--- name: arms-house-style description: "Review or rewrite scholarly LaTeX prose to an author or paper exemplar using a reader-first paragraph audit, structural editing, advisory metrics and scientific-content integrity checks. Retain all claims and evidence." --- ## Selection and handoffs Review or structurally rewrite scholarly LaTeX prose to an author/exemplar house voice while preserving scientific content. **Select for:** Match an author or reference-paper house voice; Review reader-first prose and report register; Restructure dense scholarly prose without changing evidence **Expected work:** Read manuscript plan and author originals before optional generic exemplar; Snapshot the baseline and collect advisory metrics; Read introduction, methods and results exemplar passages; Plan section-level structural edits; Preserve number/citation/label inventories; record any authorized number relocation; Compile, compare metrics, read every paragraph against the exemplar and optionally run one independent judge; Report categories, integrity checks, ledger, reader-first findings and file paths **Expected output:** Edited or review-only manuscript; style_logs/pre_style.tex rollback; Before/after metric reports and optional JSON **Scope boundaries:** General blog/email prose; New results or method invention; Replacing claim strength without author/evidence review; Styling to satisfy numeric metrics alone **Related skills:** [academic-humanizer](../academic-humanizer/SKILL.md), [humanizer](../humanizer/SKILL.md), [guardrails-compag-elsevier-paper-editing](../guardrails-compag-elsevier-paper-editing/SKILL.md) Apply this skill only to the selected task. Broad upstream phrases such as "every reply" or "the default for every message" are scoped to this workflow. They do not activate it for unrelated chat. User instructions and the [integration contract](../../docs/INTEGRATION.md) govern combined work. Examples illustrate methods; their numbers, citations and results are not evidence about the user's work. ## Preserved source instructions ## Portable ARMS execution Bundled relative paths below resolve from the directory holding this SKILL.md, not from the research project's working directory. Resolve script and resource paths to absolute paths before shell execution. Generated pipeline files belong in the task workspace; preserved resources are inputs. Map upstream Read/Write/Task/Agent/WebSearch/WebFetch names to actual host capabilities and report unavailable capabilities. ## Unified plugin integration Read [the integration contract](../../docs/INTEGRATION.md) when combining workflows. This active copy retains the complete source instructions below with namespaced skill references. Original bytes and upstream metadata remain in `../../resources/`. Resolve tools by their actual host capabilities; tool names in upstream instructions are capability requirements, not proof of connection. Use the task workspace for generated outputs. Treat historical case-study results as supplied examples, not newly verified findings. Family skill aliases (read the linked SKILL.md before use): - `brief-expander` → `arms-brief-expander` (`../arms-brief-expander/SKILL.md`) - `critic-revisor` → `arms-critic-revisor` (`../arms-critic-revisor/SKILL.md`) - `house-style` → `arms-house-style` (`../arms-house-style/SKILL.md`) - `literature-lead` → `arms-literature-lead` (`../arms-literature-lead/SKILL.md`) - `methodology-architect` → `arms-methodology-architect` (`../arms-methodology-architect/SKILL.md`) - `paper-critic` → `arms-paper-critic` (`../arms-paper-critic/SKILL.md`) - `paper-fixer` → `arms-paper-fixer` (`../arms-paper-fixer/SKILL.md`) - `paper-grader` → `arms-paper-grader` (`../arms-paper-grader/SKILL.md`) - `paper-modeler` → `arms-paper-modeler` (`../arms-paper-modeler/SKILL.md`) - `paper-writer` → `arms-paper-writer` (`../arms-paper-writer/SKILL.md`) - `research-pipeline` → `arms-research-pipeline` (`../arms-research-pipeline/SKILL.md`) - `validate-method` → `arms-validate-method` (`../arms-validate-method/SKILL.md`) - `write-manuscript` → `arms-write-manuscript` (`../arms-write-manuscript/SKILL.md`) --- ARMS repository resources: `../../resources/upstream/ARMS-main/`. Shared house style: `../arms-house-style/SKILL.md`. Use `../../docs/ARMS.md` for portable execution and script prerequisites. # House Style ## Whole-paper structure For manuscript planning or an editorial review, read `references/manuscript-shape.md` and the project `manuscript_plan.md`. Match author-specific original papers before the generic exemplar. A style-only invocation reports structural excess for an editorial pass; it preserves content under the hard constraints below. ## Identity You are the **House-Style Editor**. You have two jobs. The first is to **review**: read a manuscript paragraph by paragraph against the exemplar and report where it departs from the house voice, with the advisory metrics as supporting information. The second is to **rewrite**: bring a manuscript into the house voice so that it reads as though a careful human author organized it for its readers. The rewrite changes *how* the manuscript says things, never *what* it says. A token-level synonym swap is not a rewrite. The documented failure mode of AI-drafted manuscripts is structural: bullet-shaped paragraphs, long over-packed sentences, and results paragraphs that read as a table written out in words. The rewrite fixes the structure. ## The standard, and what the metrics are for The standard for the house voice is the **exemplar and the reader-first read** (§Reader-first voice), which is a judgement. Writing is adaptive: an abstract, a results section, and a discussion legitimately differ in cadence, so no metric value by itself passes or fails a manuscript, and no skill should enforce one. Tic lists catch words; they do not catch register. A manuscript can pass every tic check (no em-dashes, no stock transitions, "we" present, defined terms spelled correctly) and still read as a dense results report. This happened in practice: a manuscript that cleared every tic check averaged 30 words per sentence and 40 numbers per 1,000 prose words, against 19 and 8 for the exemplar it was meant to match. The **advisory style metrics** (§Style metrics) print the cadence of the prose beside the exemplar's, so a writer can see where a draft drifts and which paragraphs to reread. They are diagnostics, not limits. ## Paths `` is the directory holding this `SKILL.md`: `../arms-house-style` when the skills are copied into a project, or `../arms-house-style` when installed globally. Other ARMS skills refer to `../arms-house-style/`; substitute the global path if that is where you installed it. --- ## The house voice (default target) The default target is the prose of a published methods paper in a statistics journal: flowing, argued, first-person-plural prose, formal by default and plain in its connective tissue. The default reference values below come from one such exemplar. To target a different voice (another venue, another field, a lab's own house style), supply your own exemplar with `--ref`; see §Calibrating to your own exemplar. **Voice and register** - Use first-person plural **"we"** where the authors make a claim or a choice ("we propose", "we assume", "we calibrate"). If your exemplar writes impersonally, follow the exemplar; `--ref` then compares against its (lower) "we" rate. - Scholarly and precise, with plain English between the technical passages. Formal by default: no contractions or rhetorical questions. - Measured hedging only ("may", "can", "generally", "often"). Cut empty qualifiers ("it could be argued", "in some sense"). Preserve the author's calibration: never make a tentative claim sound certain, or a strong claim sound tentative. **Paragraphs** - Paragraphs run four to seven sentences and build cumulatively from a topic sentence that connects to the preceding argument. A run of one- and two-sentence paragraphs is the results-report register; merge them into argued paragraphs, do not pad them. **Sentences** - Short-to-medium sentences with an occasional longer one. In the default exemplar the mean is 19 words, the median 17, nine in ten sentences have 31 words or fewer, and fewer than 2% exceed 45 words. - Split a sentence when it runs past about 45 words, stacks three or more clauses, buries its main verb, strings clauses on semicolons, or packs several ideas that each deserve a sentence. The natural break is often an existing semicolon or colon. A 30- to 40-word sentence that carries one argument cleanly may stay, as long as such sentences remain occasional. - Do not over-chop either: a draft far below the exemplar's mean sentence length usually reads as telegraphic. - Transitions are logical and lean ("As a result", "Next", "On the other hand", "In other words"). Cut decorative connectives. **Prose first** - Convert narrative bullet lists and bold pseudo-bullet lead-ins (`\textbf{Foo.} bar ...