paperlint 7.2.0 → 9.0.0

package/skills/grade-paper-writing/prose-lint.test.ts

@@ -1,5 +1,5 @@
 /**
- * prose-lint.mjs as a process — the way the skill runs it. Every mode, on
+ * prose-lint.mjs as a process — the way the skill runs it. Every mode, on an accepted paper and on
  * fixtures written here, with the WHOLE output compared: the report is the product, so a changed
  * number or a dropped line is a changed result.
  */
@@ -8,87 +8,146 @@ import { join } from "node:path";
 import { dirname } from "node:path";
 import { fileURLToPath } from "node:url";
 import { expect, test } from "vitest";
+import { cpSync } from "node:fs";
+import { layRecord, recordedTree } from "../../test/recorded-fixture.ts";
 import { runNode, useTempDir, writeTree } from "../../test/support.ts";
 
 const SCRIPT = join(dirname(fileURLToPath(import.meta.url)), "prose-lint.mjs");
+/**
+ * One process that loads the LaTeX parser and reads a full accepted paper: about 1 s alone, 5.6 s
+ * measured under coverage with the suite running in parallel — past vitest's 5 s default.
+ */
+const SPAWN_LATEX_MS = 60_000;
+/** An accepted paper whose bibliography sits inline in `filecontents`, beside a long preamble. */
+const ACCEPTED_DIR = join(
+  dirname(fileURLToPath(import.meta.url)),
+  "..",
+  "..",
+  "fixtures",
+  "accepted-papers",
+  "agenticdev-acm26",
+);
 const root = useTempDir("prose-lint-");
+/** The accepted paper as a build left it: a copy with TeX's record laid at `_build/sources.json`. */
+const ACCEPTED = join(root, "accepted", "paper.tex");
 
 const LONG_CAPTION_SENTENCE = Array.from(
   { length: 45 },
   (_, i) => `word${String(i)}`,
 ).join(" ");
-const PAPER = `---
-title: A title the linter must not read as a sentence
----
-# The paper
-
-<!-- TIGHTEN: a working note, not prose -->
-
-## Introduction
-
-In this paper we present a linter. It might possibly help, but this is not a proof. Clearly the
-first result matters (Section 3). Prior work agrees [1, 2]. The speedup was 40% on average across the
-full suite of papers, which suggests the method might hold generally.
-
-- The first item is short.
-- The second item is not long either.
-
-\`\`\`
-# a heading inside a code block is code
-\`\`\`
-
-| a | table |
-|---|---|
-| is | data |
-
-### Method
-
-We believe that rather than guessing, we measure. Instead of prose, the numbers speak.
 
-## References
+/** A paper.tex whose body is `body`. */
+const tex = (body: string): string =>
+  `\\documentclass{article}\n\\begin{document}\n${body}\n\\end{document}\n`;
 
