diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.test.tsx
index 4a6ff77d99b..911b1df2f41 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.test.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.test.tsx
@@ -8,6 +8,7 @@ import { ApiError } from "@/lib/http/client";
 
 vi.mock("./useAutoRouterBenchmarks", () => ({ useAutoRouterBenchmarks: vi.fn() }));
 vi.mock("@/app/(dashboard)/hooks/models/useModels", () => ({ useAutoRouters: vi.fn() }));
+vi.mock("./ShadowEvalSection", () => ({ default: () => <div data-testid="shadow-eval-section" /> }));
 
 import { useAutoRouters } from "@/app/(dashboard)/hooks/models/useModels";
 
@@ -274,6 +275,14 @@ describe("AutoRouterBenchmarksTab", () => {
     expect(screen.getByText("Last 24 hours")).toBeInTheDocument();
   });
 
+  it("keeps the shadow eval section rendered while the benchmarks body is in its error state", () => {
+    mockHook({ error: new ApiError("boom", 500, {}) });
+    renderTab();
+
+    expect(screen.getByTestId("shadow-eval-section")).toBeInTheDocument();
+    expect(screen.getByText("Auto-router usage is unavailable right now")).toBeInTheDocument();
+  });
+
   it("keeps the window picker reachable while a window has no sessions", () => {
     mockHook({ data: response([]) });
     renderTab();
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.tsx
index 5d4fda765e7..ed2d8160efc 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.tsx
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/AutoRouterBenchmarksTab.tsx
@@ -31,6 +31,7 @@ import {
   type BucketRow,
 } from "./autoRouterBenchmarks";
 import { usd } from "./costOptimizationUtils";
+import ShadowEvalSection from "./ShadowEvalSection";
 import TierTurnsChart from "./TierTurnsChart";
 import { useAutoRouterBenchmarks } from "./useAutoRouterBenchmarks";
 
@@ -310,6 +311,8 @@ const AutoRouterBenchmarksTab: React.FC<AutoRouterBenchmarksTabProps> = ({ acces
         </div>
       </div>
 
+      <ShadowEvalSection accessToken={accessToken} />
+
       <BenchmarksBody
         isPending={isPending}
         error={error}
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/ShadowEvalSection.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/ShadowEvalSection.test.tsx
new file mode 100644
index 00000000000..8f2335d4ece
--- /dev/null
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/ShadowEvalSection.test.tsx
@@ -0,0 +1,343 @@
+import { render, screen } from "@testing-library/react";
+import userEvent from "@testing-library/user-event";
+import React from "react";
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+import { ApiError } from "@/lib/http/client";
+
+vi.mock("./useShadowEval", () => ({
+  useShadowEvalJobs: vi.fn(),
+  useShadowEvalJob: vi.fn(),
+  useStartShadowEval: vi.fn(),
+  useStopShadowEval: vi.fn(),
+}));
+
+const authorizedRoleMock = vi.fn(() => ({ isViewOnly: false }));
+vi.mock("@/app/(dashboard)/hooks/useAuthorized", () => ({ default: () => authorizedRoleMock() }));
+
+vi.mock("@/app/(dashboard)/hooks/keys/useKeys", () => ({
+  useInfiniteKeys: vi.fn(() => ({
+    data: {
+      pages: [
+        {
+          keys: [
+            { token: "hash-alpha", token_id: "id-1", key_name: "sk-...alpha", key_alias: "prod-alpha" },
+            { token: "hash-beta", token_id: "id-2", key_name: "sk-...beta", key_alias: "staging-beta" },
+          ],
+          total_count: 2,
+          current_page: 1,
+          total_pages: 1,
+        },
+      ],
+    },
+    isPending: false,
+    fetchNextPage: vi.fn(),
+    hasNextPage: false,
+    isFetchingNextPage: false,
+  })),
+}));
+
+vi.mock("@/app/(dashboard)/hooks/models/useModels", () => ({
+  useAutoRouters: vi.fn(() => ({
+    data: [
+      { model_name: "claude-auto", litellm_params: { model: "auto_router/claude-auto" } },
+      { model_name: "gpt-auto", litellm_params: { model: "auto_router/gpt-auto" } },
+    ],
+  })),
+}));
+
+vi.mock("@/app/(dashboard)/hooks/models/useModelCostMap", () => ({
+  useModelCostMap: vi.fn(() => ({
+    data: {
+      "claude-sonnet-5": { litellm_provider: "anthropic", mode: "chat" },
+      "gpt-4o": { litellm_provider: "openai", mode: "chat" },
+      "gemini/gemini-2.5-pro": { litellm_provider: "gemini", mode: "chat" },
+      "text-embedding-3-large": { litellm_provider: "openai", mode: "embedding" },
+    },
+  })),
+}));
+
+import ShadowEvalSection from "./ShadowEvalSection";
+import {
+  useShadowEvalJob,
+  useShadowEvalJobs,
+  useStartShadowEval,
+  useStopShadowEval,
+  type ShadowEvalJob,
