paperlint 6.2.2 → 6.2.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "paperlint",
-  "version": "6.2.2",
+  "version": "6.2.3",
   "description": "A linter for scientific papers written in LaTeX: catches mistakes before you submit to a conference or journal.",
   "type": "module",
   "license": "MIT",

package/skills/verify-citations/SKILL.md

@@ -4,7 +4,7 @@ description: Verify every citation is a real work with correct metadata, and tha
 allowed-tools: [Read, Write, Edit, Grep, Glob, Bash, WebSearch, WebFetch, Agent]
 ---
 
-<!-- vigiles:sha256:fe3634eb63db5c0d compiled from skills/verify-citations/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:faf8d55c8ddb82ed compiled from skills/verify-citations/SKILL.md.spec.ts -->
 
 # verify-citations — every cite real, the delta explicit
 
@@ -102,7 +102,7 @@ not-found, not mistaken for an unrelated paper.
 | verdict | when | meaning |
 |---|---|---|
 | `true` | any resolver **matched** (match always wins) | found, metadata matches |
-| `false` | no match **and** an *authoritative* disproof: doi.org says the DOI is `responseCode 100`, the reachable arXiv API definitively rejects the id, NVD has no such CVE, **or** a DOI/arXiv id resolves to a *confidently-different* paper (`DOI_MISMATCH`) | fabrication — the only thing we call fabrication |
+| `false` | no match **and** an *authoritative* disproof: doi.org says the DOI is `responseCode 100`, the reachable arXiv API definitively rejects the id, NVD has no such CVE, **or** the id's own registry (arXiv for an arXiv id or arXiv DOI, Crossref for any other DOI) puts it on a *confidently-different* paper (`DOI_MISMATCH`) | fabrication — the only thing we call fabrication |
 | `unresolvable` | title-misses, registry-404s with no authority disproof, uncomparable titles, and/or unreachable resolvers | couldn't verify, but **no positive disproof** |
 
 The whole point: **a title you simply can't find is NOT fabrication** → `unresolvable`, never `false`. A
@@ -217,13 +217,17 @@ by design, the gate **declines** in three situations rather than risk a false ac
   emit `DOI_MISMATCH` — the id counts as resolved (`true`) or, absent other confirmation, `unresolvable`.
   Never `false`. The `doi_mismatch → false` verdict fires **only** for two confidently-comparable titles that
   are genuinely different.
+- **An aggregator (OpenAlex, Semantic Scholar) that alone puts the id on another paper** →
+  `unresolvable` with `DOI_MISMATCH unconfirmed`. Aggregators copy metadata and do attach DOIs to the
+  wrong work; on 2026-10-01 OpenAlex mapped the arXiv DOIs of SWE-bench and ReAct to unrelated papers while
+  arXiv was rate-limiting. Only the id's own registry can disprove it.
 - **A DOI that resolves at doi.org but has no content-registry metadata** (DataCite/Zenodo, fresh 2026 DOIs)
   → `unresolvable`, because the work exists but its metadata (title/year) cannot be cross-checked here. This
   is the correct verdict for a paper's own artifact DOI; treat `unresolvable` as advisory, not a failure.
 
 **Known narrow residue (rare, treat a lone `DOI_MISMATCH → false` as advisory, eyeball it).** Two edge cases
-can still mis-fire: (a) a correct DOI whose *every* registry records junk metadata — e.g. the container /
-proceedings title instead of the paper's — reads as "unrelated" (mitigated when any one registry has the real
+can still mis-fire: (a) a correct DOI whose *own* registry records junk metadata — e.g. the container /
+proceedings title instead of the paper's — reads as "unrelated" (mitigated when any other registry has the real
 title, since a match wins); (b) a translated-journal title in the *same script* (transliterated Latin vs
 English) that shares no tokens. Both need semantics the gate doesn't have; both are near-zero frequency in an
 English CS bib. Identifier-prefix / trailing-punctuation mangling (`doi:`, `arXiv:`, a bib-swallowed `.`) is

package/skills/verify-citations/SKILL.md.spec.ts

@@ -118,7 +118,7 @@ not-found, not mistaken for an unrelated paper.
 | verdict | when | meaning |
 |---|---|---|
 | \`true\` | any resolver **matched** (match always wins) | found, metadata matches |
-| \`false\` | no match **and** an *authoritative* disproof: doi.org says the DOI is \`responseCode 100\`, the reachable arXiv API definitively rejects the id, NVD has no such CVE, **or** a DOI/arXiv id resolves to a *confidently-different* paper (\`DOI_MISMATCH\`) | fabrication — the only thing we call fabrication |
+| \`false\` | no match **and** an *authoritative* disproof: doi.org says the DOI is \`responseCode 100\`, the reachable arXiv API definitively rejects the id, NVD has no such CVE, **or** the id's own registry (arXiv for an arXiv id or arXiv DOI, Crossref for any other DOI) puts it on a *confidently-different* paper (\`DOI_MISMATCH\`) | fabrication — the only thing we call fabrication |
 | \`unresolvable\` | title-misses, registry-404s with no authority disproof, uncomparable titles, and/or unreachable resolvers | couldn't verify, but **no positive disproof** |
 
