diff --git a/.github/workflows/no-attribution.yml b/.github/workflows/no-attribution.yml
new file mode 100644
index 0000000..01fe484
--- /dev/null
+++ b/.github/workflows/no-attribution.yml
@@ -0,0 +1,27 @@
+name: No Attribution
+
+# This repository's own pull requests, checked by the action it publishes.
+#
+# It addresses the action by local path rather than by `HeroicLands/.github@main`
+# like every other repository does, for two reasons: a pull request that changes
+# the action is then checked by its own version of it rather than by the one on
+# `main`, and the action could not otherwise be introduced here without failing
+# the pull request that introduces it. The local path is what needs the checkout.
+
+on:
+  pull_request:
+    types: [opened, edited, reopened, synchronize]
+
+permissions:
+  contents: read
+  pull-requests: read
+
+jobs:
+  no-attribution:
+    name: No AI attribution
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+      - uses: ./actions/no-attribution
+        with:
+          token: ${{ github.token }}
diff --git a/README.md b/README.md
index 7ba458a..faef291 100644
--- a/README.md
+++ b/README.md
@@ -114,3 +114,70 @@ either, not even to format a finding; that contract belongs to
 [`content-build`](https://github.com/HeroicLands/content-build), but it is four
 fields joined by colons, and acquiring a build toolchain to write one line
 would be the wrong trade.
+
+## `actions/no-attribution`
+
+Fails a pull request whose title, body, or commit messages credit an AI
+assistant.
+
+```yaml
+# .github/workflows/no-attribution.yml
+name: No Attribution
+
+on:
+  pull_request:
+    types: [opened, edited, reopened, synchronize]
+
+permissions:
+  contents: read
+  pull-requests: read
+
+jobs:
+  no-attribution:
+    name: No AI attribution
+    runs-on: ubuntu-latest
+    steps:
+      # Needs no checkout: the subjects are the pull request, not the tree.
+      - uses: HeroicLands/.github/actions/no-attribution@main
+        with:
+          token: ${{ github.token }}
+```
+
+| Input | Default | |
+| --- | --- | --- |
+| `token` | — | needs `contents: read` and `pull-requests: read` |
+
+Two forms are refused: a `Co-Authored-By:` trailer naming an assistant, and a
+signature line saying the work was generated by one. Three subjects are read —
+the title and body, which a human edits, and the commit messages, which survive
+the merge. Every finding is reported in one run, so a two-line fix is not two
+round trips.
+
+**The pattern is anchored to the start of a line, and that is the whole trick.**
+Real attribution is a trailer or signature standing at column zero, so anchoring
+lets a pull request *describe* the rule — this section, for instance — without
+tripping it. The first version was unanchored and failed its own pull request.
+
+Findings are `address:line:column: severity: message`. The address is not a file
+path, because none of these subjects is a file; it is the subject's own address
+in the repository — `pull/123/body`, `commit/<sha>` — so a finding names
+something you can open or `git show`, and the column is the trailer's own rather
+than the start of the line.
+
+Nothing is edited. The check reports and fails, and the fix stays a human
+decision — which matters more here than elsewhere, since the thing being removed
+is a claim about who wrote the work.
+
+### Why this is an Action
+
+The same reason as the other two, arrived at from the opposite direction.
+Every repository already had this check, as a 60-line block of `bash` and `gh`
+embedded in a workflow — copied seven times, with the regex, the anchoring
+rationale and the failure message duplicated in each. Nothing about it varies by
+repository, so there was nothing for an input to capture; what there was, was
+seven chances for the pattern to drift and no way to fix it once.
+
+The repositories that also want the rule *before* a commit exists keep a local
+`.githooks/commit-msg` carrying the same pattern. That one cannot move here — a
+git hook runs on a developer's machine, from their checkout — so those two are
+the pair to keep in sync, and the only pair.
diff --git a/actions/no-attribution/action.yml b/actions/no-attribution/action.yml
new file mode 100644
index 0000000..862ada8
--- /dev/null
+++ b/actions/no-attribution/action.yml
@@ -0,0 +1,26 @@
+name: No AI attribution
+description: >-
+  Fail when a pull request's title, body, or any of its commit messages carries
+  AI/assistant attribution — a Co-Authored-By trailer naming an assistant, or a
+  "Generated with Claude Code"-style signature. Reports and fails; never edits.
