paperlint 6.2.2 → 6.3.0

package/skills/submission-portal/SKILL.md

@@ -0,0 +1,69 @@
+---
+name: submission-portal
+description: "Read or change a paper's submission on the venue's HotCRP portal via `paperlint submission show|update`: replace the PDF or abstract, mark it submitted, check it is really \"Submitted\", and verify by sha256 that the portal holds the latest build. Use when the PDF or abstract on HotCRP needs replacing, when asked whether the paper is truly submitted or which PDF reviewers see, or before a deadline. Composes with submit-paper (what to submit); this gets it onto the portal and proves it landed."
+allowed-tools: [Read, Grep, Glob, Bash]
+---
+
+<!-- vigiles:sha256:deb3d74f5c3f93a8 compiled from skills/submission-portal/SKILL.md.spec.ts -->
+
+# submission-portal — read and update the submission on the portal, then prove it landed
+
+`paperlint submission` talks to the venue's submission portal through its API. Only HotCRP is
+supported; for any other portal the command says so and the upload is done by hand.
+
+## What must be in place
+
+- **The venue preset declares the portal**: `"portal": { "kind": "hotcrp", "url": "<the site>" }`
+  (`paperlint:aidc` has it). For your own venue preset, add it there.
+- **The paper's `paperlint.json` names its submission**: `"submission": { "id": <number> }` — the
+  number HotCRP shows for the paper.
+- **A token in `HOTCRP_TOKEN`**, created by the author on the HotCRP site under Account settings →
+  Developer (scopes `submeta:read` and `document:read` to read; `submeta:write` and
+  `document:write` to update). Never ask for it in chat, never print it, never write it to a file:
+  the command reads it from the environment and sends it only as the `Authorization` header.
+
+A missing key or token stops the command with exit 2 and a line naming the file and key, or the
+variable — fix exactly that and run again.
+
+## The procedure — always in this order
+
+1. **Read it.** `npx paperlint submission show papers/<paper>` — status, title, type, topics,
+   abstract length, the PDF the portal holds, and `match`: whether that PDF's sha256 equals the
+   local `paper.pdf`. Rebuild first (`npx paperlint build papers/<paper>`) if the local PDF is stale.
+2. **Dry run.** `npx paperlint submission update papers/<paper>` — sends the PDF (add
+   `--abstract <file>` for a plain-text abstract, `--submit` to mark it submitted) with HotCRP's
+   `dry_run=1`: the portal checks the change and **keeps nothing**. Read `valid`, `changes` and every
+   `portal says:` line. Exit 1 means the portal would refuse it — report why, do not go on.
+3. **Ask the author.** Show the dry run's output and ask for an explicit yes to change the real
+   submission.
+4. **Save — only after that yes.** `npx paperlint submission update papers/<paper> --save` (same
+   flags as the dry run). 🔴 `--save` CHANGES A REAL SUBMISSION that reviewers will see. Never run it
+   without the author's explicit yes in this conversation, never "to test", and never because a
+   dry run was valid.
+5. **Prove it landed.** `npx paperlint submission show papers/<paper>` again: `match` must read
+   `yes — the portal holds this build`, and the status must be what was intended (`submitted` if
+   `--submit` was sent). Anything else is not done — say so.
+
+The same five steps as commands, in order (`<paper>` is the paper's folder):
+
+```bash
+npx paperlint submission show papers/<paper>
+npx paperlint submission update papers/<paper>
+# … the author reads the dry run and says yes …
+npx paperlint submission update papers/<paper> --save
+npx paperlint submission show papers/<paper>
+```
+
+## Reading the answers
+
+- `match yes` is the only proof the reviewers get the local build; a status of `submitted` alone
+  says nothing about WHICH PDF.
+- `the portal refused (HTTP 401): …` — the token is missing, wrong or expired; `403` — the token
+  lacks the scope, or the deadline has passed for this change. The text after the colon is
+  HotCRP's own; quote it.
+- `SAVE WITHHELD by the portal` — HotCRP did not save; nothing changed.
+
+## What this skill does not do
+
+It does not decide whether the paper is ready (that is `harden-paper`), anonymize it, or fill the
+submission form's other fields (authors, conflicts) — see `submit-paper`.