-[1] A reference that is not prose.
-`;
-
-writeTree(root, {
-  "p/paper.md": PAPER,
+const TREE = {
+  "p/paper.tex": tex(
+    "\\section{Introduction}\nIn this paper we present a linter. It might possibly help.\n\\input{figures/fig1}\n\\input{figures/fig2}\n\\input{figures/fig4}\n\\input{figures/fig5}",
+  ),
   "p/figures/fig1.tex": `\\begin{figure}\\caption{\\textbf{Big.} ${LONG_CAPTION_SENTENCE}. ${"More words here. ".repeat(20)}}\\end{figure}`,
   // Empty once LaTeX commands and braces are stripped: skipped, not counted as a zero-word caption.
   "p/figures/fig2.tex": "\\caption{\\centering}",
-  // A lone "#": no word at all, and no sentence either.
+  // A lone "#": no word at all — measured as zero words, never flagged.
   "p/figures/fig4.tex": "\\caption{{#}}",
+  // Rules only: seen by the reader, but no word once the dashes are stripped — skipped.
+  "p/figures/fig5.tex": "\\caption{---}",
   "p/figures/notes.txt": "not a figure",
-  // Reads as a markdown heading once LaTeX is stripped ("# Big"), so it splits into NO sentence:
-  // the caption scan must not flag it and must not crash.
-  "p/figures/fig3.tex": "\\caption{{#} Big}",
-  "clean/paper.md": "## Intro\n\nShort and plain. Nothing else.\n",
+  "clean/paper.tex": tex("Short and plain. Nothing else."),
+  // Captions in the paper itself and in a section it includes — not under figures/.
+  "inline/paper.tex": tex(
+    `Prose here. \\begin{figure}\\caption{${LONG_CAPTION_SENTENCE}.}\\end{figure}\n\\input{sec/results}`,
+  ),
+  "inline/sec/results.tex": `\\begin{figure}\\caption{${"Short words here. ".repeat(40)}}\\end{figure}`,
+  // One real over-long caption, and the same caption commented out and inside verbatim.
+  "commented/paper.tex": tex(
+    `Prose here.\n% \\caption{Commented ${LONG_CAPTION_SENTENCE}.}\n\\begin{verbatim}\n\\caption{Quoted ${LONG_CAPTION_SENTENCE}.}\n\\end{verbatim}\n\\begin{figure}\\caption{Real ${LONG_CAPTION_SENTENCE}.}\\end{figure}`,
+  ),
+  "ends-on-marks/paper.tex": tex(
+    "The method holds on every input~\\cite{knuth}. The proof is given in full (Section~\\ref{proof}). The result is new.",
+  ),
+  "empty/paper.tex": "\\documentclass{article}",
   "page.txt":
     "A rendered page. It was hyphen-\nated across a line.\n\n12\n\fNext page text here.\nReferences\n[1] cut here.\n",
-  "paper.tex": "\\documentclass{article}",
-});
+  "notes.rtf": "not a paper",
+};
+/** What each paper's last build read: its paper.tex, and the files it includes. */
+const READ: Readonly<Record<string, readonly string[]>> = {
+  p: [
+    "paper.tex",
+    "figures/fig1.tex",
+    "figures/fig2.tex",
+    "figures/fig4.tex",
+    "figures/fig5.tex",
+  ],
+  clean: ["paper.tex"],
+  inline: ["paper.tex", "sec/results.tex"],
+  commented: ["paper.tex"],
+  "ends-on-marks": ["paper.tex"],
+  empty: ["paper.tex"],
+};
+writeTree(
+  root,
+  Object.entries(READ).reduce<Record<string, string>>(
+    (tree, [dir, files]) =>
+      recordedTree(
+        tree,
+        dir,
+        files.map((path) => ({ path, role: "body" as const })),
+      ),
+    TREE,
+  ),
+);
+cpSync(ACCEPTED_DIR, join(root, "accepted"), { recursive: true });
+layRecord(join(root, "accepted"), join(ACCEPTED_DIR, "tex-truth.json"));
 
 test("no arguments: the usage line, exit 0", () => {
   assert.deepEqual(runNode(SCRIPT), {
     status: 0,
     stdout: "",
-    stderr: "usage: node prose-lint.mjs <file.md|file.txt> [more files]\n",
+    stderr:
+      "usage: node prose-lint.mjs <paper.tex|page.txt> [more files] [--flags-only|--headings]\n",
   });
 });
 
-test("a .tex file is refused with exit 2, and nothing is measured", () => {
-  const r = runNode(SCRIPT, [join(root, "paper.tex")]);
-  assert.deepEqual(
-    { status: r.status, stdout: r.stdout },
-    { status: 2, stdout: "" },
-  );
-  expect(r.stderr.replaceAll(root, "<root>")).toMatchSnapshot();
-});
+test(
+  "🔴 a paper.tex is measured on the prose a reader sees — not its preamble, comments or inline .bib",
+  { timeout: SPAWN_LATEX_MS },
+  () => {
+    const r = runNode(SCRIPT, [ACCEPTED]);
+    // Guards: a .tex read as raw text counted the preamble, the comments and the `filecontents`
+    // bibliography — 8127 "words" on this paper's submitted version against about 4636 real ones.
+    const words = Number(
+      /^=== paper\.tex — (\d+) words, \d+ sentences ===$/m.exec(r.stdout)?.[1],
+    );
+    assert.deepEqual(
+      {
+        status: r.status,
+        stderr: r.stderr,
+        inBand: words > 4000 && words < 5500,
+      },
+      { status: 0, stderr: "", inBand: true },
+    );
+    // No BibTeX field and no preamble command reaches the measured text.
+    assert.doesNotMatch(r.stdout, /author\s*=|\\usepackage|\\documentclass/);
+    expect(r.stdout).toMatchSnapshot();
+  },
+);
 