+} from "./useShadowEval";
+
+const job = (overrides: Partial<ShadowEvalJob> = {}): ShadowEvalJob => ({
+  job_id: "job-1",
+  status: "running",
+  router_name: "claude-auto",
+  shadow_percentage: 10,
+  request_count: 500,
+  completed_count: 42,
+  failed_count: 1,
+  results: {
+    by_tier: [
+      {
+        group: "SIMPLE",
+        turn_count: 30,
+        real_win_rate_pct: 20.0,
+        shadow_win_rate_pct: 55.0,
+        tie_rate_pct: 25.0,
+        avg_judge_confidence: 0.81,
+      },
+      {
+        group: "REASONING",
+        turn_count: 12,
+        real_win_rate_pct: 50.0,
+        shadow_win_rate_pct: 33.3,
+        tie_rate_pct: 16.7,
+        avg_judge_confidence: 0.74,
+      },
+    ],
+    by_current_model: [
+      {
+        group: "gpt-4o",
+        turn_count: 42,
+        real_win_rate_pct: 30.0,
+        shadow_win_rate_pct: 45.0,
+        tie_rate_pct: 25.0,
+        avg_judge_confidence: 0.8,
+      },
+    ],
+    overall_shadow_win_rate_pct: 48.0,
+    overall_tie_rate_pct: 22.0,
+  },
+  cost_estimate: 45.0,
+  cost_actual: 3.21,
+  created_at: "2026-08-07T00:00:00Z",
+  ends_at: null,
+  completed_at: null,
+  api_key_id: "hashed-key-abc",
+  last_error: null,
+  ...overrides,
+});
+
+const mockHooks = ({
+  jobs = [],
+  detailsById = {},
+  error = null,
+  detailError = false,
+}: {
+  jobs?: ShadowEvalJob[];
+  detailsById?: Record<string, ShadowEvalJob>;
+  error?: Error | null;
+  detailError?: boolean;
+}) => {
+  vi.mocked(useShadowEvalJobs).mockReturnValue({
+    data: error ? undefined : jobs,
+    error,
+    isPending: false,
+  } as unknown as ReturnType<typeof useShadowEvalJobs>);
+  vi.mocked(useShadowEvalJob).mockImplementation(
+    (_token, jobId) =>
+      ({
+        data: jobId ? detailsById[jobId] : undefined,
+        isError: detailError ?? false,
+      }) as unknown as ReturnType<typeof useShadowEvalJob>,
+  );
+  const start = { mutate: vi.fn(), isPending: false };
+  const stop = { mutate: vi.fn(), isPending: false };
+  vi.mocked(useStartShadowEval).mockReturnValue(start as unknown as ReturnType<typeof useStartShadowEval>);
+  vi.mocked(useStopShadowEval).mockReturnValue(stop as unknown as ReturnType<typeof useStopShadowEval>);
+  return { start, stop };
+};
+
+describe("ShadowEvalSection", () => {
+  beforeEach(() => {
+    authorizedRoleMock.mockReturnValue({ isViewOnly: false });
+  });
+
+  it("re-offers the start form when the polled detail sees the job finish before the list does", () => {
+    mockHooks({
+      jobs: [job({ status: "running" })],
+      detailsById: { "job-1": job({ status: "completed" }) },
+    });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByText("Start a shadow eval")).toBeInTheDocument();
+  });
+
+  it("gives every active job its own card with a stop button, with the form still offered", () => {
+    mockHooks({
+      jobs: [
+        job({ job_id: "job-a", status: "running", api_key_id: "key-a" }),
+        job({ job_id: "job-b", status: "pending", api_key_id: "key-b" }),
+      ],
+    });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getAllByRole("button", { name: "Stop" })).toHaveLength(2);
+    expect(screen.getByText("Start a shadow eval")).toBeInTheDocument();
+    expect(screen.queryByText(/Previous evaluations/)).not.toBeInTheDocument();
+  });
+
+  it("renders the active card from the list row while its detail is still loading", () => {
+    mockHooks({ jobs: [job({ status: "running" })], detailsById: {} });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByRole("button", { name: "Stop" })).toBeInTheDocument();
+  });
+
+  it("hides the start form and stop button from view-only admins", () => {
+    authorizedRoleMock.mockReturnValue({ isViewOnly: true });
+    mockHooks({ jobs: [job({ status: "running" })] });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.queryByText("Start a shadow eval")).not.toBeInTheDocument();
+    expect(screen.queryByRole("button", { name: "Stop" })).not.toBeInTheDocument();
+    expect(screen.getByText("running")).toBeInTheDocument();
+  });
+
+  it("surfaces a non-403 list failure instead of posing as an empty state", () => {
+    mockHooks({ error: new Error("boom") });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByText(/Existing evaluations could not be loaded/)).toBeInTheDocument();
+    expect(screen.getByText("Start a shadow eval")).toBeInTheDocument();
+  });
+
+  it("shows a failure line instead of loading forever when the detail fetch errors", () => {
