diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 3f89edb7..a58b9e74 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -37,6 +37,9 @@ jobs:
       - name: Typecheck
         run: npx tsc --noEmit
 
+      - name: Check content invariants
+        run: npm run check:content
+
       - name: Build static export
         run: npm run build
 
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 7c3abf03..810f07ba 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -159,6 +159,18 @@ The repository's lint baseline is clean: `npm run lint` prints nothing and exits
 0. Pull-request CI runs it, so a new warning or error will fail the build. Fix
 the finding rather than suppressing it.
 
+When you change people, cohorts, or counts, also run:
+
+```powershell
+npm run check:content
+```
+
+It asserts what a build cannot: that the cohort head-counts in
+`team.researchersCount`, the `cohorts` cards, and `README.md` all match the
+number of researchers actually on the roster, and that no two people share a
+`/team/[slug]`. Those numbers are maintained by hand in separate places, and
+have twice been published wrong.
+
 Always run the production build:
 
 ```powershell
diff --git a/UPDATES-NEEDED.md b/UPDATES-NEEDED.md
index ea745528..e231597a 100644
--- a/UPDATES-NEEDED.md
+++ b/UPDATES-NEEDED.md
@@ -257,18 +257,32 @@ changed.
 
 ### UPD-008 - Add focused smoke checks
 
-**Priority:** Medium
+**Priority:** Medium — **started: content invariants added.**
 
 There is no automated test script. The production build succeeds, but a build
 alone does not prove that key routes, base-path assets, publication anchors, or
 generated slugs are correct.
 
-**Proposed update:** Start with a small check over critical exported files and
-known content invariants. Add browser automation only when it protects a
-specific high-value interaction and can remain reliable.
+`npm run check:content` now runs in CI and asserts the invariants a build
+cannot: every cohort card's head-count matches the number of researchers
+actually on the roster for that term, `team.researchersCount` and the README's
+prose count agree with the current cohort, and no two people share a
+`/team/[slug]`.
+
+This targets a failure this repository has published twice. The Fall 2025 count
+read seven when the roster should have held eight, because Pegi Bracaj had
+never been added. The README read "eight applied researchers" against a Summer
+2026 roster of seven. Both are valid strings, so both survived every build and
+lint. The check reads the real `site.ts` exports rather than parsing text, and
+was verified against four seeded faults: a wrong README word, a stale cohort
+card, a researcher moved between terms, and two people sharing a slug.
+
+**Still open:** route reachability, base-path asset references, and publication
+anchors are not covered.
+
+**Acceptance:** partially met — invalid generated content mappings and cohort
+counts are caught before merge. Not met for routes and assets.
 
-**Acceptance:** CI detects at least a missing critical route, broken local
-asset reference, or invalid generated content mapping before merge.
 
 ### UPD-009 - Separate website issues from wider CoLab operations
 
diff --git a/docs/CONTENT-GUIDE.md b/docs/CONTENT-GUIDE.md
index 2807631b..aa8f8a9a 100644
--- a/docs/CONTENT-GUIDE.md
+++ b/docs/CONTENT-GUIDE.md
@@ -86,7 +86,9 @@ Each person is a `TeamMember` object. Fields:
   image.
 - Changing the cohort size means updating the count in **three** places:
   `team.researchersCount`, the matching `cohorts` entry's `items`, and
-  [`README.md`](../README.md).
+  [`README.md`](../README.md). Run `npm run check:content` to confirm all three
+  agree with the roster — CI runs it, and it will fail the pull request if they
+  do not.
 - Never publish a name, role, affiliation, or biography you have not verified
   against an authoritative source.
 
diff --git a/package.json b/package.json
index 9bc48040..b87d4f37 100644
--- a/package.json
+++ b/package.json
@@ -10,6 +10,7 @@
     "build": "next build",
     "start": "next start",
     "lint": "eslint",
+    "check:content": "node --disable-warning=MODULE_TYPELESS_PACKAGE_JSON scripts/check-content-invariants.mjs",
     "sync:static": "bash scripts/sync-static-site.sh",
     "sync:newsletter": "node scripts/sync-newsletter.mjs",
     "export:haste-paper": "node scripts/export-haste-paper.mjs",
diff --git a/scripts/check-content-invariants.mjs b/scripts/check-content-invariants.mjs
new file mode 100644