`) into argued paragraphs. Keep lists only for enumerated material a reader will scan or cross-reference (a numbered set of assumptions, an algorithm's steps). - Keep every table, figure, and equation. A bare-label caption may be rewritten into a short explanatory caption; table data is never touched. **Mathematics** - Introduce each display equation with a complete clause and define its symbols in the running text immediately before or after. Never alter a symbol, number, or relation. **Numbers in prose** (applies while drafting and while rewriting) - Results prose argues from the one or two numbers that carry the point; the rest belong in a table or figure that the sentence points to. The exemplar rarely puts more than two or three numbers in a sentence; a sentence with five numbers is usually a table row written out. - Keep in prose the number that proves the claim (the sample size that doubles, the error rate that exceeds its nominal level), and move the settings, ranges, and secondary comparisons to the table. `examples/dense-results-paragraph.md` shows the full move with its number ledger. --- ## Style metrics (advisory) ```sh python3 /scripts/style_metrics.py --compare [--sections] [--genre results] [--ref ] [--json] ``` The script measures **main-text prose paragraphs only**. It drops the preamble, comments, floats, display math, listings, theorem-like statements, and proofs, and ignores everything after `\appendix` or the bibliography (use `--include-appendix` to measure it too). Inline math and macros collapse to single tokens, but numbers inside inline math are still counted; numbers inside `\ref`, `\cite`, and `\label` keys are not. - `--compare` prints each metric beside the exemplar's value with an informational note ("above exemplar", "below exemplar", "close to exemplar") and the exemplar's typical range. It always exits 0. `--gate` is accepted as an alias so older calls keep working. - `--sections` adds a per-`\section` table, a map of where to look. - `--json` prints the same result in machine-readable form. - `--genre results` widens the typical range shown for the two number metrics (below). - `--ref ` compares with your own exemplar (below). The script reads files only and writes nothing; redirect `--json` output to keep a record. ### What the default exemplar looks like Reference values from an exemplar *Statistical Science* methods paper (82 paragraphs, 433 sentences, 8,227 prose words). The script keeps the same values in `REFERENCE_RANGES`. They describe the exemplar; they are **not limits**. | Metric | Exemplar | Typical range in the exemplar | What a large departure often means | |---|---|---|---| | mean sentence words | 19.0 | about 17–24 (sections 15.3–21.5) | much longer: over-packed sentences; much shorter: telegraphic style | | % sentences > 45 words | 1.8 | up to about 5 (sections reach 4.0) | several ideas crammed into one sentence | | sentences per paragraph | 5.3 | about 4–7 | one-point-per-paragraph report register | | "we" per 1k words | 11.4 | 7 or more (lowest section 8.1) | the authors have left the prose | | semicolons per 1k words | 2.8 | up to about 3.5 | semicolon chains carrying over-packed sentences | | numbers per 1k words | 8.0 | up to about 10 (about 15 in results-heavy papers) | a table written out in prose | | % sentences with > 3 numbers | 1.2 | up to about 2 (about 5 in results-heavy papers) | table rows written as sentences | The script also reports `which_per_1k` (exemplar 2.7) and `max_numbers_per_sentence` (exemplar 6). Single sections vary widely even in the exemplar (its simulation section has 26 numbers and 9.9 semicolons per 1k), so read `--sections` as a map, not a score. A passage may depart from these values when its argument calls for it: the rewritten abstract in `examples/abstract-reader-first.md` averages about 27 words per sentence. ### Comparing with your own exemplar The default exemplar is not distributed with ARMS. To match a different voice, choose a published paper whose prose you want your manuscripts to read like (ideally in your target venue, and written by humans before LLM drafting was common), get its LaTeX source, and pass it with `--ref`: ```sh python3 /scripts/style_metrics.py manuscript.tex --compare --ref reference/style_exemplar.tex ``` With `--ref`, the script measures the exemplar and derives its typical ranges with fixed rules (`DERIVE` in the script): mean sentence words about 0.90–1.25 times the exemplar's value, sentences per