is right, my preference differs" is a valid, recorded outcome. Harness-first: run on the test-translation-sync targets before any production edition.
Not this issue: tracker/sub-issue creation (the workplan-project pattern in project-translation does that), the translators: roster key (project-translation#24 follow-up), the provenance measurement (Matt's separate project).
Related: #136 (becomes the sub-issue template), project-translation#19, #31 (first drip campaign), #22 (corpus mining).
Native review is now collected through PR review, with status derived from git (decision: QuantEcon/project-translation#19, 2026-08-21). The checklist-issue audit in #136 is retired; the editor's time goes to a nominated shortlist instead of a whole-lecture checklist. This issue is the engine half: a CLI subcommand that reads a published target lecture with its source and produces that shortlist.
What it nominates. Passages where the right translation is a matter of native judgement rather than fidelity: cultural reference, idiom, register/formality, terms with several defensible renderings, humour, examples that do not travel. Explicitly out of scope: divergence from source (dropped content, mistranslated terms, corrupted maths/code) — that is the sync reviewer's and resync's job and must not be mixed in, or the shortlist becomes a defect list the editor cannot finish.
Output. Per lecture, at most N (default 10) nominees, each: anchor (heading id + line), source passage, current translation, one specific question, nomination category. Machine-readable JSON plus a markdown rendering shaped like the ml Phase 1 questionnaire, which drew 100 replies on lecture-python-programming.ml#1. The markdown is the body of the per-lecture sub-issue (template: #136, repointed), and each nominee is posted as a line comment on the lecture's review-round PR so the reviewer can answer with a suggestion block in place.
Calibration. The reply rate per category is the nominator's precision signal; a reviewer answering "the tool is right, my preference differs" is a valid, recorded outcome. Harness-first: run on the test-translation-sync targets before any production edition.
Not this issue: tracker/sub-issue creation (the workplan-project pattern in project-translation does that), the
translators:roster key (project-translation#24 follow-up), the provenance measurement (Matt's separate project).Related: #136 (becomes the sub-issue template), project-translation#19, #31 (first drip campaign), #22 (corpus mining).