paperlint 7.2.0 → 8.0.0

package/skills/paper-pipeline/grade-paper-writing-ablation.eval.mjs

@@ -1,6 +1,11 @@
 /**
  * grade-paper-writing-ablation.eval.mjs — WHICH PART OF A DESCRIPTION KILLS A SKILL?
  *
+ * ⚠️ 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/grade-paper-writing-ablation.eval.mjs [--trials N] [--mode main|oracle|preflight|setupdiff] [--arms A0,A6]
  *
  *   --mode preflight   every guard + the arm-by-arm word-set deltas; spends NO tokens   <- run this FIRST
@@ -415,42 +420,47 @@ if (!["main", "oracle", "preflight", "setupdiff"].includes(MODE))
 // Byte-identical to the parent's (and therefore to both of ITS siblings). If it drifts, a difference
 // between this run and the 12% it exists to explain 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",
 };

package/skills/paper-pipeline/pipeline-firing.eval.mjs

@@ -1,6 +1,11 @@
 /**
  * pipeline-firing.eval.mjs — do the paper-pipeline skills actually FIRE when they should?
  *
+ * ⚠️ 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/pipeline-firing.eval.mjs [flags]
  *   --only <skill>        run one case (repeatable: --only tighten-paper --only argument-arc)
  *   --trials N            trials per prompt (default 1)
@@ -117,7 +122,8 @@
  *
  * ─────────────────────────────────────────────────────────────────────────────
  * FIRST RUN — 2026-08-07, sonnet, 64 runs, 1 trial/prompt, 36 competitors, ~$9.55 API-equivalent
- * (billed to a Claude subscription, $0 metered). Recorded in pipeline-firing.baseline.json.
+ * (billed to a Claude subscription, $0 metered). Its baseline file was dropped when the fixture
+ * changed; `--update-baseline` records the next one.
  *
  *   skill                      recall  FP-rate  precision   n
  *   tighten-paper                75%     0%      100%       4
@@ -235,42 +241,47 @@ const GATE = !flag("no-gate");
 // therefore a LOWER bound, and a low number for that case is partly an artifact of this gap, not
 // necessarily a description defect. Do not report it as one.
 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",
 };

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

@@ -1,6 +1,11 @@
 /**
  * pipeline-language.eval.mjs — does prompt LANGUAGE change whether a skill fires?
  *
+ * ⚠️ 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/pipeline-language.eval.mjs [flags]
  *   --trials N        trials per prompt (default 3)
  *   --concurrency N   parallel runs (default 3)
@@ -184,42 +189,47 @@ const CONCURRENCY = Number(val("concurrency", "3"));
 // Byte-identical to the sibling eval's FIXTURE on purpose. If the two files drift, a difference
 // between this run and that one stops being attributable to language.
 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",
 };

package/skills/paper-pipeline/references/acceptance-gate.md

@@ -70,8 +70,8 @@ drive**.
 
 "The PDF compiled" needs no taste. "The PDF is a good paper" is nothing but taste. "The number
 recomputes from the committed data" needs no taste. "The number supports the sentence around it"
-needed taste right up until we built `check-provenance.mjs`, which is the interesting move — see the
-fourth rung.
+needed taste right up until we built the numbers gate's `requires`/`forbids` guards, which is the
+interesting move — see the fourth rung.
 
 ### Compound gates: split, never average
 
@@ -119,7 +119,7 @@ it caught a defect the judgement gates had passed:
 