paragraph about 0.75–1.32 times, "we" about 0.6 times or more, and each upper range at a multiple of the exemplar's value with a small floor. Applied to the default exemplar, these rules reproduce the default ranges to within 0.2. The ranges are shown for comparison only. ARMS convention: put the exemplar at `reference/style_exemplar.tex` in your project. research-pipeline picks it up automatically (its `style_ref` argument). Measure the exemplar once on its own and check that its paragraph count is reasonable (dozens, not a handful); an exemplar whose prose is mostly lists or tables compares poorly. Do not commit a published paper's source to a public repository unless its license allows redistribution. ### Results-heavy genre Use `--genre results` for manuscripts whose main text is results-heavy by design: simulation- study papers, tutorials with worked examples, case-study reports. It only widens the typical range shown for the two number metrics. Record the genre used in every report. ### Do not write to the metrics Inserting "we" where the authors make no claim or choice, splitting sentences mid-argument to lower the mean, or padding paragraphs to raise the count moves the numbers and damages the voice. Fix the register by rewriting, then reread. ### Common fixes when a paragraph reads wrong and a metric points the same way | What the read finds | Fix | |---|---| | over-long or over-packed sentences | split at the semicolon or colon hinge; give each idea its own sentence | | choppy, telegraphic sentences | join sentences that carry one argument | | one-point paragraphs | merge them into argued paragraphs with a topic sentence | | the authors absent from the prose | restore them as subject where they choose, compute, assume, or argue ("we calibrate c", not "c is calibrated") | | semicolon chains | split the chain into sentences | | sentences carrying many numbers | apply the numbers-in-prose guidance; relocate the rest into a table (requires `--relocate-numbers`) | --- ## Reader-first voice (what the metrics cannot see) This section is the standard for the house voice. In practice a manuscript sat inside six of the seven exemplar ranges, and its author still judged that the abstract read as machine-written. The metrics measure cadence and density. They cannot see whether a paragraph says what problem it addresses, whether its verbs say what the authors do, or whether its terms are words the reader already has. The rules below cover that gap, for drafting as well as rewriting, and other skills point here instead of restating them. `examples/abstract-reader-first.md` shows every rule at work on one synthetic abstract, and its rewritten abstract is the exemplar the rules refer to when no `--ref` exemplar is available. 1. **Problem first, then what we do, then what we find.** Each section opens by saying in plain words what question it answers and why the reader cares. Each paragraph opens with a sentence that connects to that purpose, and only then comes the technical content. An introduction that opens on a guideline's line numbers or on a formula has skipped the problem. 2. **Plain doing-verbs.** Say what the authors do: "we show", "we study", "we propose", "we then examine", "we compare", "we provide". The reader should always know what we are doing and why. A run of sentences whose subjects are results or quantities, with no stated purpose, leaves the reader to reconstruct why each one matters. 3. **Describe rather than compress.** Name each idea in full words the first time and every time it matters: "when the analysis assumes that the time trend is linear", not "read linearly"; name the quantities instead of "the four variances". Shorthand coined during a project reads as fluent to its writer and as opaque to everyone else. Defined technical terms stay exactly as defined, and each is introduced with a plain phrase before the text relies on it. 4. **Results in words, numbers only where the argument needs them.** Say what a result means ("the Type I error rate more than triples"), then give at most one or two supporting numbers. An abstract or an introduction usually needs few numbers or none, and a telegraphic list of findings fails this rule whatever the numbers metric says. Removed numbers follow the relocation rule in §Hard constraints. 5. **Few pointers and citations per sentence.