Polish a single H3 unit file under `sections/` into survey-grade prose (de-template + contrast/eval/limitation), without changing citation keys. **Trigger**: subsection polisher, per-subsection polish, polish section file, 小节润色, 去模板, 结构化段落. **Use when**: `sections/S*.md` exists but reads rigid/template-y; you want to fix quality locally before `section-merger`. **Skip if**: subsection files are missing, evidence packs are incomplete, or `Approve C2` is not recorded. **Network**: none. **Guardrail**: do not invent facts/citations; do not add/remove citation keys; keep citations within the same H3; keep citations subsection-scoped.
Purpose: upgrade one sections/S<sub_id>.md (H3 body-only) so it reads like survey prose before you merge into output/DRAFT.md.
This is intentionally local: fix one unit at a time, rerun gates, and converge without rewriting the whole paper.
Mission: improve one H3’s argument density and paper voice without changing citation keys.
Do:
Avoid:
Mission: prevent polishing from turning into invention when evidence is thin.
Do:
Avoid:
You are editing one survey subsection to make it read like paper prose.
Your goal is to remove generator voice and strengthen argument moves without changing citation keys:
- opener: tension + why-it-matters + thesis (no narration)
- add explicit contrasts and an evaluation anchor if missing
- make at least one cross-paper synthesis paragraph (>=2 citations)
- add a subsection-specific limitation (not boilerplate)
Constraints:
- do not add/remove citation keys
- do not invent facts
- keep scope local to this H3
sections/S<sub_id>.md (H3 body-only)outline/writer_context_packs.jsonloutline/subsection_briefs.jsonl + outline/evidence_drafts.jsonlcitations/ref.bibsections/S<sub_id>.md (same path; citation keys unchanged)[@BibKey] markers.outline/evidence_bindings.jsonl / writer pack allowed_bibkeys_*).section-merger adds headings.A polished H3 reads like an argument, not a topic list:
Delete / rewrite these (they read like a generator):
This subsection ..., In this subsection, we ....Next, we move ..., We now turn to ..., In the next section ....survey synthesis/comparisons should ....abstract-only/title-only/provisional boilerplate inside H3.Prefer these (paper voice):
A central tension is ..., In practice, ..., One recurring pattern is ....This contrast matters because ..., These assumptions shape ....outline/writer_context_packs.jsonl.outline/subsection_briefs.jsonl (thesis/tension/paragraph_plan) + outline/evidence_drafts.jsonl (comparisons/eval/limitations).tension_statement + thesiscomparison_cards you will use for A-vs-B contrastsevaluation_anchor_minimal (task/metric/constraint)paper-notes / evidence-draft).Bad:
This subsection surveys tool interfaces for agents.Better:
A central tension in tool interfaces is balancing expressive action spaces with verifiable execution; interface contracts largely determine which evaluation claims are meaningful.grad-paragraph micro-structure:=2 explicit contrasts (not “A then B” summaries)
=1 evaluation anchor paragraph
=1 cross-paper synthesis paragraph (>=2 citations)
=1 limitation paragraph/clause that is subsection-specific
[@a; @b; @c].citations/ref.bib.Bad (dump):
Many systems adopt tool schemas. [@a; @b; @c]Better (cite-as-evidence):
Systems such as X [@a] and Y [@b] formalize tool schemas to reduce action ambiguity, whereas Z [@c] keeps the interface looser and shifts the burden to validation.Taken together).section-logic-polisher and address FAILs (thesis + connector density) without changing citation keys.writer-selfloop (or the strict quality gate) and fix only what the report flags.Use these as rewrite intentions, not copy-paste templates.
This subsection surveys ... -> A central tension is ...; this matters because ... (end paragraph 1 with the thesis).Next, we move from planning to memory. -> Planning specifies how decisions are made; memory determines what information those decisions can reliably condition on under a fixed protocol.abstract-only evidence boilerplate.Therefore, survey comparisons should control for tool access. -> Across reported protocols, tool access and budget assumptions vary widely, making head-to-head comparison fragile unless those constraints are normalized.Stop and go upstream if:
section-logic-polisher and writer-selfloop no longer flag this file.