package/skills/submission-portal/SKILL.md.spec.ts

@@ -0,0 +1,71 @@
+// Compiled to SKILL.md by `vigiles compile`. Edit THIS file, never the markdown.
+import { experimental_skill } from "vigiles/spec";
+
+export default experimental_skill({
+  name: "submission-portal",
+  description:
+    "Read or change a paper's submission on the venue's HotCRP portal via `paperlint submission show|update`: replace the PDF or abstract, mark it submitted, check it is really \"Submitted\", and verify by sha256 that the portal holds the latest build. Use when the PDF or abstract on HotCRP needs replacing, when asked whether the paper is truly submitted or which PDF reviewers see, or before a deadline. Composes with submit-paper (what to submit); this gets it onto the portal and proves it landed.",
+  tools: ["Read", "Grep", "Glob", "Bash"],
+  body: `
+# submission-portal — read and update the submission on the portal, then prove it landed
+
+\`paperlint submission\` talks to the venue's submission portal through its API. Only HotCRP is
+supported; for any other portal the command says so and the upload is done by hand.
+
+## What must be in place
+
+- **The venue preset declares the portal**: \`"portal": { "kind": "hotcrp", "url": "<the site>" }\`
+  (\`paperlint:aidc\` has it). For your own venue preset, add it there.
+- **The paper's \`paperlint.json\` names its submission**: \`"submission": { "id": <number> }\` — the
+  number HotCRP shows for the paper.
+- **A token in \`HOTCRP_TOKEN\`**, created by the author on the HotCRP site under Account settings →
+  Developer (scopes \`submeta:read\` and \`document:read\` to read; \`submeta:write\` and
+  \`document:write\` to update). Never ask for it in chat, never print it, never write it to a file:
+  the command reads it from the environment and sends it only as the \`Authorization\` header.
+
+A missing key or token stops the command with exit 2 and a line naming the file and key, or the
+variable — fix exactly that and run again.
+
+## The procedure — always in this order
+
+1. **Read it.** \`npx paperlint submission show papers/<paper>\` — status, title, type, topics,
+   abstract length, the PDF the portal holds, and \`match\`: whether that PDF's sha256 equals the
+   local \`paper.pdf\`. Rebuild first (\`npx paperlint build papers/<paper>\`) if the local PDF is stale.
+2. **Dry run.** \`npx paperlint submission update papers/<paper>\` — sends the PDF (add
+   \`--abstract <file>\` for a plain-text abstract, \`--submit\` to mark it submitted) with HotCRP's
+   \`dry_run=1\`: the portal checks the change and **keeps nothing**. Read \`valid\`, \`changes\` and every
+   \`portal says:\` line. Exit 1 means the portal would refuse it — report why, do not go on.
+3. **Ask the author.** Show the dry run's output and ask for an explicit yes to change the real
+   submission.
+4. **Save — only after that yes.** \`npx paperlint submission update papers/<paper> --save\` (same
+   flags as the dry run). 🔴 \`--save\` CHANGES A REAL SUBMISSION that reviewers will see. Never run it
+   without the author's explicit yes in this conversation, never "to test", and never because a
+   dry run was valid.
+5. **Prove it landed.** \`npx paperlint submission show papers/<paper>\` again: \`match\` must read
+   \`yes — the portal holds this build\`, and the status must be what was intended (\`submitted\` if
+   \`--submit\` was sent). Anything else is not done — say so.
+
+The same five steps as commands, in order (\`<paper>\` is the paper's folder):
+
+\`\`\`bash
+npx paperlint submission show papers/<paper>
+npx paperlint submission update papers/<paper>
+# … the author reads the dry run and says yes …
+npx paperlint submission update papers/<paper> --save
+npx paperlint submission show papers/<paper>
+\`\`\`
+
+## Reading the answers
+
+- \`match yes\` is the only proof the reviewers get the local build; a status of \`submitted\` alone
+  says nothing about WHICH PDF.
+- \`the portal refused (HTTP 401): …\` — the token is missing, wrong or expired; \`403\` — the token
+  lacks the scope, or the deadline has passed for this change. The text after the colon is
+  HotCRP's own; quote it.
+- \`SAVE WITHHELD by the portal\` — HotCRP did not save; nothing changed.
+
+## What this skill does not do
+
+It does not decide whether the paper is ready (that is \`harden-paper\`), anonymize it, or fill the
+submission form's other fields (authors, conflicts) — see \`submit-paper\`.`,
+});