-test("the full report over markdown and rendered text", () => {
-  const r = runNode(SCRIPT, [
-    join(root, "p", "paper.md"),
-    join(root, "page.txt"),
-  ]);
+test(
+  "--headings on a paper.tex lists its section titles, in order",
+  { timeout: SPAWN_LATEX_MS },
+  () => {
+    const r = runNode(SCRIPT, ["--headings", ACCEPTED]);
+    assert.deepEqual(
+      { status: r.status, stderr: r.stderr },
+      { status: 0, stderr: "" },
+    );
+    expect(r.stdout).toMatchSnapshot();
+  },
+);
+
+test("the report over a rendered page: page numbers, hyphenation and the references are dropped", () => {
+  const r = runNode(SCRIPT, [join(root, "page.txt")]);
   assert.deepEqual(
     { status: r.status, stderr: r.stderr },
     { status: 0, stderr: "" },
@@ -96,17 +155,26 @@ test("the full report over markdown and rendered text", () => {
   expect(r.stdout).toMatchSnapshot();
 });
 
-test("--headings lists every heading from level 2 to 4, in order", () => {
-  const r = runNode(SCRIPT, ["--headings", join(root, "p", "paper.md")]);
-  assert.deepEqual(
-    { status: r.status, stderr: r.stderr },
-    { status: 0, stderr: "" },
-  );
-  expect(r.stdout).toMatchSnapshot();
+test("--headings on a rendered page is refused: it marks no heading", () => {
+  const file = join(root, "page.txt");
+  assert.deepEqual(runNode(SCRIPT, ["--headings", file]), {
+    status: 2,
+    stdout: "",
+    stderr: `prose-lint: --headings reads a paper.tex — a rendered page (${file}) marks no heading\n`,
+  });
+});
+
+test("a file that is neither a paper.tex nor a rendered page is refused with exit 2, and nothing is measured", () => {
+  const file = join(root, "notes.rtf");
+  assert.deepEqual(runNode(SCRIPT, [file]), {
+    status: 2,
+    stdout: "",
+    stderr: `prose-lint reads a paper.tex or the text of its rendered PDF (.txt), not: ${file}\n`,
+  });
 });
 
 test("--flags-only: an over-long caption flags and exits 1", () => {
-  const r = runNode(SCRIPT, ["--flags-only", join(root, "p", "paper.md")]);
+  const r = runNode(SCRIPT, ["--flags-only", join(root, "p", "paper.tex")]);
   assert.deepEqual(
     { status: r.status, stdout: r.stdout },
     { status: 1, stdout: "" },
@@ -116,38 +184,121 @@ test("--flags-only: an over-long caption flags and exits 1", () => {
 
 test("--flags-only: a paper with no figures directory is clean, exit 0 and silent", () => {
   assert.deepEqual(
-    runNode(SCRIPT, ["--flags-only", join(root, "clean", "paper.md")]),
+    runNode(SCRIPT, ["--flags-only", join(root, "clean", "paper.tex")]),
+    { status: 0, stdout: "", stderr: "" },
+  );
+});
+
+test("🔴 a paper no build has recorded is read as paper.tex alone, and the run says so", () => {
+  writeTree(root, {
+    "unbuilt/paper.tex": tex("Short and plain. Nothing else."),
+  });
+  const file = join(root, "unbuilt", "paper.tex");
+  const r = runNode(SCRIPT, ["--flags-only", file]);
+  assert.deepEqual(
+    { status: r.status, stdout: r.stdout, stderr: r.stderr },
     {
       status: 0,
       stdout: "",
-      stderr: "",
+      stderr: `prose-lint: ${file} has no current build record (_build/sources.json), so only paper.tex was read, not the files it includes — run \`npx paperlint build\` first\n`,
     },
   );
 });
 
