diff --git a/src/lib/ai/__tests__/server-provider-health.test.ts b/src/lib/ai/__tests__/server-provider-health.test.ts
new file mode 100644
index 000000000..7e6c58909
--- /dev/null
+++ b/src/lib/ai/__tests__/server-provider-health.test.ts
@@ -0,0 +1,137 @@
+/**
+ * v1.38.19 (wave D) — the instance-wide health of the operator's shared
+ * provider, as a tri-state a FRESH account is allowed to read.
+ *
+ * The ledger itself is per-user and fails OPEN (an optimisation, never a
+ * gate). This projection is the opposite on both counts: it folds rows
+ * across every account into one instance-level verdict, and it fails
+ * CLOSED — anything but a demonstrated recent success is not an offer.
+ *
+ * What it must never leak: how many accounts are tracked, when any of them
+ * last succeeded, or which account a row belongs to. The return type is the
+ * whole contract.
+ */
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+vi.mock("@/lib/db", () => ({
+  prisma: {
+    providerHealth: { findMany: vi.fn() },
+  },
+}));
+
+vi.mock("@/lib/logging/context", () => ({
+  annotate: vi.fn(),
+}));
+
+import { prisma } from "@/lib/db";
+import { annotate } from "@/lib/logging/context";
+import {
+  SERVER_PROVIDER_FRESH_WINDOW_MS,
+  readServerProviderHealth,
+} from "../server-provider-health";
+
+const NOW = new Date("2026-09-11T12:00:00.000Z");
+
+function ago(ms: number): Date {
+  return new Date(NOW.getTime() - ms);
+}
+
+function ahead(ms: number): Date {
+  return new Date(NOW.getTime() + ms);
+}
+
+function rows(
+  ...list: Array<{
+    lastResult: string;
+    lastOkAt?: Date | null;
+    nextRetryAt?: Date | null;
+  }>
+) {
+  vi.mocked(prisma.providerHealth.findMany).mockResolvedValue(
+    list.map((row) => ({
+      lastResult: row.lastResult,
+      lastOkAt: row.lastOkAt ?? null,
+      nextRetryAt: row.nextRetryAt ?? null,
+    })) as never,
+  );
+}
+
+beforeEach(() => {
+  vi.mocked(prisma.providerHealth.findMany).mockReset();
+  vi.mocked(annotate).mockReset();
+});
+
+describe("readServerProviderHealth", () => {
+  it("is healthy when some account saw the shared provider work inside the window", async () => {
+    rows({ lastResult: "ok", lastOkAt: ago(60 * 60 * 1000) });
+    await expect(readServerProviderHealth(NOW)).resolves.toBe("healthy");
+  });
+
+  it("is unhealthy when the last success is older than the window", async () => {
+    rows({
+      lastResult: "ok",
+      lastOkAt: ago(SERVER_PROVIDER_FRESH_WINDOW_MS + 60_000),
+    });
+    await expect(readServerProviderHealth(NOW)).resolves.toBe("unhealthy");
+  });
+
+  it("is unhealthy when rows exist but none ever succeeded", async () => {
+    rows({ lastResult: "hard_failed", lastOkAt: null });
+    await expect(readServerProviderHealth(NOW)).resolves.toBe("unhealthy");
+  });
+
+  it("lets a 401 in cooldown beat a fresh ok — a dead key means the same for everyone", async () => {
+    rows(
+      { lastResult: "ok", lastOkAt: ago(60 * 1000) },
+      {
+        lastResult: "auth_failed",
+        lastOkAt: ago(3 * 60 * 60 * 1000),
+        nextRetryAt: ahead(60 * 60 * 1000),
+      },
+    );
+    await expect(readServerProviderHealth(NOW)).resolves.toBe("unhealthy");
+  });
+
+  it("ignores an auth failure whose cooldown has run out", async () => {
+    rows(
+      { lastResult: "ok", lastOkAt: ago(60 * 1000) },
+      {
+        lastResult: "auth_failed",
+        lastOkAt: null,
+        nextRetryAt: ago(60 * 1000),
+      },
+    );
+    await expect(readServerProviderHealth(NOW)).resolves.toBe("healthy");
+  });
+
+  it("is unknown when the instance has no rows for either shared provider", async () => {
+    rows();
+    await expect(readServerProviderHealth(NOW)).resolves.toBe("unknown");
+  });
+
+  it("is unknown when the read throws, and says so in the event", async () => {
+    vi.mocked(prisma.providerHealth.findMany).mockRejectedValue(
+      new Error("connection terminated"),
+    );
+    await expect(readServerProviderHealth(NOW)).resolves.toBe("unknown");
+    expect(vi.mocked(annotate)).toHaveBeenCalledWith(
+      expect.objectContaining({
+        action: { name: "ai.server_health.read_failed" },
+      }),
+    );
+  });
+
+  it("reads only the operator-funded provider tags, and no per-account column", async () => {
+    rows({ lastResult: "ok", lastOkAt: ago(1000) });
+    await readServerProviderHealth(NOW);
+    const args = vi.mocked(prisma.providerHealth.findMany).mock.calls[0]![0]!;
+    expect(args.where).toEqual({
+      providerType: { in: ["admin-openai", "admin-codex"] },
+    });
+    expect(Object.keys(args.select ?? {}).sort()).toEqual([
+      "lastOkAt",
+      "lastResult",
+      "nextRetryAt",
+    ]);
+  });
+});
diff --git a/src/lib/ai/server-provider-health.ts b/src/lib/ai/server-provider-health.ts
new file mode 100644
index 000000000..aea613a71
--- /dev/null
+++ b/src/lib/ai/server-provider-health.ts
@@ -0,0 +1,112 @@
+/**
+ * v1.38.19 (wave D) — "does the shared provider of THIS instance work?",
+ * answered for a user who has never made a single AI call.