+    mockHooks({
+      jobs: [job({ status: "completed", completed_count: 12, results: null })],
+      detailsById: {},
+      detailError: true,
+    });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByText(/Results could not be loaded/)).toBeInTheDocument();
+    expect(screen.queryByText("Loading results...")).not.toBeInTheDocument();
+  });
+
+  it("shows the failure line over the collecting copy when an active job's detail errors", () => {
+    mockHooks({ jobs: [job({ status: "running", results: null })], detailsById: {}, detailError: true });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByText(/Results could not be loaded/)).toBeInTheDocument();
+    expect(screen.queryByText(/Collecting verdicts/)).not.toBeInTheDocument();
+  });
+
+  it("never claims no verdicts for a judged job whose results have not loaded yet", () => {
+    mockHooks({ jobs: [job({ status: "completed", completed_count: 12, results: null })], detailsById: {} });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByText("Loading results...")).toBeInTheDocument();
+    expect(screen.queryByText(/No verdicts were recorded/)).not.toBeInTheDocument();
+  });
+
+  it("shows the start form when there are no jobs", () => {
+    mockHooks({});
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByText("Start a shadow eval")).toBeInTheDocument();
+    expect(screen.getByText("Start shadow eval")).toBeInTheDocument();
+  });
+
+  it("renders the latest job's results with the headline stat, verdict split, and both stratifications", () => {
+    const j = job();
+    mockHooks({ jobs: [j], detailsById: { "job-1": j } });
+    render(<ShadowEvalSection accessToken="token" />);
+
+    expect(screen.getByText("Router matched or beat your current model")).toBeInTheDocument();
+    expect(screen.getByText("70.0%")).toBeInTheDocument();
+    expect(screen.getByText("of 42 judged responses")).toBeInTheDocument();
+    expect(screen.getByText(/Tie 22.0%/)).toBeInTheDocument();
+    expect(screen.getByText(/Current model won 30.0%/)).toBeInTheDocument();
+    expect(screen.getByText("gpt-4o")).toBeInTheDocument();
+    expect(screen.getByText("SIMPLE")).toBeInTheDocument();
+    expect(screen.getByText("REASONING")).toBeInTheDocument();
+    expect(screen.getByText("55.0%")).toBeInTheDocument();
+  });
+
+  it("shows the ends-in text while a job is still sampling", () => {
+    const j = job({ ends_at: new Date(Date.now() + 3 * 86_400_000).toISOString() });
+    mockHooks({ jobs: [j], detailsById: { "job-1": j } });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByText(/ends in 3 days/)).toBeInTheDocument();
+  });
+
+  it("flags rows with fewer than 30 judged turns as low sample", () => {
+    const j = job();
+    mockHooks({ jobs: [j], detailsById: { "job-1": j } });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getAllByText("(low sample)")).toHaveLength(1);
+  });
+
+  it("surfaces the last failure so a growing failed_count is diagnosable", () => {
+    const j = job({ failed_count: 7, last_error: "judge call failed: LLM Provider NOT provided" });
+    mockHooks({ jobs: [j], detailsById: { "job-1": j } });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.getByText(/LLM Provider NOT provided/)).toBeInTheDocument();
+  });
+
+  it("stops the running job from the stop button", async () => {
+    const user = userEvent.setup();
+    const j = job();
+    const { stop } = mockHooks({ jobs: [j], detailsById: { "job-1": j } });
+    render(<ShadowEvalSection accessToken="token" />);
+
+    await user.click(screen.getByText("Stop"));
+
+    expect(stop.mutate).toHaveBeenCalledWith("job-1");
+  });
+
+  it("hides the stop button and offers the start form once the latest job completed", () => {
+    const done = job({ status: "completed" });
+    mockHooks({ jobs: [done], detailsById: { "job-1": done } });
+    render(<ShadowEvalSection accessToken="token" />);
+    expect(screen.queryByText("Stop")).not.toBeInTheDocument();
+    expect(screen.getByText("Start a shadow eval")).toBeInTheDocument();
+  });
+
+  it("renders nothing for non-admins when the proxy answers 403", () => {
+    mockHooks({ error: new ApiError("forbidden", 403, {}) });
+    const { container } = render(<ShadowEvalSection accessToken="token" />);
+    expect(container).toBeEmptyDOMElement();
+  });
+
+  it("keeps the start button disabled until key, router, and judge model are picked, then submits them", async () => {
+    const user = userEvent.setup();
+    const { start } = mockHooks({});
