diff --git a/src/app/api/insights/chat/route.ts b/src/app/api/insights/chat/route.ts
index 512af8e14..27b46004f 100644
--- a/src/app/api/insights/chat/route.ts
+++ b/src/app/api/insights/chat/route.ts
@@ -90,6 +90,7 @@ import {
   findUnverifiedCoachNumbers,
   stripUnverifiedNumbers,
 } from "@/lib/ai/coach/coach-prose-grounding";
+import { scrubUnknownLearnLinks } from "@/lib/ai/coach/learn-link-guard";
 import { parseSuggestReminder } from "@/lib/ai/coach/suggest-reminder";
 import { gateSuggestion } from "@/lib/ai/coach/suggest-gate";
 import {
@@ -768,6 +769,28 @@ Reply now as the assistant, in ${locale === "de" ? "German" : "English"}. Fetch
     }
   }
 
+  // v1.21.0 (NEW-C C-3) — Learn-link post-filter. The prompt instructs the
+  // model to only link a published `/learn/<slug>`, but that is guidance, not
+  // enforcement: a fabricated `/learn/<invented-slug>` would otherwise ship as
+  // a dead link. Scrub any reference whose slug is not in the catalog (a real
+  // one is kept verbatim). A blocked turn carries canned fallback prose with no
+  // links, so skip it.
+  if (!outbound.block && replyText.includes("/learn/")) {
+    const scrubbed = scrubUnknownLearnLinks(replyText);
+    if (scrubbed.dropped.length > 0) {
+      replyText = scrubbed.text;
+      annotate({
+        action: { name: "coach.learn.link_dropped" },
+        meta: {
+          dropped: scrubbed.dropped.length,
+          // Truncated slug tokens for ops triage — no user content.
+          slugs: scrubbed.dropped.slice(0, 6),
+          promptVersion: PROMPT_VERSION,
+        },
+      });
+    }
+  }
+
   let surfacedSuggestion: CoachSuggestion | null = null;
   if (!outbound.block && suggestParse.cadence) {
     const decision = await gateSuggestion({
diff --git a/src/lib/ai/coach/__tests__/learn-link-guard.test.ts b/src/lib/ai/coach/__tests__/learn-link-guard.test.ts
new file mode 100644
index 000000000..d47d5d901
--- /dev/null
+++ b/src/lib/ai/coach/__tests__/learn-link-guard.test.ts
@@ -0,0 +1,63 @@
+/**
+ * v1.21.0 (NEW-C C-3) — the Coach Learn-link post-filter. A published
+ * `/learn/<slug>` is kept verbatim; a fabricated one (a slug not in the
+ * catalog) is neutralised, so the catalog's "impossible by construction" claim
+ * becomes an enforced guarantee rather than a prompt promise.
+ */
+import { describe, expect, it } from "vitest";
+
+import { scrubUnknownLearnLinks } from "@/lib/ai/coach/learn-link-guard";
+import { LEARN_GUIDES } from "@/lib/ai/coach/learn-catalog";
+
+const REAL = LEARN_GUIDES.find((g) => g.slug === "resting-heart-rate")!;
+
+describe("scrubUnknownLearnLinks", () => {
+  it("returns the input unchanged when there is no /learn reference", () => {
+    const reply = "Your resting heart rate is settling nicely this week.";
+    const out = scrubUnknownLearnLinks(reply);
+    expect(out.text).toBe(reply);
+    expect(out.dropped).toEqual([]);
+  });
+
+  it("keeps a published /learn link verbatim (absolute URL)", () => {
+    const reply = `Worth a read — more on this: ${REAL.url}.`;
+    const out = scrubUnknownLearnLinks(reply);
+    expect(out.text).toContain(REAL.url);
+    expect(out.dropped).toEqual([]);
+  });
+
+  it("drops a fabricated /learn slug and reports it", () => {
+    const reply =
+      "That can drift with stress. More on this: https://healthlog.dev/learn/lower-your-cortisol-fast — give it a look.";
+    const out = scrubUnknownLearnLinks(reply);
+    expect(out.dropped).toContain("lower-your-cortisol-fast");
+    expect(out.text).not.toContain("lower-your-cortisol-fast");
+    expect(out.text).not.toContain("/learn/");
+    // The surrounding sentence survives.
+    expect(out.text).toContain("That can drift with stress");
+    expect(out.text).toContain("give it a look");
+  });
+
+  it("drops a fabricated relative /learn link", () => {
+    const reply = "See /learn/made-up-topic for the details.";
+    const out = scrubUnknownLearnLinks(reply);
+    expect(out.dropped).toContain("made-up-topic");
+    expect(out.text).not.toContain("/learn/");
+  });
+
+  it("keeps the real link and drops the fake one in the same reply", () => {
+    const reply = `Two reads: ${REAL.url} and https://healthlog.dev/learn/fictional-guide.`;
+    const out = scrubUnknownLearnLinks(reply);
+    expect(out.text).toContain(REAL.url);
+    expect(out.text).not.toContain("fictional-guide");
+    expect(out.dropped).toEqual(["fictional-guide"]);
+  });
+
+  it("is case-insensitive on the slug match", () => {
+    // A real slug in mixed case still resolves as published (kept).