-test("a file with no prose left to measure is refused with exit 2, not reported as NaN", () => {
-  writeTree(root, {
-    "empty/paper.md":
-      "---\ntitle: only frontmatter\n---\n<!-- and a note -->\n",
-  });
-  const file = join(root, "empty", "paper.md");
+test("a .tex with no prose in its body is refused with exit 2, not reported as NaN", () => {
+  const file = join(root, "empty", "paper.tex");
   assert.deepEqual(runNode(SCRIPT, [file]), {
     status: 2,
     stdout: "",
     stderr:
-      `prose-lint: ${file} has no prose to measure — after the frontmatter, comments, code ` +
-      "blocks, tables and the reference section are stripped, no word is left.\n",
+      `prose-lint: ${file} has no prose to measure — after the preamble, comments, floats and ` +
+      "the back matter are left out, no word is left.\n",
   });
 });
 
-test("a body of headings only has no sentence to measure, and is refused the same way", () => {
-  writeTree(root, { "headings/paper.md": "## Intro\n\n### Method\n" });
-  const file = join(root, "headings", "paper.md");
+test("a rendered page with no sentence before its references is refused the same way", () => {
+  writeTree(root, { "blank.txt": "12\nReferences\n[1] only this.\n" });
+  const file = join(root, "blank.txt");
   assert.deepEqual(runNode(SCRIPT, [file]), {
     status: 2,
     stdout: "",
-    stderr:
-      `prose-lint: ${file} has no prose to measure — after the frontmatter, comments, code ` +
-      "blocks, tables and the reference section are stripped, no word is left.\n",
+    stderr: `prose-lint: ${file} has no prose to measure — no sentence is left before the references.\n`,
+  });
+});
+
+test("🔴 in a paper.tex, a sentence ending on a citation or a cross-reference wastes its stress position", () => {
+  const r = runNode(SCRIPT, [join(root, "ends-on-marks", "paper.tex")]);
+  // Guards: the marks used to be dropped from the prose, so neither ending could be seen.
+  assert.match(
+    r.stdout,
+    /^FLAG sentences ending on a cross-ref\/citation\/hedge \(wasted stress position\): 2$/m,
+  );
+});
+
+test("🔴 --flags-only reads the captions of the paper.tex and of the files it includes", () => {
+  const r = runNode(SCRIPT, [
+    "--flags-only",
+    join(root, "inline", "paper.tex"),
+  ]);
+  // Guards: captions were read only from figures/*.tex beside the paper, so a caption set inline,
+  // or in an included section, was never checked and the run exited clean.
+  assert.deepEqual(
+    {
+      status: r.status,
+      stdout: r.stdout,
+      flagged: r.stderr
+        .split("\n")
+        .filter((l) => l.startsWith("   "))
+        .map((l) => l.split(":")[0]),
+    },
+    {
+      status: 1,
+      stdout: "",
+      flagged: [
+        "   paper.tex",
+        "   sec/results.tex",
+        "   run prose-lint.mjs on the file without --flags-only for the sentences",
+      ],
+    },
+  );
+});
+
+test("--flags-only on a rendered page reads the captions in figures/*.tex beside it", () => {
+  writeTree(root, {
+    "rendered/page.txt": "A rendered page. Its prose is fine.\n",
+    "rendered/figures/f.tex": `\\caption{${LONG_CAPTION_SENTENCE}.}`,
+    "rendered/figures/notes.txt": "not a figure",
   });
+  const r = runNode(SCRIPT, [
+    "--flags-only",
+    join(root, "rendered", "page.txt"),
+  ]);
+  assert.deepEqual(
+    { status: r.status, first: r.stderr.split("\n")[1]?.split(":")[0] },
+    { status: 1, first: "   figures/f.tex" },
+  );
+  // …and a page with no figures/ beside it has no caption to flag.
+  assert.deepEqual(runNode(SCRIPT, ["--flags-only", join(root, "page.txt")]), {
+    status: 0,
+    stdout: "",
+    stderr: "",
+  });
+});
+
+test("🔴 a caption commented out or quoted in verbatim is not measured; a real one still is", () => {
+  const r = runNode(SCRIPT, [
+    "--flags-only",
+    join(root, "commented", "paper.tex"),
+  ]);
+  // Guards: the scan matched every `\caption{` in the raw text, comments and verbatim included.
+  assert.deepEqual(
+    {
+      status: r.status,
+      findings: r.stderr
+        .split("\n")
+        .filter((l) => l.startsWith("   paper.tex:"))
+        .map((l) => /— "(\w+)/.exec(l)?.[1]),
+    },
+    { status: 1, findings: ["Real"] },
+  );
 });

