diff --git a/packages/loopover-engine/src/index.ts b/packages/loopover-engine/src/index.ts
index 3d31fb9c87..68f89951dd 100644
--- a/packages/loopover-engine/src/index.ts
+++ b/packages/loopover-engine/src/index.ts
@@ -643,6 +643,16 @@ export {
   type LoopHealthTier,
   type LoopRunOutcome,
 } from "./loop-escalation.js";
+// The internal ops fleet view (#4808) over the escalation vocabulary re-exported above: it calls the same
+// evaluateEscalation per loop rather than restating what "needs a human" means.
+export {
+  buildActiveLoopFleetSummary,
+  LOOP_HEALTH_TIERS,
+  LOOP_RUN_STATUSES,
+  type ActiveLoopFacts,
+  type ActiveLoopFleetSummary,
+  type FleetLoopRow,
+} from "./loop-fleet-summary.js";
 export {
   buildMetadataRankInput,
   computeMetadataDupRisk,
diff --git a/packages/loopover-engine/src/loop-fleet-summary.ts b/packages/loopover-engine/src/loop-fleet-summary.ts
new file mode 100644
index 0000000000..90fccbacb7
--- /dev/null
+++ b/packages/loopover-engine/src/loop-fleet-summary.ts
@@ -0,0 +1,109 @@
+// Active rented-loop fleet summary (pure) — #4808, part of the Rent-a-Loop path #4778.
+//
+// Deterministic and side-effect-free: given every rented loop the internal ops team currently knows about, it
+// produces the at-a-glance view an operator needs — how many loops are live, how they break down by run status
+// and health tier, and which ones are misbehaving badly enough to need a human right now. That is exactly
+// #4808's acceptance criterion ("an internal operator can see every currently active rented loop and its status
+// at a glance") as a decision core: what a dashboard panel renders and what an alert rule fires on, computed
+// once, the same way, for both.
+//
+// It reuses loop-escalation.ts's (#4806) already-merged vocabulary rather than restating it — the same
+// LoopRunOutcome/LoopHealthTier a loop is already described by, and evaluateEscalation itself to decide whether
+// a given loop needs attention. So the fleet view can never disagree with the per-loop escalation path about
+// what "needs a human" means: there is one rule, called once per loop, not a second copy that drifts.
+//
+// It summarizes only: no dashboard, no alert delivery, no IO, no clock read. Wiring panels/alert rules into the
+// self-host observability stack is the separate integration this issue is blocked on (#4793) — this core has no
+// opinion about Grafana or Alertmanager, so it stays correct whatever renders it.
+
+import { evaluateEscalation, type EscalationDecision, type LoopEscalationInput, type LoopHealthTier, type LoopRunOutcome } from "./loop-escalation.js";
+
+/** Every run status a loop can report, in the order an operator reads them (live work first). */
+export const LOOP_RUN_STATUSES: readonly LoopRunOutcome[] = ["running", "converged", "abandoned", "error"];
+/** Health tiers, worst-first — the order the summary surfaces them in. */
+export const LOOP_HEALTH_TIERS: readonly LoopHealthTier[] = ["critical", "degraded", "healthy"];
+
+/** One rented loop as ops currently knows it. Mirrors LoopEscalationInput's signals so the same rule applies. */
+export type ActiveLoopFacts = LoopEscalationInput & {
+  loopId: string;
+  tenantId: string;
+};
+
+/** One row of the operator's view: the loop, plus the escalation decision computed for it. */
+export type FleetLoopRow = {
+  loopId: string;
+  tenantId: string;
+  runStatus: LoopRunOutcome;
+  /** Absent when nothing has computed a health tier for this loop yet — reported as "unknown", never guessed. */
+  healthStatus: LoopHealthTier | "unknown";
+  needsAttention: boolean;
+  escalation: EscalationDecision;
+};
+
+export type ActiveLoopFleetSummary = {
+  /** Loops still running — the "currently active" count an operator reads first. */
+  activeCount: number;
+  /** Every loop handed in, however it ended. */
+  totalCount: number;
+  /** Count per run status. Every status is always present (0 when none), so a panel never renders a hole. */
+  byStatus: Record<LoopRunOutcome, number>;
+  /** Count per health tier, plus `unknown` for loops with no tier computed yet. Always fully populated. */
+  byHealth: Record<LoopHealthTier | "unknown", number>;
+  /** Loops needing a human, worst-first — what an alert rule fires on. A subset of `loops`, never a copy that drifts. */
+  needingAttention: FleetLoopRow[];
+  /** Every loop, in a stable order. */
+  loops: FleetLoopRow[];
+};
+
+/** Highest severity first; ties broken by loopId so the view is stable across renders, never reshuffled. */
+const SEVERITY_RANK: Record<EscalationDecision["severity"], number> = { high: 0, medium: 1, low: 2, none: 3 };
+
+function compareRows(a: FleetLoopRow, b: FleetLoopRow): number {
+  const bySeverity = SEVERITY_RANK[a.escalation.severity] - SEVERITY_RANK[b.escalation.severity];
+  return bySeverity !== 0 ? bySeverity : a.loopId.localeCompare(b.loopId);
+}
+
+function toRow(loop: ActiveLoopFacts): FleetLoopRow {
+  const escalation = evaluateEscalation(loop);
+  return {
+    loopId: loop.loopId,
+    tenantId: loop.tenantId,
+    runStatus: loop.runStatus,
+    healthStatus: loop.healthStatus ?? "unknown",
+    needsAttention: escalation.shouldEscalate,
+    escalation,
+  };
+}
+
+/**
+ * Summarize the rented-loop fleet for the internal ops view (#4808). Pure: reads only the loops it is handed
+ * and returns a summary without mutating, fetching, or notifying anything.