index 00000000..47357a6f
--- /dev/null
+++ b/scripts/check-content-invariants.mjs
@@ -0,0 +1,133 @@
+// Content invariants that a successful build does not prove.
+//
+// The cohort head-counts are written by hand in three places — the roster
+// itself, `team.researchersCount`, the `cohorts` cards, and the README — and
+// nothing has ever forced them to agree. Both times this repository has
+// published a wrong cohort size, the cause was the same: a researcher was
+// added or missed in one place and the other numbers were never touched. A
+// build cannot catch that, because every one of those values is a valid
+// string.
+//
+// Run with `npm run check:content`.
+
+import fs from "node:fs";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+
+const root = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
+const { team, cohorts, cohortTerms } = await import(
+  new URL("../src/content/site.ts", import.meta.url).href
+);
+
+const failures = [];
+const fail = (msg) => failures.push(msg);
+
+const currentTerm = cohortTerms[cohortTerms.length - 1];
+
+/** Researchers actually on the roster for a term. */
+function rosterCount(term) {
+  return team.researchers.filter((r) => (r.term ?? currentTerm) === term).length;
+}
+
+/** "7 researchers" -> 7 */
+function statedCount(text) {
+  const m = /^(\d+)\s+researchers?$/.exec(text.trim());
+  return m ? Number(m[1]) : null;
+}
+
+// 1. Every cohort card's head-count matches the roster for that term.
+for (const cohort of cohorts) {
+  const counts = cohort.items
+    .filter((item) => typeof item === "string")
+    .map(statedCount)
+    .filter((n) => n !== null);
+
+  if (counts.length === 0) {
+    fail(`cohorts["${cohort.term}"] has no "N researchers" item to check.`);
+    continue;
+  }
+  for (const stated of counts) {
+    const actual = rosterCount(cohort.term);
+    if (stated !== actual) {
+      fail(
+        `cohorts["${cohort.term}"] says ${stated} researchers, but the roster ` +
+          `holds ${actual}. Either a researcher is missing from team.researchers ` +
+          `or the count is stale.`
+      );
+    }
+  }
+}
+
+// 2. team.researchersCount matches the current cohort.
+{
+  const stated = statedCount(team.researchersCount);
+  const actual = rosterCount(currentTerm);
+  if (stated === null) {
+    fail(`team.researchersCount ("${team.researchersCount}") is not "N researchers".`);
+  } else if (stated !== actual) {
+    fail(
+      `team.researchersCount says ${stated}, but the ${currentTerm} roster holds ${actual}.`
+    );
+  }
+}
+
+// 3. The README's prose count matches the current cohort. This is the exact
+//    claim that was wrong: it read "eight applied researchers" against a roster
+//    of seven.
+{
+  const WORDS = [
+    "zero", "one", "two", "three", "four", "five", "six", "seven", "eight",
+    "nine", "ten", "eleven", "twelve",
+  ];
+  const readme = fs.readFileSync(path.join(root, "README.md"), "utf8");
+  const m = /(\w+)\s+applied researchers/i.exec(readme);
+  if (!m) {
+    fail("README.md no longer states a '<number> applied researchers' count.");
+  } else {
+    const actual = rosterCount(currentTerm);
+    const stated = WORDS.indexOf(m[1].toLowerCase());
+    if (stated !== actual) {
+      fail(
+        `README.md says "${m[1]} applied researchers", but the ${currentTerm} ` +
+          `roster holds ${actual} (${WORDS[actual] ?? actual}).`
+      );
+    }
+  }
+}
+
+// 4. A slug must identify one person. The same person may legitimately appear
+//    in more than one collection (a fellow who is also a researcher), but two
+//    different people sharing a slug would collide on /team/[slug].
+{
+  const bySlug = new Map();
+  const everyone = [
+    team.founder,
+    ...team.advisors,
+    ...team.residentFellows,
+    ...team.researchers,
+    ...team.collaborators,
+  ];
+  for (const person of everyone) {
+    if (!person?.slug) continue;
+    const seen = bySlug.get(person.slug);
+    if (seen && seen !== person.name) {
+      fail(
+        `slug "${person.slug}" is used by two different people: ` +
+          `"${seen}" and "${person.name}".`
+      );
+    }
+    bySlug.set(person.slug, person.name);
+  }
+}
+
+if (failures.length > 0) {
+  console.error("Content invariants failed:\n");
+  for (const f of failures) console.error("  - " + f);
+  console.error("");
+  process.exit(1);
+}
+
+console.log(
+  `Content invariants OK (${cohorts.length} cohorts, ` +
+    `${team.researchers.length} researcher records).`
+);