 | the taste question                              | the surrogate we built                                                                              | what it caught                                                                                 |
 | ----------------------------------------------- | --------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- |
-| does this number support this sentence          | `check-provenance.mjs` — arm word in the data path vs condition word in the sentence                | the 2026-08-05 abstract defect, 29 hours before a deadline; five judgement gates had passed it |
+| does this number support this sentence          | `repro/numbers.tsv` guards — the arm word the paragraph must name, the condition it must not claim  | the 2026-08-05 abstract defect, 29 hours before a deadline; five judgement gates had passed it |
 | is this quantity really bound to its data       | `repro/arm_permutation.py` — permute the arm labels, rebuild, an unchanged quantity was never bound | a whole class the artifact reviewer recomputes and passes, because the number is _correct_     |
 | did the number survive into the delivered paper | `repro/delivered_pdf.py` — read the built PDF back                                                  | a superscript minus lost in typesetting                                                        |
 | is this reference real                          | `extract-ref-facts.mjs` + rules `refs/*` — arXiv / CrossRef                                         | nine fabricated titles carrying correct arXiv ids, and one invented award                      |
@@ -248,21 +248,19 @@ Everything in `.claude/skills/paper-pipeline/scripts/`, `.claude/lib/` and `<pap
 construction — that is why it is trustworthy, and it is not a coincidence but a selection effect: a
 program is what you can write only when the question is Type A.
 
-| unit                                                 | question                                                                                 | rung                                                                                                                  |
-| ---------------------------------------------------- | ---------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- |
-| `repro/report-submission.py`                         | body pages, overfull boxes, unresolved refs, dropped characters                          | A (3)                                                                                                                 |
-| `repro/check-anon.sh`, `check-figure.sh`             | deanonymising strings; does the figure fit the column                                    | A (3)                                                                                                                 |
-| `structure.mjs`, `prose-lint.mjs`                    | section weights, unjustified blocks, thresholds with published baselines                 | A (3) — _measuring_ a B question, which is legitimate and is not the same as answering it                             |
-| `check-provenance.mjs`                               | untraced number; arm word in the path vs condition word in the sentence                  | A (3)                                                                                                                 |
-| `population-map.mjs`, `artifact-coverage.mjs`        | which set does this number count; does the bundle hold data for what the paper points at | A (3)                                                                                                                 |
-| rules `refs/*` (facts by `extract-ref-facts.mjs`)    | does the registry have this record                                                       | A (3), **external adjudicator**                                                                                       |
-| `paper-lint.mjs`                                     | appendix ratio, shaved passages                                                          | A (3)                                                                                                                 |
-| `repro/arm_permutation.py`, `repro/delivered_pdf.py` | is the quantity bound to its arm; did it survive typesetting                             | A (3), metamorphic                                                                                                    |
-| **`repro/paper_numbers.py` + `numbers.tsv`**         | —                                                                                        | **rung 1: no PDF exists in which the number is wrong**                                                                |
-| `paper-edit-guard.hook.ts`                           | writing to a paper from Bash                                                             | rung 2 — **leaky**, an interpreter reading its program from stdin still gets through (`papers/CLAUDE.md`, 2026-08-06) |
-
-`structure.mjs` and `prose-lint.mjs` deserve their note. They put a **number with a threshold** on
-questions that are Type B — does this section earn its place, is this sentence readable. That is not
+| unit                                                 | question                                                        | rung                                                                                                                  |
+| ---------------------------------------------------- | --------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- |
+| `repro/report-submission.py`                         | body pages, overfull boxes, unresolved refs, dropped characters | A (3)                                                                                                                 |
+| `repro/check-anon.sh`, `check-figure.sh`             | deanonymising strings; does the figure fit the column           | A (3)                                                                                                                 |
+| `prose-lint.mjs`                                     | thresholds with published baselines                             | A (3) — _measuring_ a B question, which is legitimate and is not the same as answering it                             |
+| rules `refs/*` (facts by `extract-ref-facts.mjs`)    | does the registry have this record                              | A (3), **external adjudicator**                                                                                       |
+| `paper-lint.mjs`                                     | appendix ratio, shaved passages                                 | A (3)                                                                                                                 |
+| `repro/arm_permutation.py`, `repro/delivered_pdf.py` | is the quantity bound to its arm; did it survive typesetting    | A (3), metamorphic                                                                                                    |
+| **`repro/paper_numbers.py` + `numbers.tsv`**         | —                                                               | **rung 1: no PDF exists in which the number is wrong**                                                                |
+| `paper-edit-guard.hook.ts`                           | writing to a paper from Bash                                    | rung 2 — **leaky**, an interpreter reading its program from stdin still gets through (`papers/CLAUDE.md`, 2026-08-06) |
+
+`prose-lint.mjs` deserves its note. It puts a **number with a threshold** on a question that is
+Type B — is this sentence readable. That is not
 cheating and it is not acquittal either: a threshold breach is a fact, and the fact drives a human or
 a judgement pass. The failure to avoid is reading a green threshold as an acquittal, which is what
 happened when a prose metric stood at 50.8 against a machine norm of 7.8 and printed as a neutral
@@ -318,7 +316,5 @@ questions into programs.
 ## See also
 