package/skills/paper-adversarial-review/SKILL.md

@@ -5,7 +5,7 @@ context: fork
 allowed-tools: [Read, Write, Grep, Glob, Agent, Skill]
 ---
 
-<!-- vigiles:sha256:fe41e7c71eb4672d compiled from skills/paper-adversarial-review/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:cb852f6fb5fb9988 compiled from skills/paper-adversarial-review/SKILL.md.spec.ts -->
 
 # paper-adversarial-review — be the reviewer who wants to reject it
 
@@ -22,7 +22,7 @@ reviewer and returns an actionable critique.
 
 ## How to run it
 Prefer a **separate model as the reviewer** (e.g. Fable via a subagent) so it's not the author
-grading itself. Give the reviewer the full paper text (Read the `.tex`/`.md`/PDF) and the target
+grading itself. Give the reviewer the full paper text (Read the `.tex`/PDF) and the target
 venue + paper type (short/full/position/benchmark). Ask for the structured output below.
 
 ## The review must cover (ranked by how often it kills a paper)

package/skills/paper-adversarial-review/SKILL.md.spec.ts

@@ -28,7 +28,7 @@ reviewer and returns an actionable critique.
 
 ## How to run it
 Prefer a **separate model as the reviewer** (e.g. Fable via a subagent) so it's not the author
-grading itself. Give the reviewer the full paper text (Read the \`.tex\`/\`.md\`/PDF) and the target
+grading itself. Give the reviewer the full paper text (Read the \`.tex\`/PDF) and the target
 venue + paper type (short/full/position/benchmark). Ask for the structured output below.
 
 ## The review must cover (ranked by how often it kills a paper)

package/skills/paper-pipeline/SKILL.md

@@ -4,7 +4,7 @@ description: The orchestrator for writing a research paper end-to-end, from idea
 allowed-tools: [Read, Write, Edit, Grep, Glob, Bash, WebSearch, WebFetch, Agent, Skill]
 ---
 
-<!-- vigiles:sha256:a8a20289badf1fa8 compiled from skills/paper-pipeline/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:67a3989629952dc1 compiled from skills/paper-pipeline/SKILL.md.spec.ts -->
 
 # paper-pipeline — the conductor for the whole organism
 
@@ -191,7 +191,7 @@ These have no position, because a position is what made them look finished. Each
 | any `\cite` added or moved | **`verify-citations`** | An edit on the last day adds a citation the last run never saw. "We checked last cycle" does not count — the project rule, and it was numbered as a one-shot anyway. |
 | a genuine sibling surfaces | **`analyze-sibling-paper`** | Deep-read, scoop/delta, saved in `<paper-dir>/siblings/`. |
 | the contribution's framing moves | **`map-prior-work`** (delta mode) | A cheap re-sweep. ⚠️ If a rival first surfaces during citation checking, the SETUP sweep was skipped and you are reshaping the contribution under deadline. |
-| any `.tex` / `paper.md` edit | **`render-paper`** + page count | **Not a stage — a command.** Render, then eyeball pages 1–2 before believing anything about the paper. 🔴 **Rendering is continuous; ACTING on the page count is not** — see the rule below. |
+| any `.tex` edit | **`render-paper`** + page count | **Not a stage — a command.** Render, then eyeball pages 1–2 before believing anything about the paper. 🔴 **Rendering is continuous; ACTING on the page count is not** — see the rule below. |
 | the deadline approaches | the `access` row | Re-check that the profile is still ACTIVE and the portal still opens. External state rots without telling you. |
 
 ## GATES — late, expensive, each blocked on a named input
@@ -310,25 +310,22 @@ binds every row from now on, and the existing rows are *enumerated rather than e
 says out loud how many figures still have no machine-checked path to data. On the paper that has it:
 7 guarded, 248 grandfathered. That ratio is a finding, not a score.
 
