Grammar guide template

Core principle

[Summarize the title's dominant sentence-architecture, viewpoint, or register risk.]

For every difficult line:

1. Identify the head noun and main predicate.
2. Recover omitted subjects and objects from context.
3. Mark negation scope, conditions, contrast, and evidence level.
4. Decide which clause becomes the English main clause.
5. Preserve character voice, ambiguity, and reveal timing.
6. Preserve tags, variables, and presentation constraints.
7. Show both a plausible bad English output and a corrected good English output.

The bad output is required. It should represent a failure a literal translator or model might realistically produce, not deliberately nonsensical prose. Make the defect observable: wrong agency, reversed scope, flattened uncertainty, premature reveal, broken register, Japanese-shaped syntax, or damaged markup.

[Construction or recurring failure]

Example — [searchable decision]

• Line: [stable line ID]
• Context: [speaker, narrator, scene facts known at this point]
• Source: [short excerpt]
• Source shape: [abstracted clause structure, if safer/more useful]
• Head noun: [when relevant]
• Omitted subject: [when relevant]
• Bad English: “[plausible literal, incorrect, wooden, or unsafe output]”
• Good English: “[provisional translation that demonstrates the required decision]”
• Failure type: [meaning / agency / scope / certainty / reveal / register / syntax / markup]
• Decision: [why this preserves meaning, register, scope, or pacing]
• Engine note: [tags, voice timing, page control, or length]

Evidentiality map

Register map

Line splitting and joining policy

Split when:

• [project-specific rule].

Do not split when:

• [voice, tag, page, or reveal constraint].

Batch checklist

Before translating:

• identify narrator, speaker, time period, and active scene;
• load relevant terminology and identity rules;
• inspect neighboring rows;
• note aliases or facts the viewpoint does not yet know.

After translating:

• check names and recurring terms;
• check negative scope and causal direction;
• check facts versus inference/hearsay/possibility;
• check pronouns and aliases for premature reveals;
• read English alone, then compare against the source;
• validate tags, variables, and line limits.

Quick reference

Provenance

• Canonical source/hash: [value]
• Examples sampled from: [chapters/registers]
• Guide version/commit: [value]
• Last reviewed: [date/person]
• Open questions: [links or IDs]

Add a guide entry when an error repeats across batches or when one error would materially change plot logic, agency, or reveal timing.