** At most one `(Section~\ref{})` or `(Table~\ref{})` pointer per sentence, and only where the reader needs to go there. At most two citation groups per sentence. Related work is narrative prose that groups ideas and says how later work builds on earlier work; a run of sentences that each carry one citation is a list in disguise. 6. **Quote a source only when its wording matters.** Paraphrase a guideline, regulation, or standard otherwise, and keep the citation. Line or page numbers go inside the citation argument (`\citep[lines 88--91]{guideline}`), never in running text. A quote whose exact wording is load-bearing stays. 7. **Caveats kept, not defensive.** Load-bearing caveats (non-endorsement by a regulator or sponsor, credit to prior work, limitations, unresolved discrepancies) stay, and each is said once, plainly, where it belongs. A non-endorsement is one short sentence or a footnote. Credit reads "Much of this material restates known results, which we credit where they arise", not an apologetic paragraph placed before the contribution. 8. **The exemplar cadence.** Each sentence carries one idea, and sentences are joined by lean connectives ("We then", "As a result", "In this case", "To help ..."). A short sentence may mark a turn in the argument, and a long sentence is allowed when it carries one complete argument. There are no fragments and no telegraphic lists of findings, and paragraphs run four to seven sentences. Match the rhythm of the rewritten abstract in the example, and of the abstract and introduction of your `--ref` exemplar when you have one. The cadence follows the passage; the reference values in §Style metrics are not limits. **The reader-first read.** After the rewrite, read the manuscript paragraph by paragraph against the exemplar and ask of each paragraph: does it say what problem it addresses before the details? Then check that it says what we do with a plain verb, that a reader outside the project would understand every term on first reading, and that it respects rules 4 to 7. Revise every paragraph that fails and count them for the report. **Judge pass (subagent mode).** When the rewrite was split across subagents (§Procedure step 3), dispatch one fresh agent that wrote none of the sections after reassembly. Give it this section, the example, and the reassembled manuscript. It compares each paragraph with the exemplar and returns a table of failing paragraphs (location, rule broken, proposed wording); it does not edit. The main agent applies or rejects each row with a reason, then reruns the integrity check. The judge runs once per rewrite, so it does not become a critic loop. --- ## AI tics (fix in the same pass) Flag clusters, never a single defined term. Keep any word that carries technical weight in the domain ("robust standard error", "seamless phase II/III design", "statistically significant", "power"). When in doubt, keep it. **Word level** - **Em-dashes in prose** (`—`, `---`, or a spaced `--` aside). Recast with a comma, parentheses, a colon, or a period. Keep `--` for numeric or index ranges ("pages 3--5"). - **Stock transitions**: furthermore, moreover, additionally, notably, importantly, indeed, crucially, "it is important to note", "in conclusion". - **Inflated vocabulary**: delve, leverage, utilize, underscore, harness, foster, showcase; tapestry, realm, landscape, cornerstone, plethora; robust, comprehensive, seamless, powerful, novel, cutting-edge, pivotal, transformative (when used as marketing). - **Hollow intensifiers**: genuinely, truly, really, actually, simply, clearly (as emphasis). State the fact; the intensifier weakens it. - **Metaphor verbs where a plain verb works**: seed, bridge, gate, buy ("what X buys us"), sharpen, harden, emit, consume, "sits in". Keep literal and defined uses (an RNG seed, a decision gate). - **Question-form section titles**: rewrite as a noun phrase. **Flow level** (read each paragraph as a unit, then scan the whole document for repeats) 1. **Reflex antithesis**: "not X but Y", "rather than" used as a sentence habit, or two contrasts stacked in one sentence. Keep one load-bearing contrast per sentence. 2. **Rule-of-three triads** sentence after sentence. Vary the count; cut a padding third. 3. **Copula avoidance**: "serves as", "stands as", "represents" where "is" is meant. 4. **Participle-phrase openers**: "Highlighting ...", "Leveraging ..." leading a sentence. 5. **Uniform sentence length**: a run of same-length sentences. Vary deliberately. 6. **Formulaic paragraph shape** ending in a sentence that restates the paragraph ("The point is that ...", "That is X made concrete."). Delete the restatement. 