-**Adopting it in another paper:** copy `repro/paper_numbers.py`, `repro/paper_numbers.selftest.py` and
-the three lines in `md2submission.py` that call `expand()`; generate the grandfather list from the
-current draft; register the numbers a reviewer will actually check first. `check-provenance.mjs`
-reports `no-numbers-gate` for any paper that has not.
+**Adopting it in another paper:** copy `repro/paper_numbers.py`, `repro/paper_numbers.selftest.py`
+and the build step that calls `expand()` on `paper.tex` before `pdflatex`; generate the
+grandfather list from the current draft; register the numbers a reviewer will actually check first.
 
 🔴 **Do not write the substitution half — it is `jinja2`, and `{{name}}` is already its syntax.**
-Eleven candidates were installed and run on a real markdown-source ACL paper on 2026-08-05. The
-result is not a matter of taste:
+Eleven candidates were installed and run on a real ACL paper on 2026-08-05. The result is not a
+matter of taste:
 
 | candidate | verdict, from an actual run |
 |---|---|
 | **jinja2** | **adopt.** Our `{{annotated.failRate}}` was already valid Jinja2; dotted names are attribute access. Zero characters changed in the paper, byte-identical `.tex`. This is Manubot's mechanism, taken without Manubot. Use `StrictUndefined`, and move the block/comment delimiters off `{%`/`{#` if the paper might ever quote a config file |
 | **knitr** | equally proven, also byte-identical. Rejected only for putting an R runtime into an artifact that otherwise needs `python3` + `pdflatex` |
-| `\newcommand` + `numbers.tex` | works (1 line in the converter); the folk convention, ~12 GitHub hits, all hand-rolled. No package exists |
-| **Quarto**, **pandoc lua filters** | ❌ **destroy the paper.** Markdown→markdown is a pandoc AST round-trip: HTML comments become raw blocks and `[58]` becomes `\[58\]`, so every citation renders as literal `\textbackslash{}[58]`. 3,120 diff lines on the real source. No ACL format exists for Quarto either |
+| `\newcommand` + `numbers.tex` | works; the folk convention, ~12 GitHub hits, all hand-rolled. No package exists |
 | **cog.py** | ❌ line-oriented; cannot substitute mid-sentence |
 | **pythontex** | ❌ not in Ubuntu TeX Live at any level, not on PyPI; CTAN zip + `.ins` build |
-| **showyourwork** | ❌ LaTeX-only, and it manages *figures*, not in-text numbers |
-| **Pweave** | ❌ dead — the one candidate with a markdown native format, broken against modern IPython |
+| **showyourwork** | ❌ it manages *figures*, not in-text numbers |
 | DVC · datalad · papermill | orthogonal and **worth adding**: they pin the data file to the pipeline that made it, a layer below anything here |
 