 - `review-ratchet.md` — what a caught overclaim costs, and the pass that pays it back.
-- `../scripts/round-diff.mjs` — the per-round edit gate: a driving pass may only change what it
-  declared, and the whole document's growth is measured every round.
 - the papers tree's own `CLAUDE.md` — the four-rung ladder, and the rule that a rule written in prose is
   a construction somebody did not build.

package/skills/paper-pipeline/references/artifact-checklist.md

@@ -65,15 +65,14 @@ Observed: a paper's released bundle contained **no data at all** for the section
 with. Six `harden-paper` passes and five review panels had not noticed. It surfaced because the
 author asked a plain question — _"did we ever actually run the artifact?"_
 
-**Run `paper-pipeline/scripts/artifact-coverage.mjs <paper-dir>` — wired into pre-commit and CI.**
-Two checks, deliberately dumb:
+**Check two things, deliberately dumb:**
 
 1. every body section that prints a figure is named in the bundle's index (`NUMBERS.md`);
 2. every path the index names exists in the bundle.
 
-Its first run found **three** gaps: one real absence, and two experiments whose data shipped but
-which the index never named — so a reviewer following the index would have concluded they were
-unreleased. Absence and unfindability fail the same way for a reader.
+The first time they were checked they found **three** gaps: one real absence, and two experiments
+whose data shipped but which the index never named — so a reviewer following the index would have
+concluded they were unreleased. Absence and unfindability fail the same way for a reader.
 
 ## Running `aclpubcheck` locally
 

package/skills/paper-pipeline/references/pipeline-status-template.md

@@ -103,7 +103,7 @@ existed. Neither is expressible now. Keep new ids lowercase, one word, and say w
 ---
 
 The template itself is a FILE, not this page: `templates/paper/PIPELINE-STATUS.md` at the package
-root (beside `templates/paper/paper.tex` and `paper.md`). `paperlint new <name>` copies it with the paper's
+root (beside `templates/paper/paper.tex`). `paperlint new <name>` copies it with the paper's
 name filled in, and a project that keeps its own richer scorecard puts one at
 `<papers>/.template/PIPELINE-STATUS.md`, which `paperlint new` prefers. The copy carries no `stages` — a new paper has shipped nothing.
 

package/skills/paper-pipeline/scripts/README.md

@@ -7,12 +7,8 @@ installed package — see `consumer.mjs`).
 
 | file                    | what it answers                                                                                       |
 | ----------------------- | ----------------------------------------------------------------------------------------------------- |
-| `artifact-coverage.mjs` | does the released bundle actually contain what the paper points at?                                   |
-| `check-provenance.mjs`  | does the paper have a build-time numbers gate, and what does it still not cover?                      |
 | `extract-ref-facts.mjs` | parse the bibliography, ask the registries, write facts to JSON for the `refs/*` ESLint rules         |
 | `generated-code.mjs`    | lint the analysis scripts before they produce a number (unseeded randomness, absolute paths, …)       |
-| `population-map.mjs`    | does the population registry keep its link to the body of the paper?                                  |
-| `round-diff.mjs`        | did a review round change only what it declared?                                                      |
 | `consumer.mjs`          | where the consumer's repository is, and whether a script was executed or imported through the symlink |
 
 The checkers with a `--flags-only` mode print nothing on a clean paper and exit 0 — that mode is

package/skills/paper-pipeline/scripts/extract-ref-facts.harness.mjs

@@ -308,71 +308,12 @@ process.on("exit", () => rmSync(TMP, { recursive: true, force: true }));
   );
 }
 