package/skills/submission-portal/submission-portal.eval.mjs

@@ -0,0 +1,15 @@
+/**
+ * submission-portal — the PAID tier: does this skill's description actually fire?
+ *
+ * The prompts live in `lib/skill-trigger-cases.mjs` beside every other skill's, so collisions with
+ * `submit-paper` and `osf-artifact-upload` are visible in one table. Measures recall on its own
+ * territory and precision against those siblings, against the real harness.
+ *
+ * ⚠️ NEVER RUN. Written 2026-10-03 with the skill; its rate is UNKNOWN, not assumed good.
+ *
+ * Costs money; not CI.
+ *   node skills/submission-portal/submission-portal.eval.mjs [trials]
+ */
+import { runSkillTriggerEval } from "../../lib/skill-eval-kit.mjs";
+
+await runSkillTriggerEval("submission-portal");

package/skills/submission-portal/submission-portal.test.ts

@@ -0,0 +1,81 @@
+/**
+ * submission-portal — the free tier: the skill's commands are commands the CLI accepts, in the
+ * order the skill promises, and the skill passes the shared per-skill checks. No model, no network,
+ * no portal.
+ *
+ * 🔴 WHY THE COMMANDS ARE PARSED, NOT GREPPED: a skill that tells the agent to run a flag the CLI
+ * refuses fails at the worst moment — before a deadline, on a real submission. So every
+ * `paperlint submission …` the skill writes goes through the CLI's own argument parser and its
+ * usage check, and must come out runnable.
+ */
+import assert from "node:assert/strict";
+import { readFileSync } from "node:fs";
+import { dirname, join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { test } from "vitest";
+import { commandsIn } from "vigiles";
+import { checkSkill } from "../../lib/skill-checks.mjs";
+import { parseArgs } from "../../src/cli.ts";
+import { usageProblem } from "../../src/submission.ts";
+
+const SKILL = readFileSync(
+  join(dirname(fileURLToPath(import.meta.url)), "SKILL.md"),
+  "utf8",
+);
+const commands = commandsIn(
+  SKILL,
+  /(^|[|;&]\s*)npx paperlint submission\b/,
+).map((c) => c.text);
+
+/** A command as the CLI receives it: the words after `npx paperlint`. */
+const argvOf = (command: string): string[] =>
+  command
+    .replace(/^.*?npx paperlint /, "")
+    .split(/\s+/)
+    .filter(Boolean);
+
+test("the skill names show, a dry run and a save — and nothing else", () => {
+  const kinds = commands.map((c) => {
+    const a = parseArgs(argvOf(c));
+    return `${String(a.paths[0])}${a.save ? " --save" : ""}`;
+  });
+  assert.deepEqual([...new Set(kinds)], ["show", "update", "update --save"]);
+});
+
+test("every `paperlint submission` command in the skill is one the CLI runs", () => {
+  assert.ok(
+    commands.length > 0,
+    "the skill names no `paperlint submission` command",
+  );
+  for (const c of commands) {
+    const a = parseArgs(argvOf(c));
+    assert.equal(a.cmd, "submission", c);
+    assert.equal(a.unknownFlag, undefined, c);
+    assert.equal(a.missingValue, undefined, c);
+    const problem = usageProblem({
+      sub: a.paths[0],
+      paperDir: a.paths[1] ?? ".",
+      extra: a.paths.slice(2),
+      pdf: a.pdf,
+      abstract: a.abstract,
+      submit: a.submit,
+      save: a.save,
+    });
+    assert.equal(problem, null, c);
+  }
+});
+
+test("🔴 the procedure reads, dry-runs, then saves, then reads again — never saves first", () => {
+  const order = commands.map((c) => {
+    const a = parseArgs(argvOf(c));
+    return a.paths[0] === "show" ? "show" : a.save ? "save" : "dry";
+  });
+  const firstSave = order.indexOf("save");
+  assert.ok(firstSave > order.indexOf("dry"), order.join(" → "));
+  assert.ok(order.indexOf("dry") > order.indexOf("show"), order.join(" → "));
+  assert.ok(order.lastIndexOf("show") > firstSave, order.join(" → "));
+});
+
+test("the shared per-skill checks pass", async () => {
+  await checkSkill("submission-portal");
+});