 The whole point: **a title you simply can't find is NOT fabrication** → \`unresolvable\`, never \`false\`. A
@@ -233,13 +233,17 @@ by design, the gate **declines** in three situations rather than risk a false ac
   emit \`DOI_MISMATCH\` — the id counts as resolved (\`true\`) or, absent other confirmation, \`unresolvable\`.
   Never \`false\`. The \`doi_mismatch → false\` verdict fires **only** for two confidently-comparable titles that
   are genuinely different.
+- **An aggregator (OpenAlex, Semantic Scholar) that alone puts the id on another paper** →
+  \`unresolvable\` with \`DOI_MISMATCH unconfirmed\`. Aggregators copy metadata and do attach DOIs to the
+  wrong work; on 2026-10-01 OpenAlex mapped the arXiv DOIs of SWE-bench and ReAct to unrelated papers while
+  arXiv was rate-limiting. Only the id's own registry can disprove it.
 - **A DOI that resolves at doi.org but has no content-registry metadata** (DataCite/Zenodo, fresh 2026 DOIs)
   → \`unresolvable\`, because the work exists but its metadata (title/year) cannot be cross-checked here. This
   is the correct verdict for a paper's own artifact DOI; treat \`unresolvable\` as advisory, not a failure.
 
 **Known narrow residue (rare, treat a lone \`DOI_MISMATCH → false\` as advisory, eyeball it).** Two edge cases
-can still mis-fire: (a) a correct DOI whose *every* registry records junk metadata — e.g. the container /
-proceedings title instead of the paper's — reads as "unrelated" (mitigated when any one registry has the real
+can still mis-fire: (a) a correct DOI whose *own* registry records junk metadata — e.g. the container /
+proceedings title instead of the paper's — reads as "unrelated" (mitigated when any other registry has the real
 title, since a match wins); (b) a translated-journal title in the *same script* (transliterated Latin vs
 English) that shares no tokens. Both need semantics the gate doesn't have; both are near-zero frequency in an
 English CS bib. Identifier-prefix / trailing-punctuation mangling (\`doi:\`, \`arXiv:\`, a bib-swallowed \`.\`) is

package/skills/verify-citations/scripts/verify-cites.mjs

@@ -159,6 +159,15 @@ export function normalizeArxiv(arxiv) {
 /** arXiv registers a DataCite DOI for every paper: `10.48550/arXiv.<id>`. */
 const ARXIV_DOI = /^10\.48550\/arxiv\.(.+)$/i;
 
+/** The one source whose record of an identifier is the identifier's own metadata: arXiv for an
+ *  arXiv id or an arXiv DOI, Crossref for any other DOI. OpenAlex and Semantic Scholar are
+ *  aggregators: they copy that metadata and can attach a DOI to the wrong work (OpenAlex did, for
+ *  SWE-bench and ReAct, on 2026-10-01), so a title they disagree on is not proof of anything. */
+function identifierAuthority(citation, query) {
+  if (query === "arxiv") return "arxiv";
+  return citation.doi && ARXIV_DOI.test(citation.doi) ? "arxiv" : "crossref";
+}
+
 /** Shallow copy of a citation with doi/arxiv identifiers normalized. Applied at the
  *  verify boundary so BOTH JSON input and .bib-parsed input are clean before any
  *  network / authority lookup.
@@ -367,6 +376,7 @@ export function classifyResolver(citation, response) {
           db,
           status: "doi_mismatch",
           foundTitle: response.record.title,
+          authoritative: identifierAuthority(citation, response.query) === db,
         };
       }
     }
@@ -473,8 +483,9 @@ export function checkCommit(citation) {
  * commit flags, produce the final verdict.
  *
  *   true          iff ANY evidence is `matched`  (a match always WINS)
- *   false         iff NOT matched AND ≥1 evidence is `id_unmatched` or `doi_mismatch`
- *                 (a resolvable id that fails, or a DOI on the wrong paper)
+ *   false         iff NOT matched AND ≥1 evidence is an authoritative `id_unmatched` or
+ *                 `doi_mismatch` (a resolvable id that fails, or an id that its own registry
+ *                 puts on the wrong paper)
  *   unresolvable  otherwise (only `title_miss` and/or `unreachable` — no disproof)
  */
 export function reduceVerdict(citation, evidence, commitFlags = []) {
@@ -500,8 +511,13 @@ export function reduceVerdict(citation, evidence, commitFlags = []) {
     );
   }
 
