ts-paper-refine

ts-paper-refine — right-size & polish (one pass)

Do this as one reasoning pass over the whole draft (you see every section at once, so coherence and de-duplication are free). The original's purpose was conference page-compression — that is wrong for a journal. Traitement du Signal papers are substantial (≈8–14 two-column pages); right-size, do not aggressively cut.

What to enforce

• Length toward the template's targets — the table below is the ts_iieta default; for another template, right-size to its template.json sections[].words (e.g. NeurIPS bands differ). The spec is authoritative (draft_lint reads bands from template.json). Expand thin sections, trim only genuine bloat — never compress below the floor:

  section (ts_iieta)	final words
  method	2000–3000
  experiments	1000–1500
  introduction	800–1200
  related_work	800–1200 (3 themes)
  analysis	900–1400
  conclusion	200–280 (1 paragraph)
  abstract	150–220 (1 paragraph)

  These bands target a substantial ~10–12 page two-column TS article. Expand by adding real substance (more method detail/derivation, more components, deeper analysis, more ablation discussion, fuller related-work lineage) — never by padding, restating, or filler sentences.

• Coherence: the same method/component/dataset/baseline names everywhere (use the blueprint terminology); no contradictions between sections; intro/abstract accurately summarize what Method/Experiments actually say.

• No redundancy: each mechanism explained once (Method), not re-derived in Analysis; remove repeated motivation across intro/related/method.

• Preserve exactly: every \cite{}, every equation, every \ref{}, every table/figure label and the -- result placeholders. Refinement changes prose, never citations/structure/math semantics.

• Honesty intact: still zero fabricated numbers in any sentence; results stay forward-looking; tables stay blank.

• TS tone: no bold in body, hedged conclusion language, no hype, no filler phrases.

De-AI pass (this is AI-written prose — scrub the tells)

Every section was drafted by an LLM, so it is prone to AI tells. Read each section and rewrite where it improves (don't churn natural sentences). Scan for and replace:

• Prose flow (the #1 fix). Turn any bare-comma comma soup into connected prose: punctuate an appositive/enumeration with a colon or em-dash, not a bare comma — "a pipeline: retrieval, planning" or "a pipeline — retrieval, planning — in which …", NOT "a pipeline, retrieval, planning, … in which …" (that comma soup is what reads as fragmented "碎句子"). A legitimate appositive em-dash is GOOD and encouraged; only remove em-dash abuse (dramatic-pause dashes). Vary sentence length; each paragraph must read as a flowing whole, not a clause list. Avoid the recipe rhythm of "X does A [cite], Y does B [cite]" in related work — synthesize across papers.
• Tell words/phrases: leverage, delve, utilize, showcase, underscore, intricate, pivotal, seamless, holistic, nuanced, realm, tapestry, testament, "it is worth noting", "plays a crucial role", "in order to" (→ "to"), rule-of-three triplets, connective stacks (firstly/moreover/furthermore — a single appropriate transition is fine), empty "-ing" wind-ups, vague attribution ("studies show"), "not only … but also" overuse.
• Translationese (these drafts trend that way): stacked attributive chains ("the … of the … of the …"), gratuitous passive voice, hollow rhetoric (paradigm shift, disruptive, profound, in essence). Prefer plain, precise words; keep \cite/\ref/math intact; add no new emphasis. draft_lint code-enforces (hard-fails, ai_tell) ONLY the CONTEXT-FREE canned phrases — tapestry, testament, "plays a crucial role", "it is worth noting", "delve into", "realm of", "paradigm shift", "navigating the landscape", "ever-evolving", … plus in order to — those MUST be gone. The CONTEXT-DEPENDENT cohesion tells — em-dashes, sentence-initial firstly/moreover/furthermore/additionally,, not only … but also, rule-of-three triplets, leverage/utilize — are deliberately NOT gated (a regex would kill the good appositive em-dash along with the abuse — precisely what produced the comma soup): you remove the abuse by judgment while keeping legitimate uses. (Source: PaperJury writing-toolkit de-ai.)