package/skills/submit-paper/SKILL.md

@@ -4,7 +4,7 @@ description: End-to-end playbook for submitting a peer-reviewed paper to a doubl
 allowed-tools: [Read, Write, Edit, Grep, Glob, Bash, WebSearch, WebFetch, Skill]
 ---
 
-<!-- vigiles:sha256:9e57bcb2ba1fb055 compiled from skills/submit-paper/SKILL.md.spec.ts -->
+<!-- vigiles:sha256:767a4757d1df67d3 compiled from skills/submit-paper/SKILL.md.spec.ts -->
 
 # submit-paper — get a reviewed paper from "done" to "ready for review"
 
@@ -110,6 +110,7 @@ Many workshop HotCRPs have **no supplementary-upload field** — check first; if
   is the step people miss.
 - You can edit until the deadline: use **Replace** on the Submission field to swap the PDF for a late
   fix (e.g. adding the artifact link after you host it). Submit early, upgrade before the deadline.
+- When the venue preset declares its HotCRP portal, replacing the PDF or abstract and proving the portal holds the latest build is `submission-portal` (`paperlint submission show|update`).
 
 ## 6. Independent-researcher / dossier notes
 - **Name consistency** is the quiet risk: publications, LinkedIn, passport, letters must resolve to one

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

@@ -126,6 +126,7 @@ Many workshop HotCRPs have **no supplementary-upload field** — check first; if
   is the step people miss.
 - You can edit until the deadline: use **Replace** on the Submission field to swap the PDF for a late
   fix (e.g. adding the artifact link after you host it). Submit early, upgrade before the deadline.
+- When the venue preset declares its HotCRP portal, replacing the PDF or abstract and proving the portal holds the latest build is \`submission-portal\` (\`paperlint submission show|update\`).
 
 ## 6. Independent-researcher / dossier notes
 - **Name consistency** is the quiet risk: publications, LinkedIn, passport, letters must resolve to one