 **🔴 Computing the value DISSOLVES the freshness problem — do not build a checker for it.** With no

package/skills/paper-pipeline/SKILL.md.spec.ts

@@ -208,7 +208,7 @@ These have no position, because a position is what made them look finished. Each
 | any \`\\cite\` added or moved | **\`verify-citations\`** | An edit on the last day adds a citation the last run never saw. "We checked last cycle" does not count — the project rule, and it was numbered as a one-shot anyway. |
 | a genuine sibling surfaces | **\`analyze-sibling-paper\`** | Deep-read, scoop/delta, saved in \`<paper-dir>/siblings/\`. |
 | the contribution's framing moves | **\`map-prior-work\`** (delta mode) | A cheap re-sweep. ⚠️ If a rival first surfaces during citation checking, the SETUP sweep was skipped and you are reshaping the contribution under deadline. |
-| any \`.tex\` / \`paper.md\` edit | **\`render-paper\`** + page count | **Not a stage — a command.** Render, then eyeball pages 1–2 before believing anything about the paper. 🔴 **Rendering is continuous; ACTING on the page count is not** — see the rule below. |
+| any \`.tex\` edit | **\`render-paper\`** + page count | **Not a stage — a command.** Render, then eyeball pages 1–2 before believing anything about the paper. 🔴 **Rendering is continuous; ACTING on the page count is not** — see the rule below. |
 | the deadline approaches | the \`access\` row | Re-check that the profile is still ACTIVE and the portal still opens. External state rots without telling you. |
 
 ## GATES — late, expensive, each blocked on a named input
@@ -327,25 +327,22 @@ binds every row from now on, and the existing rows are *enumerated rather than e
 says out loud how many figures still have no machine-checked path to data. On the paper that has it:
 7 guarded, 248 grandfathered. That ratio is a finding, not a score.
 
-**Adopting it in another paper:** copy \`repro/paper_numbers.py\`, \`repro/paper_numbers.selftest.py\` and
-the three lines in \`md2submission.py\` that call \`expand()\`; generate the grandfather list from the
-current draft; register the numbers a reviewer will actually check first. \`check-provenance.mjs\`
-reports \`no-numbers-gate\` for any paper that has not.
+**Adopting it in another paper:** copy \`repro/paper_numbers.py\`, \`repro/paper_numbers.selftest.py\`
+and the build step that calls \`expand()\` on \`paper.tex\` before \`pdflatex\`; generate the
+grandfather list from the current draft; register the numbers a reviewer will actually check first.
 
 🔴 **Do not write the substitution half — it is \`jinja2\`, and \`{{name}}\` is already its syntax.**
-Eleven candidates were installed and run on a real markdown-source ACL paper on 2026-08-05. The
-result is not a matter of taste:
+Eleven candidates were installed and run on a real ACL paper on 2026-08-05. The result is not a
+matter of taste:
 
 | candidate | verdict, from an actual run |
 |---|---|
 | **jinja2** | **adopt.** Our \`{{annotated.failRate}}\` was already valid Jinja2; dotted names are attribute access. Zero characters changed in the paper, byte-identical \`.tex\`. This is Manubot's mechanism, taken without Manubot. Use \`StrictUndefined\`, and move the block/comment delimiters off \`{%\`/\`{#\` if the paper might ever quote a config file |
 | **knitr** | equally proven, also byte-identical. Rejected only for putting an R runtime into an artifact that otherwise needs \`python3\` + \`pdflatex\` |
-| \`\\newcommand\` + \`numbers.tex\` | works (1 line in the converter); the folk convention, ~12 GitHub hits, all hand-rolled. No package exists |
-| **Quarto**, **pandoc lua filters** | ❌ **destroy the paper.** Markdown→markdown is a pandoc AST round-trip: HTML comments become raw blocks and \`[58]\` becomes \`\\[58\\]\`, so every citation renders as literal \`\\textbackslash{}[58]\`. 3,120 diff lines on the real source. No ACL format exists for Quarto either |
+| \`\\newcommand\` + \`numbers.tex\` | works; the folk convention, ~12 GitHub hits, all hand-rolled. No package exists |
 | **cog.py** | ❌ line-oriented; cannot substitute mid-sentence |
 | **pythontex** | ❌ not in Ubuntu TeX Live at any level, not on PyPI; CTAN zip + \`.ins\` build |
-| **showyourwork** | ❌ LaTeX-only, and it manages *figures*, not in-text numbers |
-| **Pweave** | ❌ dead — the one candidate with a markdown native format, broken against modern IPython |
+| **showyourwork** | ❌ it manages *figures*, not in-text numbers |
 | DVC · datalad · papermill | orthogonal and **worth adding**: they pin the data file to the pipeline that made it, a layer below anything here |
 