-  // A DOI/arXiv id that resolves to a CONFIDENTLY-DIFFERENT paper → misdirection.
-  const mismatch = ev.find((e) => e.status === "doi_mismatch");
+  // A DOI/arXiv id that resolves to a CONFIDENTLY-DIFFERENT paper AT ITS AUTHORITY → misdirection.
+  const mismatch = ev.find(
+    (e) => e.status === "doi_mismatch" && e.authoritative,
+  );
+  const aggregatorMismatch = ev.find(
+    (e) => e.status === "doi_mismatch" && !e.authoritative,
+  );
   if (mismatch) {
     return out(
       "false",
@@ -553,7 +569,9 @@ export function reduceVerdict(citation, evidence, commitFlags = []) {
   // ── No positive disproof → unresolvable, with the most informative reason ──
   const reachable = ev.some((e) => e.status !== "unreachable");
   let reason;
-  if (doiIdMiss && authorityPresent) {
+  if (aggregatorMismatch) {
+    reason = `DOI_MISMATCH unconfirmed: ${aggregatorMismatch.db} maps the identifier to "${aggregatorMismatch.foundTitle}" (claimed "${citation.title}"), but ${aggregatorMismatch.db} only aggregates metadata and the identifier's own registry did not confirm it — eyeball the entry → unresolvable, not false`;
+  } else if (doiIdMiss && authorityPresent) {
     reason = `DOI ${citation.doi} resolves at doi.org but no content registry (Crossref/OpenAlex/Semantic Scholar) has metadata for it — common for DataCite/Zenodo and freshly-minted DOIs; the work EXISTS, its metadata is just unindexed → unresolvable, NOT fabrication`;
   } else if (doiIdMiss) {
     reason = `DOI ${citation.doi} not found in any content registry, and the DOI authority (doi.org) was not reachable to confirm non-existence — cannot assert fabrication → unresolvable, never false`;

package/skills/verify-citations/scripts/verify-cites.test.mjs

@@ -104,17 +104,88 @@ console.log(
     title: "A Formal Verification Framework for Autonomous Agents",
     year: 2021,
   };
-  const openalex = {
-    db: "openalex",
+  const wrongRecord = {
+    title: "Photosynthesis in Deep-Sea Bacteria",
+    year: 2009,
+  };
+  const crossref = {
+    db: "crossref",
     transport: "ok",
     query: "doi",
-    record: { title: "Photosynthesis in Deep-Sea Bacteria", year: 2009 },
+    record: wrongRecord,
   };
-  const ev = classifyResolver(cite, openalex);
+  const ev = classifyResolver(cite, crossref);
   eq(ev.status, "doi_mismatch", "classify → doi_mismatch");
   const v = reduceVerdict(cite, [ev], checkCommit(cite));
-  eq(v.verdict, "false", "verdict false");
+  eq(v.verdict, "false", "the DOI's own registry (Crossref) disagrees → false");
   ok(/DOI_MISMATCH/.test(v.reason), "reason names DOI_MISMATCH");
+
+  // An aggregator alone is not proof: it can attach a DOI to the wrong work.
+  const openalex = {
+    db: "openalex",
+    transport: "ok",
+    query: "doi",
+    record: wrongRecord,
+  };
+  const agg = reduceVerdict(cite, [classifyResolver(cite, openalex)]);
+  eq(
+    agg.verdict,
+    "unresolvable",
+    "only OpenAlex disagrees → unresolvable, not false",
+  );
+  ok(
+    /DOI_MISMATCH unconfirmed/.test(agg.reason),
+    "reason says the mismatch is unconfirmed",
+  );
+}
+
+// ── (2b) 2026-10-01: OpenAlex mapped the SWE-bench arXiv DOI to another paper while
+//         arXiv answered 429. The entry is correct; it must not be called fabricated.
+console.log(
+  "\n(2b) aggregator mismatch on an arXiv DOI, arXiv rate-limited → unresolvable:",
+);
+{
+  const cite = normalizeIdentifiers({
+    id: "jimenez2024swebench",
+    doi: "10.48550/arXiv.2310.06770",
+    title: "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?",
+    year: 2024,
+  });
+  const openalex = {
+    db: "openalex",
+    transport: "ok",
+    query: "doi",
+    record: {
+      title: "GardenBench: A lightweight, daily evaluation of LLM capabilities",
+      year: 2025,
+    },
+  };
+  const arxiv = {
+    db: "arxiv",
+    transport: "error",
+    query: "arxiv",
+    record: null,
+  };
+  const s2 = {
+    db: "semantic_scholar",
+    transport: "error",
+    query: "doi",
+    record: null,
+  };
+  const v = reduceVerdict(
+    cite,
+    [openalex, arxiv, s2].map((r) => classifyResolver(cite, r)),
+  );
+  eq(
+    v.verdict,
+    "unresolvable",
+    "correct entry with a corrupted aggregator record → unresolvable",
+  );
+
+  // …and arXiv itself putting the id on another paper is still misdirection.
+  const arxivWrong = { ...arxiv, transport: "ok", record: openalex.record };
+  const w = reduceVerdict(cite, [classifyResolver(cite, arxivWrong)]);
+  eq(w.verdict, "false", "arXiv (the id's registry) disagrees → false");
 }
 
 // ── (3) made-up arXiv id that resolves to nothing → false ────────────────────