+
+inputs:
+  token:
+    description: >-
+      A token with `contents: read` and `pull-requests: read`, used to read
+      the pull request's commit messages. The calling workflow's own
+      `github.token` is enough — pass it wrapped in expression braces there.
+      They are absent here on purpose: an expression anywhere in a manifest,
+      a description included, is evaluated when the manifest loads, and
+      `github` is not a context that an action manifest has. Writing one in
+      this sentence fails every workflow that calls the action.
+    required: true
+
+runs:
+  using: composite
+  steps:
+    - name: Scan the pull request for attribution
+      shell: bash
+      env:
+        GITHUB_TOKEN: ${{ inputs.token }}
+      run: node "$GITHUB_ACTION_PATH/no-attribution.mjs"
diff --git a/actions/no-attribution/no-attribution.mjs b/actions/no-attribution/no-attribution.mjs
new file mode 100644
index 0000000..ae3005a
--- /dev/null
+++ b/actions/no-attribution/no-attribution.mjs
@@ -0,0 +1,211 @@
+/*
+ * Copyright (c) 2024-2026 Tom Rodriguez ("Toasty") — <toasty@heroiclands.org>
+ *
+ * SPDX-License-Identifier: GPL-3.0-or-later
+ */
+
+/**
+ * **No AI/assistant attribution on a pull request.**
+ *
+ * A `Co-Authored-By:` trailer naming an assistant, or a signature line saying
+ * the work was generated by one, is noise in a project's history: it credits a
+ * tool rather than a person, it is appended by the tool itself rather than
+ * chosen by the author, and once merged it is permanent. Every HeroicLands
+ * repository wants it gone, so the check is the same one everywhere and the
+ * rule takes no configuration.
+ *
+ * **Three subjects, because attribution reaches a pull request three ways.**
+ * The title and body are the ones a human edits, and the ones an assistant
+ * most often writes wholesale; the commit messages are the ones that survive
+ * the merge. All three are read, and every finding is reported in one run
+ * rather than one per push — a check that stops at the first hit turns a
+ * two-line fix into two round trips.
+ *
+ * **The pattern is anchored to the start of a line, and that is the whole
+ * trick.** Real attribution is a trailer or a signature standing at column
+ * zero, so anchoring lets a pull request *describe* the rule — this paragraph,
+ * for instance — without tripping it. The first version of this check was
+ * unanchored and failed its own pull request. The leading run before the match
+ * is tolerated deliberately: whitespace before a trailer, and an emoji or
+ * bullet before a signature.
+ *
+ * Findings are `address:line:column: severity: message`. The address is not a
+ * file path, because none of these subjects is a file — it is the subject's own
+ * address in the repository (`pull/123/body`, `commit/<sha>`), so a finding
+ * names something you can open or `git show`. The line and column are real
+ * positions within that text, and the column is the trailer's own rather than
+ * the start of the line.
+ *
+ * Writes nothing, and never edits a message. It reports and fails, so a human
+ * decides how to fix it. The local `.githooks/commit-msg` hook some
+ * repositories carry enforces the same pattern before a commit is written;
+ * keep the two in sync.
+ *
+ * @module
+ */
+
+import { readFileSync } from "node:fs";
+
+const API = "https://api.github.com";
+const REPO = process.env.GITHUB_REPOSITORY;
+const TOKEN = process.env.GITHUB_TOKEN;
+const EVENT_PATH = process.env.GITHUB_EVENT_PATH;
+
+/**
+ * The forms attribution takes, each split so that group 1 is the run of
+ * characters *before* the offending token. Its length is the column, which is
+ * why the split exists — anchoring the pattern at the line start would
+ * otherwise make every finding report column 1 and send a reader to the
+ * indentation instead of the trailer.
+ */
+const ATTRIBUTION = [
+    {
+        kind: "Co-Authored-By trailer naming an assistant",
+        pattern: /^([ \t]*)(co-authored-by:.*(?:claude|anthropic))/i,
+    },
+    {
+        kind: "assistant signature",
+        pattern: /^([^A-Za-z]*)(generated with .*claude code)/i,
+    },
+];
+
+/** A finding, in the form an error matcher already reads. */
+function report({ file, line, column, severity = "error", message }) {
+    const at = [file, line, column].filter((part) => part != null).join(":");
+    console.error(`${at}: ${severity}: ${message}`);
+}
+
+/**
+ * Every attributing line in one subject.