 **🔴 Computing the value DISSOLVES the freshness problem — do not build a checker for it.** With no

package/skills/paper-pipeline/description-language.eval.mjs

@@ -1,6 +1,11 @@
 /**
  * description-language.eval.mjs — does adding Russian to a skill's DESCRIPTION close the language gap?
  *
+ * ⚠️ THE FIXTURE CHANGED AFTER THE RECORDED RUNS. The paper the model is shown is now `paper.tex`;
+ * every result recorded for this eval (here and in `repro/`) was measured on the previous fixture,
+ * the same paper in a source form paperlint no longer reads. The next run is a fresh, deliberate
+ * measurement — record it, do not compare it against those numbers.
+ *
  * Run:  node .claude/skills/paper-pipeline/description-language.eval.mjs --arm before|after [--trials N]
  *
  *   --arm before   measure with the descriptions AS THEY ARE (English only)   <- run this FIRST
@@ -119,42 +124,47 @@ if (ARM !== "before" && ARM !== "after")
 // Byte-identical to the other two evals. If these drift, a difference between runs stops being
 // attributable to the descriptions.
 const FIXTURE = {
-  "paper/paper.md": [
-    "# Prose Isn't Policy: Measuring Whether Agent-Config Rules Are Enforceable",
+  "paper/paper.tex": [
+    "\\documentclass{article}",
+    "\\title{Prose Isn't Policy: Measuring Whether Agent-Config Rules Are Enforceable}",
+    "\\begin{document}",
+    "\\maketitle",
     "",
-    "## Abstract",
+    "\\begin{abstract}",
     "Agent configuration files state rules in prose and assume the model obeys them. We compile a",
-    "corpus of real rules and measure what fraction can be mechanically enforced. We find that 84%",
+    "corpus of real rules and measure what fraction can be mechanically enforced. We find that 84\\%",
     "of rules in our corpus are enforceable, and that LLM-authored checkers for the remainder leak",
-    "silently in 84-96% of adversarial cases \\cite{greshake2023}.",
+    "silently in 84--96\\% of adversarial cases~\\cite{greshake2023}.",
+    "\\end{abstract}",
     "",
-    "## 1 Introduction",
+    "\\section{Introduction}",
     "Every agent harness ships a natural-language rulebook. Nothing checks it. This is the same",
-    "mistake as a code comment that claims an invariant no test enforces \\cite{thompson1984}.",
+    "mistake as a code comment that claims an invariant no test enforces~\\cite{thompson1984}.",
     "",
-    "## 2 Method",
+    "\\section{Method}",
     "We gather rules from public repositories, classify each by enforceability, and build a",
     "two-stage adversarial gate that validates a synthesized rule against a blind gold set.",
     "",
-    "## 3 Results",
-    "See Table 1. The headline number is 84%.",
+    "\\section{Results}",
+    "See Table~1. The headline number is 84\\%.",
     "",
-    "## 4 Discussion",
+    "\\section{Discussion}",
     "The result generalizes beyond our corpus in the sense that the mechanism is not corpus-specific,",
     "though of course the specific percentages are, and it is important to note in this context that",
     "the framing itself may be what carries, rather than the measurement.",
     "",
-    "## 5 Threats to Validity",
+    "\\section{Threats to Validity}",
     "Our corpus is drawn from public repositories and may not represent private configurations.",
     "",
-    "## 6 Related Work",
+    "\\section{Related Work}",
     "TODO",
     "",
-    "## 7 Conclusion",
+    "\\section{Conclusion}",
     "Prose is not policy.",
+    "\\end{document}",
   ].join("\n"),
   "paper/repro/README.md":
-    "# Reproduction artifact\n\n`python3 paper_numbers.py` recomputes every bolded figure in paper.md.\n",
+    "# Reproduction artifact\n\n`python3 paper_numbers.py` recomputes every bolded figure in paper.tex.\n",
   "paper/PIPELINE-STATUS.md":
     "# Pipeline status\n\n| gate | state |\n|---|---|\n| numbers | pass |\n| structure | not run |\n| citations | not run |\n",
 };