+ *
+ * The question comes from the end of the setup flow. The done screen used
+ * to claim "insights work for you right away" on the strength of
+ * `managedBy === "server"`, which is a PRESENCE read: it says the operator
+ * configured a key, never that the key works. On 2026-09-11 the operator's
+ * own instance answered HTTP 500 from its OAuth proxy while the screen kept
+ * making that promise. This module is the evidence the promise needs.
+ *
+ * Why it cannot come from the user's own ledger: `provider_health` is unique
+ * on `(user_id, provider_type)`, and an account walking the wizard has NO
+ * row — for any provider. Its own ledger can only answer "unknown", forever.
+ * So the projection folds rows ACROSS accounts. That is also the reason for
+ * the shape of the return value.
+ *
+ * What leaves this module is a tri-state and nothing else: no counts, no
+ * timestamps, no user ids, not even "how many accounts are tracked". "The
+ * provider of this instance last worked" is a statement about the instance;
+ * "seven accounts used it at 06:42Z" is a statement about other people. The
+ * admin fold (`GET /api/admin/provider-health`) keeps the second kind and
+ * stays behind `requireAdmin()`; this one is readable by any signed-in user
+ * precisely because it carries none of it.
+ *
+ * Fail CLOSED, deliberately against the grain of the ledger it reads. The
+ * ledger fails open (`provider-health-ledger.ts`) because there it protects
+ * a generation already under way — contributing nothing is the safe outcome.
+ * Here the output is a promise made to a person, so a read error, a timeout
+ * or an empty table is `unknown`, and `unknown` buys nothing: the caller
+ * treats it exactly like `unhealthy`.
+ *
+ * No probe. `POST /api/ai/test` exists and would be the direct answer, but it
+ * is a real completion billed to whoever the chain resolves to — usually the
+ * operator. A probe here would charge the operator for every registration,
+ * and a sign-up bot would scale that bill. The design spec's rule holds:
+ * nothing in the flow talks to a provider.
+ */
+import { prisma } from "@/lib/db";
+import { annotate } from "@/lib/logging/context";
+
+/**
+ * The whole contract. Three words, no payload.
+ *
+ *   healthy   — the instance's shared provider demonstrably served somebody
+ *               inside the freshness window, and no credential is benched.
+ *   unhealthy — it has been tried and it is not working.
+ *   unknown   — nobody has tried it yet on this instance, or the read
+ *               failed. Never an offer; see the module note on failing closed.
+ */
+export type ServerProviderHealth = "healthy" | "unhealthy" | "unknown";
+
+/**
+ * How recent a success has to be to still count.
+ *
+ * The ledger's own cooldowns bracket the choice: 6 h for an auth failure,
+ * 5 min for a hard one. A window under those would read `unhealthy` on a
+ * quiet single-person instance simply because nothing ran that hour; a much
+ * larger one would paint a key that died on Monday as green on Thursday.
+ * A day is long enough for an instance with one briefing per morning to hold
+ * the threshold, short enough that a dead key goes grey within a day.
+ */
+export const SERVER_PROVIDER_FRESH_WINDOW_MS = 24 * 60 * 60 * 1000;
+
+/**
+ * The provider tags the operator pays for: the instance-wide OpenAI key and
+ * the shared central Codex account. These are exactly the two tags
+ * `consent-guard.ts` gates on, and that is not a coincidence — the offer
+ * this feeds is the consent for these two and for nothing else.
+ */
+const SHARED_PROVIDER_TYPES = ["admin-openai", "admin-codex"] as const;
+
+export async function readServerProviderHealth(
+  now: Date = new Date(),
+): Promise<ServerProviderHealth> {
+  let rows: Array<{
+    lastResult: string;
+    lastOkAt: Date | null;
+    nextRetryAt: Date | null;
+  }>;
+  try {
+    rows = await prisma.providerHealth.findMany({
+      where: { providerType: { in: [...SHARED_PROVIDER_TYPES] } },
+      // Three columns, none of which identifies an account. The absence of
+      // `userId` here is load-bearing, not tidiness.
+      select: { lastResult: true, lastOkAt: true, nextRetryAt: true },
+    });
+  } catch {
+    annotate({ action: { name: "ai.server_health.read_failed" } });
+    return "unknown";
+  }
+
+  if (rows.length === 0) return "unknown";
+
+  // A benched credential outranks any success. A 401/403 is not one
+  // account's bad luck — the operator's key is one object, and while it is
+  // in its cooldown it is dead for everybody, including whoever is reading
+  // this screen right now.
+  const credentialBenched = rows.some(
+    (row) =>
+      row.lastResult === "auth_failed" &&
+      row.nextRetryAt !== null &&
+      row.nextRetryAt.getTime() > now.getTime(),
+  );
+  if (credentialBenched) return "unhealthy";
+
+  const floor = now.getTime() - SERVER_PROVIDER_FRESH_WINDOW_MS;
+  const workedRecently = rows.some(
+    (row) => row.lastOkAt !== null && row.lastOkAt.getTime() >= floor,
+  );
+  return workedRecently ? "healthy" : "unhealthy";
+}