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 ────────────────────

package/src/adapters/hotcrp/cassette/README.md

@@ -0,0 +1,24 @@
+# HotCRP cassette
+
+Real answers to the requests `paperlint submission` sends, recorded from a live HotCRP instance
+through HotCRP's documented API (`/api/paper`). `portal.test.ts` replays them from its
+local mock server and parses each one through the adapter's real zod schemas: a schema that rejects
+a real answer is the defect class these files exist to catch. A hand-written mock body is what let
+a wrong upload field name through once.
+
+Only read-only requests were recorded: one `GET` and two `POST`s with `dry_run=1` — nothing was
+saved. Identifying values (the submission id, title, abstract, authors, topics, the PDF's hash, the
+timestamps) are replaced with fakes; HotCRP's keys, value types, statuses and message texts are
+verbatim.
+
+| file              | request                                                                                                                                                   | HTTP |
+| ----------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------- | ---: |
+| `get-paper.json`  | `GET /api/paper?p=<id>&word_limit=hard` with a token                                                                                                      |  200 |
+| `get-noauth.json` | the same without a token                                                                                                                                  |  401 |
+| `post-dot.json`   | `POST /api/paper?p=<id>&dry_run=1`, PDF part named `paper.pdf`, `content_file` the same — PHP renames the field to `paper_pdf`, so HotCRP finds no upload |  200 |
+| `post-ok.json`    | the same with the part and `content_file` named `paperlint_pdf`                                                                                           |  200 |
+
+To re-record: run the same four requests with `curl` against a HotCRP site where you have a
+submission and a token (`Authorization: bearer $HOTCRP_TOKEN`), always with `dry_run=1` on a `POST`,
+save each body, then replace every identifying value with a fake of the same type before committing.
+These files ship in the npm package with the rest of `src/`, like the tests beside them.

package/src/adapters/hotcrp/cassette/get-noauth.json

@@ -0,0 +1,10 @@
+{
+  "ok": false,
+  "message_list": [
+    {
+      "status": 2,
+      "message": "<0>Missing credentials"
+    }
+  ],
+  "status_code": 401
+}

package/src/adapters/hotcrp/cassette/get-paper.json

@@ -0,0 +1,39 @@
+{
+  "ok": true,
+  "paper": {
+    "object": "paper",
+    "pid": 7,
+    "title": "An Example Paper Title",
+    "submission": {
+      "mimetype": "application/pdf",
+      "hash": "sha2-b3e602c135f2455f64aecfdc2e1e7f81dad20c5d69f0981e965b1a0131d47163",
+      "timestamp": 1790000000,
+      "size": 17,
+      "pages": 2
+    },
+    "abstract": "An example abstract of exactly ten words for the test.",
+    "contacts": [
+      {
+        "email": "author@example.org",
+        "given_name": "A.",
+        "family_name": "Author",
+        "affiliation": "Example University"
+      }
+    ],
+    "paper_type": "Example paper type",
+    "authors": [
+      {
+        "email": "author@example.org",
+        "given_name": "A.",
+        "family_name": "Author",
+        "affiliation": "Example University"
+      }
+    ],
+    "topics": ["Example topic one", "Example topic two"],
+    "pc_conflicts": {},
+    "status": "submitted",
+    "submitted": true,
+    "submitted_at": 1790000100,
+    "modified_at": 1790000100
+  }
+}

package/src/adapters/hotcrp/cassette/post-dot.json

@@ -0,0 +1,21 @@
+{
+  "ok": false,
+  "message_list": [
+    {
+      "status": 2,
+      "field": "submission",
+      "message": "<0>Ignored attempt to upload document without any content"
+    },
+    {
+      "status": 3,
+      "field": "submission"
+    },
+    {
+      "status": -3
+    }
+  ],
+  "dry_run": true,
+  "valid": false,
+  "change_list": [],
+  "pid": 7
+}

package/src/adapters/hotcrp/cassette/post-ok.json

@@ -0,0 +1,12 @@
+{
+  "ok": true,
+  "message_list": [
+    {
+      "status": -3
+    }
+  ],
+  "dry_run": true,
+  "valid": true,
+  "change_list": ["submission"],
+  "pid": 7
+}

package/src/adapters/hotcrp/index.ts

@@ -0,0 +1,11 @@
+/**
+ * HotCRP, as the `SubmissionPortal` port: its REST API read and written over `fetch`, its answers
+ * parsed here. Everything that changes if HotCRP's API does lives in this folder.
+ */
+export {
+  CONTENT_FIELD,
+  changeObject,
+  hotcrpPortal,
+  type HotcrpOptions,
+} from "./portal.io.ts";
+export { parseShow, parseUpdate, sha256OfHash } from "./response.ts";