+    render(<ShadowEvalSection accessToken="token" />);
+
+    expect(screen.getByText("Start shadow eval")).toBeDisabled();
+
+    await user.click(screen.getByPlaceholderText("Search keys by alias"));
+    await user.click(await screen.findByText("prod-alpha"));
+    await user.click(screen.getByPlaceholderText("Select an auto-router"));
+    await user.click(await screen.findByText("gpt-auto"));
+
+    expect(screen.getByText("Start shadow eval")).toBeDisabled();
+
+    await user.click(screen.getByPlaceholderText("Select a judge model"));
+    await user.click(await screen.findByRole("option", { name: /anthropic\/claude-sonnet-5/ }));
+    await user.click(screen.getByText("Start shadow eval"));
+
+    const expectedBody = {
+      api_key_id: "hash-alpha",
+      router_name: "gpt-auto",
+      shadow_percentage: 10,
+      duration_days: 7,
+      judge_model: "anthropic/claude-sonnet-5",
+    };
+    expect(start.mutate).toHaveBeenCalledWith(expectedBody);
+  });
+
+  it("keeps an older job's verdicts reachable through the previous evaluations list", async () => {
+    const user = userEvent.setup();
+    const emptyOverrides: Partial<ShadowEvalJob> = {
+      job_id: "job-new",
+      status: "pending",
+      completed_count: 0,
+      failed_count: 0,
+      results: null,
+    };
+    const current = job(emptyOverrides);
+    const older = job({ job_id: "job-old", status: "completed", results: null });
+    mockHooks({ jobs: [current, older], detailsById: { "job-new": current, "job-old": job({ job_id: "job-old" }) } });
+    render(<ShadowEvalSection accessToken="token" />);
+
+    expect(screen.queryByText("SIMPLE")).not.toBeInTheDocument();
+
+    await user.click(screen.getByRole("button", { name: /Previous evaluations \(1\)/ }));
+    expect(screen.getByText("view results")).toBeInTheDocument();
+    await user.click(screen.getByRole("button", { name: /10% via claude-auto/ }));
+
+    expect(await screen.findByText("SIMPLE")).toBeInTheDocument();
+    expect(screen.getByText("REASONING")).toBeInTheDocument();
+  });
+});
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/ShadowEvalSection.tsx b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/ShadowEvalSection.tsx
new file mode 100644
index 00000000000..7252e2ad863
--- /dev/null
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/ShadowEvalSection.tsx
@@ -0,0 +1,510 @@
+"use client";
+
+import React, { useMemo, useState } from "react";
+
+import { useInfiniteKeys } from "@/app/(dashboard)/hooks/keys/useKeys";
+import useAuthorized from "@/app/(dashboard)/hooks/useAuthorized";
+import { useModelCostMap } from "@/app/(dashboard)/hooks/models/useModelCostMap";
+import { useAutoRouters } from "@/app/(dashboard)/hooks/models/useModels";
+import { PaginatedSearchSelect } from "@/components/shared/PaginatedSearchSelect";
+import { SearchSelect, type SearchSelectOption } from "@/components/shared/SearchSelect";
+import { Badge } from "@/components/ui/badge";
+import { Button } from "@/components/ui/button";
+import { Card, CardContent, CardHeader, CardTitle } from "@/components/ui/card";
+import { Input } from "@/components/ui/input";
+import { Label } from "@/components/ui/label";
+import { Select, SelectContent, SelectItem, SelectTrigger, SelectValue } from "@/components/ui/select";
+import { Table, TableBody, TableCell, TableHead, TableHeader, TableRow } from "@/components/ui/table";
+import { ApiError } from "@/lib/http/client";
+
+import { usd } from "./costOptimizationUtils";
+import {
+  useShadowEvalJob,
+  useShadowEvalJobs,
+  useStartShadowEval,
+  useStopShadowEval,
+  type ShadowEvalJob,
+  type ShadowEvalSlice,
+} from "./useShadowEval";
+
+const pct = (value: number): string => `${value.toFixed(1)}%`;
+
+const MIN_TURNS_FOR_CONFIDENCE = 30;
+
+const isActive = (job: ShadowEvalJob): boolean => job.status === "pending" || job.status === "running";
+
+const endsIn = (endsAt: string | null | undefined): string | null => {
+  if (!endsAt) return null;
+  const remainingMs = new Date(endsAt).getTime() - Date.now();
+  if (!Number.isFinite(remainingMs)) return null;
+  if (remainingMs <= 0) return "ending now";
+  const days = Math.round(remainingMs / 86_400_000);
+  return days >= 2 ? `ends in ${days} days` : "ends within a day";
+};
+
+const STATUS_STYLES: Record<string, string> = {
+  pending: "bg-secondary text-muted-foreground",
+  running: "bg-blue-50 text-blue-700",
+  completed: "bg-emerald-50 text-emerald-700",
+};
+
+const StatusBadge: React.FC<{ status: string }> = ({ status }) => (
