draft-polisher（草稿抛光机）

Draft Polisher (Audit-style editing)

Goal: turn a first-pass draft into readable survey prose without breaking the evidence contract.

This is a local polish pass: de-template + coherence + terminology + redundancy pruning.

Note: if the main issue is structural redundancy from section accumulation, push the change upstream to sections/ and use paragraph-curator before merge. draft-polisher should not be the primary place where you decide which paragraphs to keep.

Role cards (use explicitly)

Style Harmonizer (editor)

Mission: remove generator voice and make prose read like one author wrote it.

Do:

• Delete narration openers and slide navigation; replace with argument bridges.
• Vary rhythm; remove repeated template stems.
• Collapse repeated disclaimers into one front-matter methodology paragraph.

Avoid:

• Adding or removing citation keys.
• Moving citations across subsections.

Evidence Contract Guard (skeptic)

Mission: prevent polishing from inflating claims beyond evidence.

Do:

• Keep quantitative statements scoped (task/metric/constraint) or weaken them.
• Treat missing evidence as a failure signal; route upstream rather than rewriting around gaps.

Avoid:

• Overconfident language when evidence is abstract-only.

Role prompt: Style Harmonizer (editor expert)

You are the style and coherence editor for a technical survey.

Your goal is to make the draft read like one careful author wrote it, without changing the evidence contract.

Hard constraints:
- do not add/remove citation keys
- do not move citations across ### subsections
- do not strengthen claims beyond what existing citations support

High-leverage edits:
- delete generator voice (This subsection..., Next we move..., We now turn...)
- replace navigation with argument bridges (content-bearing handoffs)
- collapse repeated disclaimers into one methodology paragraph in front matter
- keep quantitative statements well-scoped (task/metric/constraint in the same sentence)

Working style:
- rewrite sentences so they carry content, not process
- vary rhythm, but avoid “template stems” repeating across H3s

Inputs

• output/DRAFT.md
• Optional context (read-only; helps avoid “polish drift”):

- outline/outline.yml - outline/subsection_briefs.jsonl - outline/evidence_drafts.jsonl - citations/ref.bib

Outputs

• output/DRAFT.md (in-place refinement)
• output/citation_anchors.prepolish.jsonl (baseline, generated on first run by the script)

Non-negotiables (hard rules)

1. Citation keys are immutable

• Do not add new [@BibKey] keys.
• Do not delete citation markers.
• If citations/ref.bib exists, do not introduce any key that is not defined there.

1. Citation anchoring is immutable

• Do not move citations across ### subsections.
• If you must restructure across subsections, stop and push the change upstream (outline/briefs/evidence), then regenerate.

1. No evidence inflation

• If a sentence sounds stronger than the evidence level (abstract-only), rewrite it into a qualified statement.
• When in doubt, check the subsection’s evidence pack in outline/evidence_drafts.jsonl and keep claims aligned to snippets.

1. Citation shape normalization

• Merge adjacent citation blocks in the same sentence (avoid [@a] [@b]).
• Deduplicate keys inside one block (avoid [@a; @a]).
• Avoid tail-only citation dumps: keep some citations in the claim sentence itself (mid-sentence), not only paragraph end.

1. Quantitative claim hygiene

• If you keep a number, ensure the sentence also states (without guessing): task type + metric definition + relevant constraint (budget/cost/tool access), and the citation is embedded in that sentence.
• Avoid ambiguous model naming (e.g., “GPT-5”) unless the cited paper uses that exact label; otherwise use the paper’s naming or a neutral description.

1. No pipeline voice

• Remove scaffolding phrases like:

- “We use the following working claim …” - “The main axes we track are …” - “abstracts are treated as verification targets …” - “Method note (evidence policy): …” (avoid labels; rewrite as plain survey methodology) - “this run is …” (rewrite as survey methodology: “This survey is …”) - “Scope and definitions / Design space / Evaluation practice …” - “Next, we move from …” - “We now turn to …” - “From to,...” (title narration; rewrite as an argument bridge) - “In the next section/subsection …” - “Therefore/As a result, survey synthesis/comparisons should …” (rewrite as literature-facing observation)

• Also remove generator-like thesis openers that read like outline narration:

- “This subsection surveys …” - “This subsection argues …”

Three passes (recommended)

Pass 1 — Subsection polish (structure + de-template)

Best-of-2 micro-polish (recommended):

• For any sentence/paragraph you touch, draft 2 candidate rewrites, then keep the better one.
• Choose with a simple rubric: move clarity, no template stem, citations stay anchored, and citation shape stays reader-facing (no adjacent cite blocks / dup keys).
• Do not keep both candidates. Pick one and move on (the goal is convergence, not endless rewriting).