-// ── 7. markdown: reference list, source lines, `et al.`, appendix ───
-{
-  const md = [
-    "# Paper",
-    "",
-    "## 1. Intro",
-    "",
-    "Prose.",
-    "",
-    "## References",
-    "",
-    "1. C. Yang, Z. Zhao, L. Zhang. *KNighter: Transforming Static Analysis.* arXiv:2503.09002, 2025.",
-    "2. Z. Xiang et al. *Another Work.* ICSE 2024. arXiv:2401.00001",
-    "",
-    "## Appendix A",
-    "",
-    "1. Not a reference at all.",
-    "",
-  ].join("\n");
-  const es = X.parseMarkdownRefs(md);
-  assert.equal(
-    es.length,
-    2,
-    `heading must close the reference list, parsed ${es.length}`,
-  );
-  assert.equal(
-    es[0].line,
-    9,
-    `line of first entry is ${es[0].line}, but it is ninth in file`,
-  );
-  assert.equal(es[0].title, "KNighter: Transforming Static Analysis");
-  assert.equal(es[1].truncated, true, "`et al.` must be marked as truncation");
-  assert.deepEqual(es[1].authors, ["Z. Xiang"]);
-  assert.equal(es[1].year, "2024");
-}
-
-// ── 8. `## References` inside ```-block does NOT open list ──────────────────
-// Papers in this repo cite their own markup in pieces; the old `split` would open
-// bibliography in the middle of a code example.
-{
-  const md = [
-    "# P",
-    "",
-    "## Body",
-    "",
-    "```md",
-    "## References",
-    "",
-    "1. Fake. *X.* arXiv:1111.11111",
-    "```",
-    "",
-  ].join("\n");
-  assert.deepEqual(
-    X.parseMarkdownRefs(md),
-    [],
-    "heading inside ```-block is not a heading",
-  );
-}
-
-// ── 9. CLI: ZERO ENTRIES — REFUSAL, not empty facts ─────────────────────────
+// ── 7. CLI: ZERO ENTRIES — REFUSAL, not empty facts ─────────────────────────
 {
   const { spawnSync } = await import("node:child_process");
   const dir = join(TMP, "empty");
   mkdirSync(dir, { recursive: true });
-  writeFileSync(join(dir, "paper.md"), "# P\n\nNo links at all.\n");
+  writeFileSync(join(dir, "refs.bib"), "% no entries at all\n");
   const r = spawnSync(
     "node",
     [join(HERE, "extract-ref-facts.mjs"), dir, "--offline"],
@@ -393,7 +334,7 @@ process.on("exit", () => rmSync(TMP, { recursive: true, force: true }));
   );
 }
 
-// ── 10. CLI: no source — also refusal, not silence ─────────────────────────
+// ── 8. CLI: no source — also refusal, not silence ─────────────────────────
 {
   const { spawnSync } = await import("node:child_process");
   const dir = join(TMP, "nosrc");
@@ -410,20 +351,15 @@ process.on("exit", () => rmSync(TMP, { recursive: true, force: true }));
   );
 }
 
-// ── 11. end-to-end: real paper → facts, freshness is computed ───────────────
+// ── 9. end-to-end: real paper → facts, freshness is computed ───────────────
 {
   const { spawnSync } = await import("node:child_process");
   const dir = join(TMP, "e2e");
   mkdirSync(join(dir, "repro"), { recursive: true });
-  const paper = [
-    "# P",
-    "",
-    "## References",
-    "",
-    "1. A. Lovelace, G. Hopper. *A Study.* arXiv:2503.09002, 2025.",
-    "",
-  ].join("\n");
-  writeFileSync(join(dir, "paper.md"), paper);
+  const paper =
+    "@misc{lovelace2025,\n  author={Lovelace, Ada and Hopper, Grace},\n  title={A Study},\n" +
+    "  note={arXiv:2503.09002},\n  year={2025}}\n";
+  writeFileSync(join(dir, "refs.bib"), paper);
   writeFileSync(
     join(dir, "repro", "refs-cache.json"),
     JSON.stringify({