+  <Badge variant="secondary" className={STATUS_STYLES[status] ?? STATUS_STYLES.pending}>
+    {status}
+  </Badge>
+);
+
+const SliceTable: React.FC<{ groupHeader: string; slices: readonly ShadowEvalSlice[] }> = ({ groupHeader, slices }) => (
+  <Table>
+    <TableHeader>
+      <TableRow>
+        <TableHead>{groupHeader}</TableHead>
+        {["Judged turns", "Router wins", "Current model wins", "Ties", "Judge confidence"].map((label) => (
+          <TableHead key={label} className="text-right">
+            {label}
+          </TableHead>
+        ))}
+      </TableRow>
+    </TableHeader>
+    <TableBody>
+      {slices.map((slice) => (
+        <TableRow key={slice.group}>
+          <TableCell className="font-medium text-foreground">
+            {slice.group}
+            {slice.turn_count < MIN_TURNS_FOR_CONFIDENCE && (
+              <span className="ml-2 text-xs font-normal text-muted-foreground">(low sample)</span>
+            )}
+          </TableCell>
+          <TableCell className="text-right tabular-nums">{slice.turn_count.toLocaleString()}</TableCell>
+          <TableCell className="text-right font-medium tabular-nums text-foreground">
+            {pct(slice.shadow_win_rate_pct)}
+          </TableCell>
+          <TableCell className="text-right tabular-nums">{pct(slice.real_win_rate_pct)}</TableCell>
+          <TableCell className="text-right tabular-nums">{pct(slice.tie_rate_pct)}</TableCell>
+          <TableCell className="text-right tabular-nums">{slice.avg_judge_confidence.toFixed(2)}</TableCell>
+        </TableRow>
+      ))}
+    </TableBody>
+  </Table>
+);
+
+const VerdictBar: React.FC<{ results: NonNullable<ShadowEvalJob["results"]> }> = ({ results }) => {
+  const routerWins = results.overall_shadow_win_rate_pct;
+  const ties = results.overall_tie_rate_pct;
+  const segments = [
+    { label: "Router won", value: routerWins, fill: "bg-emerald-500" },
+    { label: "Tie", value: ties, fill: "bg-emerald-200" },
+    { label: "Current model won", value: Math.max(0, 100 - routerWins - ties), fill: "bg-muted-foreground/30" },
+  ];
+  return (
+    <div className="space-y-2 border-b px-6 py-4">
+      <div className="flex h-2 w-full overflow-hidden rounded-full" role="img" aria-label="Verdict breakdown">
+        {segments
+          .filter((segment) => segment.value > 0)
+          .map((segment) => (
+            <div key={segment.label} className={segment.fill} style={{ width: `${segment.value}%` }} />
+          ))}
+      </div>
+      <div className="flex flex-wrap gap-x-4 gap-y-1 text-xs text-muted-foreground">
+        {segments.map((segment) => (
+          <span key={segment.label} className="flex items-center gap-1.5">
+            <span className={`size-2 rounded-full ${segment.fill}`} />
+            {segment.label} {pct(segment.value)}
+          </span>
+        ))}
+      </div>
+    </div>
+  );
+};
+
+const emptyResultsText = (job: ShadowEvalJob, resultsError: boolean): string => {
+  if (resultsError) return "Results could not be loaded. Retrying.";
+  if (isActive(job)) return "Collecting verdicts. Results appear as sampled requests are judged.";
+  if (job.completed_count === 0) return "No verdicts were recorded for this job.";
+  return "Loading results...";
+};
+
+const ResultsBody: React.FC<{ job: ShadowEvalJob; resultsError?: boolean }> = ({ job, resultsError = false }) => {
+  const results = job.results;
+  if (!results || (results.by_tier.length === 0 && results.by_current_model.length === 0)) {
+    return <p className="px-6 py-8 text-center text-sm text-muted-foreground">{emptyResultsText(job, resultsError)}</p>;
+  }
+  return (
+    <>
+      <div className="flex flex-col gap-1 border-b px-6 py-4">
+        <p className="text-[11px] uppercase tracking-wide text-muted-foreground">
+          Router matched or beat your current model
+        </p>
+        <p className="text-3xl font-semibold text-foreground">
+          {pct(results.overall_shadow_win_rate_pct + results.overall_tie_rate_pct)}
+        </p>
+        <p className="text-xs text-muted-foreground">of {job.completed_count.toLocaleString()} judged responses</p>
+      </div>
+      <VerdictBar results={results} />
+      {results.by_current_model.length > 0 && (
+        <SliceTable groupHeader="Compared against" slices={results.by_current_model} />
+      )}
+      {results.by_tier.length > 0 && (
+        <div className={results.by_current_model.length > 0 ? "border-t" : ""}>