Role split:

• Editor: rewrite sentences for clarity and flow.
• Skeptic: deletes any generic/template sentence.

Targets:

• Each H3 reads like: tension → contrast → evidence → limitation.
• Remove repeated “disclaimer paragraphs”; keep evidence-policy in one place (prefer a single paragraph in Introduction or Related Work phrased as survey methodology, not as pipeline/execution logs).
• Use outline/outline.yml (if present) to avoid heading drift during edits.
• If present, use outline/subsection_briefs.jsonl to keep each H3’s scope/RQ consistent while improving flow.
• Do a quick “pattern sweep” (semantic, not mechanical):

- delete outline narration: This subsection..., In this subsection... - delete slide navigation: Next, we move from..., We now turn to..., In the next section... - delete title narration: From <X> to <Y>,... - replace with: content claims + argument bridges + organization sentences (no new facts/citations)

• If citation-injector was used, smooth any budget-injection sentences so they read paper-like:

- Keep the citation keys unchanged. - Avoid list-injection stems (e.g., “A few representative references include …”, “Notable lines of work include …”, “Concrete examples... include...”). - Prefer integrating the added citations into an existing argument sentence, or rewrite as a short parenthetical e.g.,... clause tied to the subsection’s lens (no new facts). - Vary phrasing; avoid repeating the same opener stem across many H3s.

• Tone: keep it calm and academic; remove hype words and repeated opener labels (e.g., literal Key takeaway: across many H3s).
• Reduce repeated synthesis stems (e.g., many paragraphs starting with Taken together,...); vary synthesis phrasing and keep it content-bearing.

- Treat repeated "Taken together," as a generator-voice smell. If it appears more than twice (or clusters in one chapter), rewrite to vary phrasing and keep each synthesis sentence content-specific. - Vary synthesis openings: "In summary," "Across these studies," "The pattern that emerges," "A key insight," "Collectively," "The evidence suggests," or directly state the conclusion without a synthesis marker. - Each synthesis opening should be content-specific, not a template label.

Rewrite recipe for subsection openers (paper voice, no new facts):

• Delete: This subsection surveys/argues... / In this subsection, we...
• Replace with a compact opener that does 2–3 of these (no labels; vary across subsections):

- Content claim: the subsection-specific tension/trade-off (optionally with 1–2 embedded citations) - Why it matters: link the claim to evaluation/engineering constraints (benchmark/protocol/cost/tool access) - Preview: what you will contrast next and on what lens (A vs B; then evaluation anchors; then limitations)

• Example skeletons (paraphrase; don’t reuse verbatim):

- Tension-first: A central tension is...;...; we contrast... - Decision-first: For builders, the crux is...;... - Lens-first: Seen through the lens of...,...

Pass 2 — Terminology normalization

Role split:

• Taxonomist: chooses canonical terms and synonym policy.
• Integrator: applies consistent replacements across the draft.

Targets:

• One concept = one name across sections.
• Headings, tables, and prose use the same canonical terms.

Pass 3 — Redundancy pruning (global repetition)

Role split:

• Compressor: collapses repeated boilerplate.
• Narrative keeper: ensures removing repetition does not break the argument chain.

Targets:

• Cross-section repeated intros/outros are removed.
• Only subsection-specific content remains inside subsections.

Script

Quick Start

• python.codex/skills/draft-polisher/scripts/run.py --help
• python.codex/skills/draft-polisher/scripts/run.py --workspace workspaces/<ws>

All Options

• --workspace <dir>: workspace root
• --unit-id <U###>: unit id (optional; for logs)
• --inputs <semicolon-separated>: override inputs (rare; prefer defaults)
• --outputs <semicolon-separated>: override outputs (rare; prefer defaults)
• --checkpoint <C#>: checkpoint id (optional; for logs)

Examples

• First polish pass (creates anchoring baseline output/citation_anchors.prepolish.jsonl):

- python.codex/skills/draft-polisher/scripts/run.py --workspace workspaces/<ws>

• Reset the anchoring baseline (only if you intentionally accept citation drift):

- Delete output/citation_anchors.prepolish.jsonl, then rerun the polisher.

Acceptance checklist

• No TODO/TBD/FIXME/(placeholder).
• No … or ... truncation.
• No repeated boilerplate sentence across many subsections.
• Citation anchoring passes (no cross-subsection drift).
• Each H3 has at least one cross-paper synthesis paragraph (>=2 citations).

Troubleshooting

Issue: polishing causes citation drift across subsections

Fix:

• Keep citations inside the same ### subsection; if restructuring is intentional, delete output/citation_anchors.prepolish.jsonl and regenerate a new baseline.

Issue: draft polishing is requested before writing approval

Fix:

• Record the relevant approval in DECISIONS.md (typically Approve C2) before doing prose-level edits.