+ *
+ * `needsAttention` is not a second opinion — each row's flag IS evaluateEscalation's own `shouldEscalate` for
+ * that loop, so the fleet view and the per-loop escalation path (#4806) can never disagree about what needs a
+ * human. `needingAttention` is those rows, worst-severity first, with ties broken by `loopId` so an operator
+ * watching the panel sees a stable order rather than rows reshuffling between renders.
+ *
+ * Both breakdowns are always fully populated (0 for an absent status/tier) so a panel binds to a fixed set of
+ * keys and never renders a hole. A loop with no health tier computed yet counts as `unknown` rather than being
+ * assumed healthy — an operator must be able to tell "nothing is wrong" from "nothing has checked yet".
+ */
+export function buildActiveLoopFleetSummary(loops: readonly ActiveLoopFacts[]): ActiveLoopFleetSummary {
+  const byStatus = Object.fromEntries(LOOP_RUN_STATUSES.map((s) => [s, 0])) as Record<LoopRunOutcome, number>;
+  const byHealth = Object.fromEntries([...LOOP_HEALTH_TIERS, "unknown"].map((h) => [h, 0])) as Record<LoopHealthTier | "unknown", number>;
+
+  const rows = loops.map(toRow);
+  for (const row of rows) {
+    byStatus[row.runStatus] += 1;
+    byHealth[row.healthStatus] += 1;
+  }
+
+  return {
+    activeCount: byStatus.running,
+    totalCount: rows.length,
+    byStatus,
+    byHealth,
+    needingAttention: rows.filter((row) => row.needsAttention).sort(compareRows),
+    loops: [...rows].sort(compareRows),
+  };
+}
diff --git a/test/unit/loop-fleet-summary.test.ts b/test/unit/loop-fleet-summary.test.ts
new file mode 100644
index 0000000000..bff7ca26c9
--- /dev/null
+++ b/test/unit/loop-fleet-summary.test.ts
@@ -0,0 +1,125 @@
+import { describe, expect, it } from "vitest";
+
+import { evaluateEscalation } from "../../packages/loopover-engine/src/loop-escalation";
+import {
+  buildActiveLoopFleetSummary,
+  LOOP_HEALTH_TIERS,
+  LOOP_RUN_STATUSES,
+  type ActiveLoopFacts,
+} from "../../packages/loopover-engine/src/loop-fleet-summary";
+
+const loop = (over: Partial<ActiveLoopFacts> = {}): ActiveLoopFacts => ({
+  loopId: "loop-1",
+  tenantId: "acme",
+  runStatus: "running",
+  healthStatus: "healthy",
+  ...over,
+});
+
+describe("buildActiveLoopFleetSummary (#4808)", () => {
+  // The acceptance criterion: an operator sees every currently active rented loop and its status at a glance.
+  it("counts the live fleet and breaks it down by run status and health", () => {
+    const summary = buildActiveLoopFleetSummary([
+      loop({ loopId: "a" }),
+      loop({ loopId: "b" }),
+      loop({ loopId: "c", runStatus: "converged" }), // finished cleanly, so still healthy
+      loop({ loopId: "d", runStatus: "error", healthStatus: "critical" }),
+      loop({ loopId: "e", runStatus: "abandoned", healthStatus: "degraded" }),
+    ]);
+
+    expect(summary.activeCount).toBe(2); // only the running ones are "currently active"
+    expect(summary.totalCount).toBe(5);
+    expect(summary.byStatus).toEqual({ running: 2, converged: 1, abandoned: 1, error: 1 });
+    // Health is independent of run status: the converged loop is healthy too, so healthy counts a+b+c.