+          <SliceTable groupHeader="Prompt difficulty" slices={results.by_tier} />
+        </div>
+      )}
+    </>
+  );
+};
+
+const JobResults: React.FC<{
+  job: ShadowEvalJob;
+  onStop: () => void;
+  stopPending: boolean;
+  resultsError?: boolean;
+  readOnly?: boolean;
+}> = ({ job, onStop, stopPending, resultsError = false, readOnly = false }) => {
+  const active = isActive(job);
+  const remaining = endsIn(job.ends_at);
+  return (
+    <Card className="overflow-hidden py-0">
+      <div className="flex flex-wrap items-center justify-between gap-3 border-b px-6 py-4">
+        <div className="flex items-center gap-3">
+          <StatusBadge status={job.status} />
+          <div>
+            <p className="text-sm font-medium text-foreground">
+              Shadowing {job.shadow_percentage}% via <span className="font-mono text-xs">{job.router_name}</span>
+            </p>
+            <p className="text-xs text-muted-foreground">
+              {job.completed_count.toLocaleString()} responses judged · {job.failed_count.toLocaleString()} errored ·{" "}
+              {usd(job.cost_actual)} judge spend
+              {active && remaining ? ` · ${remaining}` : ""}
+            </p>
+          </div>
+        </div>
+        {active && !readOnly && (
+          <Button variant="outline" size="sm" onClick={onStop} disabled={stopPending}>
+            {stopPending ? "Stopping..." : "Stop"}
+          </Button>
+        )}
+      </div>
+      {job.failed_count > 0 && job.last_error && (
+        <p className="border-b bg-red-50 px-6 py-2 text-xs text-destructive">
+          Last failure: <span className="font-mono">{job.last_error}</span>
+        </p>
+      )}
+      <ResultsBody job={job} resultsError={resultsError} />
+    </Card>
+  );
+};
+
+const RECOMMENDED_JUDGE_MODELS = ["anthropic/claude-sonnet-5", "openai/gpt-4o", "gemini/gemini-2.5-pro"] as const;
+
+interface CostMapEntry {
+  litellm_provider?: string;
+  mode?: string;
+}
+
+const useJudgeModelOptions = (): SearchSelectOption[] => {
+  const { data: costMap } = useModelCostMap();
+  return useMemo(() => {
+    const pinned: SearchSelectOption[] = RECOMMENDED_JUDGE_MODELS.map((model) => ({
+      label: model,
+      value: model,
+      sublabel: "Recommended",
+    }));
+    if (!costMap) return pinned;
+    const pinnedNames = new Set<string>(RECOMMENDED_JUDGE_MODELS);
+    const chatModels = Object.entries(costMap as Record<string, CostMapEntry>)
+      .filter(([, value]) => value?.mode === "chat" && value?.litellm_provider)
+      .map(([key, value]) => (key.startsWith(`${value.litellm_provider}/`) ? key : `${value.litellm_provider}/${key}`));
+    const rest = [...new Set(chatModels)]
+      .filter((model) => !pinnedNames.has(model))
+      .toSorted((a, b) => a.localeCompare(b))
+      .map((model) => ({ label: model, value: model }));
+    return [...pinned, ...rest];
+  }, [costMap]);
+};
+
+const DURATION_OPTIONS = [
+  { value: "1", label: "1 day" },
+  { value: "3", label: "3 days" },
+  { value: "7", label: "7 days" },
+  { value: "14", label: "14 days" },
+  { value: "30", label: "30 days" },
+] as const;
+
+const Field: React.FC<{ label: string; htmlFor?: string; className?: string; children: React.ReactNode }> = ({
+  label,
+  htmlFor,
+  className,
+  children,
+}) => (
+  <div className={`space-y-1.5 ${className ?? ""}`}>
+    <Label htmlFor={htmlFor} className="text-xs">
+      {label}
+    </Label>
+    {children}
+  </div>
+);
+
+const KeySelect: React.FC<{ value: string; onChange: (token: string) => void }> = ({ value, onChange }) => {
+  const [search, setSearch] = useState("");
+  const { data, isPending, fetchNextPage, hasNextPage, isFetchingNextPage } = useInfiniteKeys(50, {
+    selectedKeyAlias: search || null,
+  });
+  const options = useMemo<SearchSelectOption[]>(
+    () =>
+      (data?.pages ?? [])
+        .flatMap((page) => page.keys)
+        .map((key) => ({
+          label: key.key_alias || key.key_name || key.token,
+          value: key.token,
+          sublabel: key.token,
+        })),
+    [data],
+  );
+  return (
+    <PaginatedSearchSelect
+      inputId="shadow-eval-key"
+      options={options}
+      value={value}
+      onValueChange={onChange}
+      onSearchChange={setSearch}
+      onLoadMore={() => void fetchNextPage()}
+      hasNextPage={hasNextPage}
+      isFetchingNextPage={isFetchingNextPage}
+      isLoading={isPending}
+      placeholder="Search keys by alias"
+      emptyText="No matching keys"