Logic self-check (after each edit, narrow self-gate — not a reviewer)

After you change a passage, re-read ONLY that passage for a show-stopper introduced by the edit: a logic contradiction, an undeclared terminology switch (a term silently renamed), or severe Chinglish. Fix it; do not raise style nits here (the de-AI pass owns those). This is the in-pass honesty check; the adversarial peer-review (ts-paper-review, stage 5, runs by default) is the heavier, separate hardening loop that runs after this stage and routes its fix-now issues back here in Review-fix mode.

DATA-AWARE branch (only when template.results_mode == "data_aware")

When the paper reports real results (data-aware — see ts-paper-data), the "remove unbacked numbers" instinct inverts for numbers that trace to results.facts.json:

• Do NOT convert real numbers to forward-looking language; do NOT remove real numbers from running text; do NOT flag real (traceable) numbers as hallucinations.
• Verify cross-section number consistency — the same metric formatted identically everywhere; the abstract/conclusion headline numbers equal the filled result-table values.
• Preserve the filled result tables and the evidence sentences; preserve ablation deltas; enforce definitive past tense; apply the TBD silence rule (no claim about a "TBD" metric).
• Keep every number consistent with results.facts.json (the number-audit re-runs after refine). Everything else (length right-sizing, coherence, de-duplication, tone) is as in proposal mode.

Match the template's venue type

If the active template is a journal (e.g. ts_iieta), strip any leaked conference framing ("to fit the page limit / 8-page / camera-ready") and don't over-compress — journals are substantial. If the template is a conference style (e.g. neurips), respect its tighter conventions instead. Let the template's word bands — not a fixed venue habit — decide length.

After refining (enforced)

Re-run python ../ts-paper-write/scripts/draft_lint.py <workdir> — it now fails the build on any section outside its template.json word band and on any broken shape contract, so "right-sized" is verified in code, not by eye. Drive every word_band violation to zero by expanding thin sections with real substance (more method detail, more discussion, a deeper failure analysis) — never by padding. Then re-run python ../ts-paper-cite/scripts/citations_lint.py <workdir> to confirm citations survived. Re-run both until they exit 0 (ok:true); do not hand off on a nonzero exit. Then run python ../ts-paper-write/scripts/reflow_tex.py <workdir> so every sections/*.tex is one logical line per paragraph (your de-AI rewrites may have re-wrapped them; idempotent, PDF-neutral). Then write logs/4_refine.io.md (INPUT: pre-refine word counts; DECISIONS: per-section deltas + what was added/cut; OUTPUT: post-refine counts, all in band). Hand off to ts-paper-review (stage 5, the default next stage): it returns a triaged issue list, and for each fix-now issue the orchestrator re-invokes THIS skill in Review-fix mode (below). Only after review is dry/closed does the chain proceed to ts-paper-figure then ts-paper-latex. If the user explicitly opted out of review, hand directly to ts-paper-figure.

Review-fix mode (ONLY when invoked with a single review issue — NOT a full pass)

This mode is triggered only when the orchestrator hands you one ts-paper-review issue {section, severity, evidence_quote, close_criterion} to close. It is a separate entry path: do not run it on a normal stage-4 invocation, and do not run the holistic right-size / de-AI pass here. The default full-pass behavior above is unchanged.

• Make the minimal edit to the named section only that makes that one close_criterion true. Bind strictly to that criterion; do not touch other sections and do not over-rewrite.
• Preserve exactly: every \cite{}, equation, \ref{}, table/figure label, the -- placeholders, and the section's template.json word band (stay in band — no compression below the floor).
• If the fix touches an equation, it must keep notation complete: any symbol the edit introduces into an equation has to be added to the notation table (or the fix renamed to an already-defined symbol) in the same edit — draft_lint now hard-fails notation_incomplete (this is exactly the path that let a stray \psi into the central equation), so an unmatched new symbol will block the re-gate below.
• Re-run python ../ts-paper-write/scripts/draft_lint.py <workdir> AND python ../ts-paper-cite/scripts/citations_lint.py <workdir> until both exit 0 (ok:true) to confirm nothing broke (including notation_incomplete); do not return a fix on a nonzero exit.
• Append the fix (issue id, section, what changed, close_criterion met) to logs/5_review.io.md (the review log — not 4_refine.io.md).
• Return control to the orchestrator for the next issue, or to advance to figure once review is dry.