+    const reply = "Read: https://healthlog.dev/learn/Resting-Heart-Rate.";
+    const out = scrubUnknownLearnLinks(reply);
+    expect(out.dropped).toEqual([]);
+    expect(out.text).toContain("Resting-Heart-Rate");
+  });
+});
diff --git a/src/lib/ai/coach/learn-link-guard.ts b/src/lib/ai/coach/learn-link-guard.ts
new file mode 100644
index 000000000..3c06b2749
--- /dev/null
+++ b/src/lib/ai/coach/learn-link-guard.ts
@@ -0,0 +1,100 @@
+/**
+ * v1.21.0 (NEW-C C-3) — Coach Learn-link post-filter.
+ *
+ * The Coach MAY point the user at a public `/learn/<slug>` guide, but only a
+ * slug that appears in `learn-catalog.ts`. The system prompt instructs this
+ * ("Only link a URL from this list — never invent a /learn URL"), but that is a
+ * prompt instruction, not enforcement: a model could still hallucinate
+ * `/learn/lower-your-cortisol` and ship a dead/wrong link.
+ *
+ * This deterministic post-filter runs on the assembled reply before
+ * persistence and streaming. It scans for every `/learn/<slug>` reference
+ * (relative or the absolute `healthlog.dev/learn/<slug>` form) and NEUTRALISES
+ * any whose slug is not a published catalog entry — stripping the fabricated
+ * URL out of the prose while leaving the surrounding sentence intact. A real
+ * slug is kept verbatim. This converts the catalog's "impossible by
+ * construction" claim from a prompt promise into an enforced guarantee.
+ *
+ * Cheap + read-only: a single regex pass over the reply, the slug set sourced
+ * from the read-only `LEARN_GUIDES` catalog. No allocation on the common path
+ * (a reply with no `/learn` mention returns the input unchanged).
+ */
+import { LEARN_GUIDES } from "./learn-catalog";
+
+/** The set of published `/learn` slugs, lower-cased for case-insensitive match. */
+const KNOWN_SLUGS: ReadonlySet<string> = new Set(
+  LEARN_GUIDES.map((g) => g.slug.toLowerCase()),
+);
+
+/**
+ * Match a `/learn/<slug>` reference, with or without the host. Captures the
+ * slug so the caller can validate it. The slug is the URL-path segment after
+ * `/learn/` — letters, digits, and hyphens (the catalog's slug grammar). A
+ * trailing slash or punctuation ends the slug.
+ *
+ * Examples matched:
+ *   https://healthlog.dev/learn/resting-heart-rate
+ *   healthlog.dev/learn/made-up-slug
+ *   /learn/another-fabrication
+ */
+const LEARN_LINK_RE =
+  /(?:https?:\/\/)?(?:[a-z0-9.-]*healthlog\.dev)?\/learn\/([a-z0-9-]+)\/?/gi;
+
+export interface LearnLinkScrubResult {
+  /** The reply with every unknown `/learn/<slug>` reference removed. */
+  text: string;
+  /** The unknown slugs that were neutralised (lower-cased, deduped). */
+  dropped: string[];
+}
+
+/**
+ * Strip every `/learn/<slug>` reference whose slug is not in the published
+ * catalog. A known slug is preserved verbatim; an unknown one (a fabrication)
+ * is removed along with a common lead-in fragment ("more on this: ", "see "),
+ * so the prose does not read as a broken link. Returns the input unchanged when
+ * the reply contains no `/learn` reference.
+ */
+export function scrubUnknownLearnLinks(reply: string): LearnLinkScrubResult {
+  if (!reply || !reply.includes("/learn/")) {
+    return { text: reply, dropped: [] };
+  }
+
+  const dropped = new Set<string>();
+  // First pass: collect the unknown slugs (so the result is reportable) and
+  // build the cleaned string in one replace.
+  const text = reply.replace(LEARN_LINK_RE, (match, slug: string) => {
+    if (KNOWN_SLUGS.has(slug.toLowerCase())) {
+      // Published guide — keep the link exactly as written.
+      return match;
+    }
+    dropped.add(slug.toLowerCase());
+    // Drop the fabricated URL. Leave the surrounding prose; a dangling
+    // "more on this:" lead-in is cleaned by the post-pass below.
+    return "";
+  });
+
+  if (dropped.size === 0) {
+    // Every /learn link was valid — return the original (the replace above is a
+    // no-op clone, but be explicit about the contract).
+    return { text: reply, dropped: [] };
+  }
+
+  // Tidy the prose around a removed link: collapse a now-empty "more on this:"
+  // / "see:" lead-in and any double spaces / stranded punctuation the removal
+  // left behind. Conservative — only the well-known lead-in phrases the prompt
+  // example seeds, plus whitespace normalisation.
+  const tidied = text
+    // Drop a now-empty "more on this:" / "read more:" / "see:" lead-in (EN+DE).
+    .replace(
+      /\b(?:more on this|read more|see|learn more|mehr dazu|mehr hier)\s*:?\s*(?=[\s).,;!?]|$)/gi,
+      "",
+    )
+    // Collapse a doubled space + tidy a space-before-punctuation the gap left.
+    .replace(/[ \t]{2,}/g, " ")
+    .replace(/\s+([).,;!?])/g, "$1")
+    // Drop an empty parenthesis pair left by a parenthesised link.
+    .replace(/\(\s*\)/g, "")
+    .trim();
+
+  return { text: tidied, dropped: [...dropped] };
+}