+    />
+  );
+};
+
+const StartForm: React.FC<{ accessToken: string | null }> = ({ accessToken }) => {
+  const [apiKeyId, setApiKeyId] = useState("");
+  const [routerName, setRouterName] = useState("");
+  const [percentage, setPercentage] = useState("10");
+  const [durationDays, setDurationDays] = useState("7");
+  const [judgeModel, setJudgeModel] = useState("");
+  const { data: autoRouters } = useAutoRouters();
+  const judgeModelOptions = useJudgeModelOptions();
+  const start = useStartShadowEval();
+
+  const routerOptions = useMemo<SearchSelectOption[]>(() => {
+    const names = new Set<string>();
+    for (const deployment of autoRouters ?? []) {
+      if (deployment.model_name) names.add(deployment.model_name);
+    }
+    return [...names].sort().map((name) => ({ label: name, value: name }));
+  }, [autoRouters]);
+
+  const parsedPct = Number.parseFloat(percentage);
+  const percentageValid = parsedPct >= 0.1 && parsedPct <= 100;
+  const filled = [apiKeyId, routerName, judgeModel].every((field) => field !== "");
+  const valid = Boolean(accessToken) && filled && percentageValid;
+  const startBody = {
+    api_key_id: apiKeyId,
+    router_name: routerName,
+    shadow_percentage: parsedPct,
+    duration_days: Number.parseInt(durationDays, 10),
+    judge_model: judgeModel,
+  };
+
+  return (
+    <Card size="sm">
+      <CardHeader>
+        <CardTitle className="text-sm font-medium text-foreground">Start a shadow eval</CardTitle>
+        <p className="text-xs text-muted-foreground">
+          Duplicates a sampled slice of the key&apos;s traffic through the auto-router and has an LLM judge compare both
+          answers blind. The router&apos;s answers are never served to users; judge calls bill to the shadowed key.
+        </p>
+      </CardHeader>
+      <CardContent className="space-y-3">
+        <div className="grid gap-3 sm:grid-cols-3">
+          <Field label="Key to shadow" htmlFor="shadow-eval-key">
+            <KeySelect value={apiKeyId} onChange={setApiKeyId} />
+          </Field>
+          <Field label="Auto-router">
+            <SearchSelect
+              options={routerOptions}
+              value={routerName}
+              onValueChange={setRouterName}
+              placeholder="Select an auto-router"
+              emptyText="No auto-routers configured"
+            />
+          </Field>
+          <Field label="Traffic sampled" htmlFor="shadow-eval-pct">
+            <div className="flex items-center gap-2">
+              <Input
+                id="shadow-eval-pct"
+                type="number"
+                min={0.1}
+                max={100}
+                step={0.1}
+                className="w-24"
+                value={percentage}
+                onChange={(e) => setPercentage(e.target.value)}
+              />
+              <span className="text-sm text-muted-foreground">% of traffic</span>
+            </div>
+            <div>
+              {percentage.trim() !== "" && !percentageValid && (
+                <p className="text-xs text-destructive">Enter a value from 0.1 to 100</p>
+              )}
+            </div>
+          </Field>
+          <Field label="Duration">
+            <Select value={durationDays} onValueChange={(v: string | null) => setDurationDays(v ?? "7")}>
+              <SelectTrigger className="w-full">
+                <SelectValue>{DURATION_OPTIONS.find((o) => o.value === durationDays)?.label}</SelectValue>
+              </SelectTrigger>
+              <SelectContent>
+                {DURATION_OPTIONS.map((option) => (
+                  <SelectItem key={option.value} value={option.value}>
+                    {option.label}
+                  </SelectItem>
+                ))}
+              </SelectContent>
+            </Select>
+          </Field>
+          <Field label="Judge model" className="sm:col-span-2">
+            <SearchSelect
+              options={judgeModelOptions}
+              value={judgeModel}
+              onValueChange={setJudgeModel}
+              placeholder="Select a judge model"
+              emptyText="No chat models available"
+            />
+          </Field>
+        </div>
+        <Button disabled={!valid || start.isPending} onClick={() => start.mutate(startBody)}>
+          {start.isPending ? "Starting..." : "Start shadow eval"}
+        </Button>
+      </CardContent>
+    </Card>
+  );
+};
+
+const previousSummary = (job: ShadowEvalJob): string => {
+  const results = job.results;
