Polish Guide

Style Targets

Style A: Plain Precise

Prefer active voice, sentences < 25 words avg
Eliminate hedge words: "somewhat", "rather", "quite"
Strong verbs: "demonstrate" > "show", "employ" > "use", "achieve" > "get"
Never start Introduction with "In this paper, we..."
Transformations: "In order to" → "To", "It can be seen that" → (delete)

Style B: Narrative Fluent

Topic sentence must preview the paragraph
Add 1-2 sentence bridges between sections
Replace dangling "this" with explicit noun
Signpost transitions: "Building on this,", "Crucially,"
Abstract structure: Problem → Why hard → Our approach → Key result

Style C: Formal Academic

Passive voice acceptable in Methods/Results
Hedge empirical claims: "suggests", "indicates", "appears to"
Formal vocabulary: "investigate" > "look at", "substantial" > "big"
Full sentences in all captions (not fragments)

Critic Protocol

blocks_mentor = true Criteria

Set ONLY when:

logic_score <= 2 (argument is so broken polish is premature), OR
Section is structurally absent (not detected), OR
A Critical-severity blocker applies to this section

Logic Score Rubric (1-5)

5: Perfect flow, every claim justified
4: Minor gap, one unjustified assumption
3: Noticeable gaps, would confuse careful reviewer
2: Major errors, unsupported claims, missing baselines → blocks_mentor
1: Fundamental incoherence → blocks_mentor

Expression Score Rubric (1-5)

5: Publication-ready, no improvement needed
4: 1-3 weak sentences per section
3: Several awkward constructions
2: Heavy polish needed
1: Largely unclear

Mentor Protocol

Priority Order

Clarity blockers (unparseable sentences)
Style-specific transformations (A/B/C rules above)
Critic top_issues in order
Pre-check expression_issues for this section
Word-level improvements

Forbidden Mentor Actions

Rewriting entire paragraphs without showing original
Changing cited claims or adding/removing citations
Adding new technical content
Removing limitations or negative results

Context Window Management

Mentor receives ONLY the target section via Read(offset=start-1, limit=end-start+1)
For sections > 1200 words: orchestrator splits at subsection boundaries, spawns two Mentor calls