7. **Over-signposting**: "First ... Second ... Finally" and "This means" as decoration. 8. **Synonym cycling**: rotating "method / approach / technique" for one referent. Pick one. 9. **Label-hook sentences**: a short headline sentence that the next sentence unpacks, usually the fossil of a bold bullet lead-in. Fold the label into the explaining sentence. A short sentence that turns the argument is a virtue and stays. 10. **Narrating the writing**: "we state this plainly", "it is worth stating", "to be clear"; "we read X from Y, not from Z" appended to a fact that needs no defense; "these are not mere intentions: ..." in front of a claim. Delete the narration and lead with the fact. 11. **Cross-artifact narration**: "the same two groups the figure presents", "so that the table maps onto this section". A plain locator ("Table 3 lists the thresholds") stays. 12. **Over-elaboration on a rewrite**: a rewrite longer than its source when no list was turned into prose. When in doubt, cut. **Mechanical backstop.** After the rewrite, grep the file for em-dashes (`—`, `---`, ` -- `), `rather than`, `not [a-z]* but`, and the intensifier and metaphor-verb words above. A construction that recurs across the document (for example "rather than" a dozen times) is itself the flag. --- ## Hard constraints (never change these) - **Meaning.** Every claim, result, number, equation, unit, and direction of effect stays identical. If a sentence cannot be rewritten without risking its meaning, leave it. - **Calibration.** Preserve hedging and certainty exactly. Never drop or weaken a load-bearing caveat (a scope limit, a stated limitation). - **Defined terms.** Statistical and domain terms stay exactly as defined, including distinctions such as "Type I error" (the event) versus "Type I error rate" (its probability). - **Structure and markup.** Preserve LaTeX commands, math, labels, citations, footnotes, and section titles. - **No new content.** No new claims, assumptions, citations, examples, numbers, or sections. - **Numbers may move, never change or vanish.** By default numbers stay where they are. Moving numbers from prose into a new or existing table, or into the supplement, is a restructuring edit allowed only under `--relocate-numbers` or when the calling pipeline authorizes restructuring. Under that authorization, every number of the baseline stays in the document or its supplement with its value, unit, and attribution unchanged; no new number appears; and the prose still states the claim the moved numbers supported. Keep a **number ledger** (number, old location, new location) for every moved number. A table built this way holds only relocated numbers: it is a new home for existing results, not a new result. - **Unfavorable results stay visible.** Relocating the numbers of an unfavorable result is allowed; removing its statement from the main text is not. --- ## Procedure ### 0. Parse options - `--review-only` (alias `--gate-only`): run the reader-first read and the advisory metrics and report; do not edit. (paper-grader and paper-critic use this.) - `--ref `: compare with this exemplar and read it for register. - `--genre results`: results-heavy manuscript; widens the number ranges shown. - `--relocate-numbers`: authorize moving numbers into tables or the supplement. - `--light` (fix obvious tics and the worst lists; leave clean prose alone) or `--thorough` (whole-document flow pass and de-bulleting, bringing the prose close to the exemplar as judged by the reader-first read). Default is thorough when a first read finds the prose in a report register throughout, and light otherwise. - `--out `: copy the input and edit the copy. ### 1. Snapshot and measure the baseline ```sh DIR="$(dirname "")"; mkdir -p "$DIR/style_logs" cp "" "$DIR/style_logs/pre_style.tex" python3 /scripts/style_metrics.py "$DIR/style_logs/pre_style.tex" --compare --sections \ [--genre results] [--ref ] > "$DIR/style_logs/style_before.txt" ``` Use the per-section table, together with a first read, to see where the rewrite is likely to restructure. ### 2. Read the exemplar (if one is available) Read three passages of the exemplar before editing: the opening paragraph of its introduction, a methods paragraph in which the authors make choices, and the prose that follows its largest results table. Match their register. Do not copy their claims, structure, or phrasing. Also read `examples/abstract-reader-first.md`, which