+ *
+ * @param {string} address the subject's address, used as the finding's locator
+ * @param {string | null | undefined} text the subject itself; an empty body is
+ *   ordinary and yields nothing
+ * @returns {{address: string, line: number, column: number, kind: string, text: string}[]}
+ */
+function scan(address, text) {
+    const findings = [];
+    for (const [index, line] of (text ?? "").split(/\r?\n/).entries()) {
+        for (const { kind, pattern } of ATTRIBUTION) {
+            const match = line.match(pattern);
+            if (!match) continue;
+            findings.push({
+                address,
+                line: index + 1,
+                column: match[1].length + 1,
+                kind,
+                text: match[2].trim(),
+            });
+            break;
+        }
+    }
+    return findings;
+}
+
+/** The pull request this run is about, from the event that triggered it. */
+function pullRequest() {
+    if (!EVENT_PATH) {
+        report({
+            file: "GITHUB_EVENT_PATH",
+            message:
+                "no event payload in the environment. This action reads the " +
+                "pull request that triggered it, so it belongs on a " +
+                "`pull_request` workflow",
+        });
+        process.exit(1);
+    }
+    const event = JSON.parse(readFileSync(EVENT_PATH, "utf8"));
+    if (!event.pull_request) {
+        report({
+            file: process.env.GITHUB_EVENT_NAME ?? "event",
+            message:
+                "not a pull request event. Trigger this action on " +
+                "`pull_request`, whose payload carries the title, body and " +
+                "commits it examines",
+        });
+        process.exit(1);
+    }
+    return event.pull_request;
+}
+
+/**
+ * The pull request's commit messages, paginated.
+ *
+ * A failed read exits rather than reporting success on the subjects it did
+ * manage to read: the commit messages are the ones that survive the merge, so
+ * a run that skipped them has not checked the thing that matters most.
+ */
+async function commitMessages(number) {
+    const headers = {
+        Accept: "application/vnd.github+json",
+        "X-GitHub-Api-Version": "2022-11-28",
+        Authorization: `Bearer ${TOKEN}`,
+    };
+    const commits = [];
+    for (let page = 1; ; page++) {
+        const res = await fetch(
+            `${API}/repos/${REPO}/pulls/${number}/commits?per_page=100&page=${page}`,
+            { headers },
+        );
+        if (!res.ok) {
+            report({
+                file: `pull/${number}/commits`,
+                message:
+                    `could not be read: ${res.status} ${await res.text()}. ` +
+                    "The token needs `contents: read` and `pull-requests: read`",
+            });
+            process.exit(1);
+        }
+        const batch = await res.json();
+        for (const entry of batch) {
+            commits.push({ sha: entry.sha, message: entry.commit.message });
+        }
+        if (batch.length < 100) return commits;
+    }
+}
+
+if (!TOKEN) {
+    report({
+        file: "token",
+        message:
+            "no token supplied, so the commit messages cannot be read. Pass " +
+            "`token: ${{ github.token }}`",
+    });
+    process.exit(1);
+}
+
+const pr = pullRequest();
+const commits = await commitMessages(pr.number);
+
+const findings = [
+    ...scan(`pull/${pr.number}/title`, pr.title),
+    ...scan(`pull/${pr.number}/body`, pr.body),
+    ...commits.flatMap((commit) => scan(`commit/${commit.sha}`, commit.message)),
+];
+
+if (findings.length) {
+    console.error(
+        `\nno-attribution: ${findings.length} attributing line(s) on ` +
+            `pull request #${pr.number}:\n`,
+    );
+    for (const finding of findings) {
+        report({
+            file: finding.address,
+            line: finding.line,
+            column: finding.column,
+            message: `${finding.kind}: ${finding.text}`,
+        });
+    }
+    console.error(
+        "\nThis project does not credit an assistant in its history. Edit the " +
+            "pull request's\ntitle and body, amend or rebase any commit whose " +
+            "message carries the line, and push.\nNothing was changed for " +
+            "you — the fix is yours to make.\n",
+    );
+    process.exit(1);
+}
+
+console.log(
+    `no-attribution: pull request #${pr.number} is clean — title, body and ` +
+        `${commits.length} commit message(s) carry no attribution.`,
+);