+    expect(summary.byHealth).toEqual({ critical: 1, degraded: 1, healthy: 3, unknown: 0 });
+  });
+
+  it("an empty fleet totals to zeros, never undefined", () => {
+    const summary = buildActiveLoopFleetSummary([]);
+    expect(summary.activeCount).toBe(0);
+    expect(summary.totalCount).toBe(0);
+    expect(summary.needingAttention).toEqual([]);
+    expect(summary.loops).toEqual([]);
+  });
+
+  it("INVARIANT: every status and tier key is always present, so a panel never renders a hole", () => {
+    const summary = buildActiveLoopFleetSummary([]);
+    expect(Object.keys(summary.byStatus).sort()).toEqual([...LOOP_RUN_STATUSES].sort());
+    expect(Object.keys(summary.byHealth).sort()).toEqual([...LOOP_HEALTH_TIERS, "unknown"].sort());
+    expect(Object.values(summary.byStatus).every((n) => n === 0)).toBe(true);
+    expect(Object.values(summary.byHealth).every((n) => n === 0)).toBe(true);
+  });
+
+  it("a loop with no computed health tier reads as unknown, never assumed healthy", () => {
+    const summary = buildActiveLoopFleetSummary([loop({ healthStatus: undefined })]);
+    expect(summary.loops[0]!.healthStatus).toBe("unknown");
+    expect(summary.byHealth.unknown).toBe(1);
+    expect(summary.byHealth.healthy).toBe(0);
+  });
+
+  describe("needsAttention", () => {
+    it("surfaces only the loops that need a human, worst-severity first", () => {
+      const summary = buildActiveLoopFleetSummary([
+        loop({ loopId: "healthy-one" }),
+        loop({ loopId: "degraded-one", healthStatus: "degraded" }), // low → notify
+        loop({ loopId: "killed-one", killRequested: true }), // high → stop
+        loop({ loopId: "abandoned-one", runStatus: "abandoned" }), // medium → human_review
+      ]);
+
+      expect(summary.needingAttention.map((r) => r.loopId)).toEqual(["killed-one", "abandoned-one", "degraded-one"]);
+      expect(summary.needingAttention.map((r) => r.escalation.severity)).toEqual(["high", "medium", "low"]);
+      expect(summary.needingAttention.every((r) => r.needsAttention)).toBe(true);
+    });
+
+    it("a fully healthy fleet needs no attention", () => {
+      const summary = buildActiveLoopFleetSummary([loop({ loopId: "a" }), loop({ loopId: "b" })]);
+      expect(summary.needingAttention).toEqual([]);
+      expect(summary.loops.every((r) => r.needsAttention === false)).toBe(true);
+      expect(summary.loops.every((r) => r.escalation.action === "none")).toBe(true);
+    });
+
+    // The point of reusing evaluateEscalation instead of re-deciding: the fleet view and the per-loop
+    // escalation path (#4806) cannot drift apart about what "needs a human" means.
+    it("INVARIANT: each row's decision IS evaluateEscalation's own, not a second opinion", () => {
+      const facts = loop({ loopId: "x", runStatus: "error", healthStatus: "critical", customerFlagged: true });
+      const row = buildActiveLoopFleetSummary([facts]).loops[0]!;
+      expect(row.escalation).toEqual(evaluateEscalation(facts));
+      expect(row.needsAttention).toBe(evaluateEscalation(facts).shouldEscalate);
+    });
+  });
+
+  describe("stable ordering (an operator's panel must not reshuffle between renders)", () => {
+    it("breaks severity ties by loopId, deterministically", () => {
+      const summary = buildActiveLoopFleetSummary([
+        loop({ loopId: "zeta", runStatus: "abandoned" }),
+        loop({ loopId: "alpha", runStatus: "abandoned" }),
+        loop({ loopId: "mid", runStatus: "abandoned" }),
+      ]);
+      expect(summary.needingAttention.map((r) => r.loopId)).toEqual(["alpha", "mid", "zeta"]);
+    });
+
+    it("input order never changes the output", () => {
+      const loops = [loop({ loopId: "b", healthStatus: "degraded" }), loop({ loopId: "a", runStatus: "error" }), loop({ loopId: "c" })];
+      const forward = buildActiveLoopFleetSummary(loops);
+      const reversed = buildActiveLoopFleetSummary([...loops].reverse());
+      expect(reversed).toEqual(forward);
+    });
+
+    it("does not mutate the caller's array", () => {
+      const loops = [loop({ loopId: "z" }), loop({ loopId: "a", runStatus: "error" })];
+      const before = loops.map((l) => l.loopId);
+      buildActiveLoopFleetSummary(loops);
+      expect(loops.map((l) => l.loopId)).toEqual(before);
+    });
+  });
+
+  it("keeps each loop attributed to its own tenant, so ops can see whose loop is misbehaving", () => {
+    const summary = buildActiveLoopFleetSummary([
+      loop({ loopId: "a", tenantId: "acme", runStatus: "error" }),
+      loop({ loopId: "b", tenantId: "globex" }),
+    ]);
+    expect(summary.loops.find((r) => r.loopId === "a")!.tenantId).toBe("acme");
+    expect(summary.loops.find((r) => r.loopId === "b")!.tenantId).toBe("globex");
+    expect(summary.needingAttention.map((r) => r.tenantId)).toEqual(["acme"]);
+  });
+});