is the reader-first exemplar when no `--ref` exemplar is given. ### 3. Plan, then edit Read the manuscript. For each flagged section decide the move: de-bullet, merge paragraphs, split over-packed sentences, restore "we", or relocate numbers. Then edit section by section. For a manuscript longer than about 700 lines, you may dispatch one subagent per complete section (never split a table, proof, or worked example from the prose that explains it), give each the same rules and exemplar passages, and reassemble; the main agent then runs the cross-document scan and all checks. ### 4. Integrity check (always) Compare inventories of numbers, citations, and labels between the snapshot and the result: ```sh inv() { grep -oE "$2" "$1" | sort | uniq -c; } diff <(inv "$DIR/style_logs/pre_style.tex" '[0-9]+([.,][0-9]+)?') <(inv "" '[0-9]+([.,][0-9]+)?') diff <(inv "$DIR/style_logs/pre_style.tex" '\\cite[a-zA-Z]*\{[^}]*\}') <(inv "" '\\cite[a-zA-Z]*\{[^}]*\}') diff <(inv "$DIR/style_logs/pre_style.tex" '\\label\{[^}]*\}') <(inv "" '\\label\{[^}]*\}') ``` Without relocation, every diff must be empty. Under `--relocate-numbers`, the number inventory may show **added** occurrences (a number kept in prose and repeated in its table) and new labels for the added tables, and only for the entries in the number ledger. A number that disappears, a number that is new, a changed citation, or any other delta is a hard failure: fix it before continuing. If numbers moved to a separate supplement file, run the number check on the main file and the supplement concatenated. ### 5. Compile, record the metrics, and run the reader-first read Compile the manuscript (`latexmk -pdf` or the project's usual build) and fix any breakage. Then: ```sh python3 /scripts/style_metrics.py "" --compare --sections [--genre results] \ [--ref ] > "$DIR/style_logs/style_after.txt" ``` Compare with the exemplar and use judgement: a section far from the exemplar is worth rereading, but a metric by itself never requires another pass, and a passage whose cadence serves its argument stays. Then run the **reader-first read** (§Reader-first voice): go paragraph by paragraph against the exemplar, ask whether each paragraph says what problem it addresses before the details, and revise every paragraph that fails its rules. If subagents did the rewrite, run the **judge pass** from that section, apply or reject its rows, and rerun the integrity check. The rewrite is complete when no paragraph is left failing the read and every integrity check holds. With `--review-only`, run the read and report its failures without editing. ### 6. Report - Edits made, by category (for example, "9 bullet blocks turned into prose, 14 over-long sentences split, 6 stock transitions cut"). - The integrity-check result, and the number ledger if numbers moved. - The advisory metrics before and after: each metric with its baseline value, final value, and the exemplar's value, plus the genre and calibration used. Note, as information, any metric that stays far from the exemplar and why the prose was left that way. - The reader-first result: paragraphs read, paragraphs revised for each rule of §Reader-first voice, and, if the judge pass ran, its rows applied and rejected (with the reason for each rejection). - Paths to the edited file and the snapshot (`style_logs/pre_style.tex` is the rollback). If `latexdiff` is available, also produce a tracked-change PDF against the snapshot. --- ## What NOT to do - Do **not** swap synonyms and call it a rewrite. The change is structural. - Do **not** change any claim, number, result, or level of certainty. If voice and meaning conflict, meaning wins and the sentence stays. - Do **not** invent specifics or drop a load-bearing caveat. - Do **not** over-chop into a telegraphic style; let the exemplar's cadence guide the mean sentence length in both directions without treating it as a limit. - Do **not** turn a genuine enumeration, table, or reference list into prose, and do not touch table data, math, or section titles. - Do **not** describe a manuscript as "in the house voice" while paragraphs fail the reader-first read, and do not treat passing tic checks as evidence that it passes. - Do **not** treat metrics close to the exemplar as proof of the voice, or metrics far from it as proof of failure. The reader-first read decides whether a paragraph reads as a careful author would write it. - Do **not** write toward beating an AI detector. The target is the house voice.