+  if (results) return pct(results.overall_shadow_win_rate_pct + results.overall_tie_rate_pct);
+  return job.completed_count > 0 ? "view results" : "no verdicts";
+};
+
+const PreviousJob: React.FC<{ accessToken: string | null; job: ShadowEvalJob }> = ({ accessToken, job }) => {
+  const [expanded, setExpanded] = useState(false);
+  const { data: detail, isError } = useShadowEvalJob(accessToken, expanded ? job.job_id : null);
+  const shown = detail ?? job;
+  return (
+    <div className="border-b last:border-b-0">
+      <button
+        type="button"
+        aria-expanded={expanded}
+        onClick={() => setExpanded((open) => !open)}
+        className="flex w-full flex-wrap items-center justify-between gap-3 px-6 py-3 text-left hover:bg-muted/50"
+      >
+        <div className="flex items-center gap-3">
+          <StatusBadge status={shown.status} />
+          <div>
+            <p className="text-sm font-medium text-foreground">
+              {shown.shadow_percentage}% via <span className="font-mono text-xs">{shown.router_name}</span>
+            </p>
+            <p className="text-xs text-muted-foreground">
+              {shown.completed_count.toLocaleString()} judged · {shown.failed_count.toLocaleString()} failed ·{" "}
+              {usd(shown.cost_actual)} judge spend · {new Date(shown.created_at).toLocaleDateString()}
+            </p>
+          </div>
+        </div>
+        <span className="text-sm font-medium text-foreground">{previousSummary(shown)}</span>
+      </button>
+      {expanded && (
+        <div className="border-t">
+          <ResultsBody job={shown} resultsError={isError} />
+        </div>
+      )}
+    </div>
+  );
+};
+
+const PreviousJobs: React.FC<{ accessToken: string | null; jobs: readonly ShadowEvalJob[] }> = ({
+  accessToken,
+  jobs,
+}) => {
+  const [open, setOpen] = useState(false);
+  if (jobs.length === 0) return null;
+  return (
+    <Card className="overflow-hidden py-0">
+      <button
+        type="button"
+        aria-expanded={open}
+        onClick={() => setOpen((prev) => !prev)}
+        className="flex w-full items-center justify-between gap-3 px-6 py-3 text-left hover:bg-muted/50"
+      >
+        <span className="text-sm font-medium text-foreground">Previous evaluations ({jobs.length})</span>
+        <span className="text-xs text-muted-foreground">{open ? "Hide" : "Show"}</span>
+      </button>
+      {open && (
+        <div className="border-t">
+          {jobs.map((job) => (
+            <PreviousJob key={job.job_id} accessToken={accessToken} job={job} />
+          ))}
+        </div>
+      )}
+    </Card>
+  );
+};
+
+const JobCard: React.FC<{ accessToken: string | null; job: ShadowEvalJob; readOnly: boolean }> = ({
+  accessToken,
+  job,
+  readOnly,
+}) => {
+  const { data: detail, isError } = useShadowEvalJob(accessToken, job.job_id);
+  const stop = useStopShadowEval();
+  const shown = detail ?? job;
+  return (
+    <JobResults
+      job={shown}
+      onStop={() => stop.mutate(shown.job_id)}
+      stopPending={stop.isPending}
+      resultsError={isError}
+      readOnly={readOnly}
+    />
+  );
+};
+
+const ShadowEvalSection: React.FC<{ accessToken: string | null }> = ({ accessToken }) => {
+  const { data: jobs, error, isPending } = useShadowEvalJobs(accessToken);
+  const { isViewOnly } = useAuthorized();
+  const active = (jobs ?? []).filter(isActive);
+  const finished = (jobs ?? []).filter((job) => !isActive(job));
+  const showcased = active.length > 0 ? active : finished.slice(0, 1);
+  const listed = finished.filter((job) => !showcased.includes(job));
+
+  if (error instanceof ApiError && error.status === 403) return null;
+
+  return (
+    <div className="space-y-4">
+      <div className="flex flex-wrap items-baseline gap-2">
+        <h3 className="text-lg font-semibold text-foreground">Shadow eval</h3>
+        <p className="text-sm text-muted-foreground">
+          Would the auto-router have answered as well as the models you use today? Find out on your real traffic, before
+          switching anything.
+        </p>
+      </div>
+
+      {error != null && (
+        <p className="text-sm text-destructive">Existing evaluations could not be loaded. Refresh the page to retry.</p>
+      )}
+
+      {isPending && error == null && <p className="text-sm text-muted-foreground">Loading evaluations...</p>}
+
+      {showcased.map((job) => (
+        <JobCard key={job.job_id} accessToken={accessToken} job={job} readOnly={isViewOnly} />
+      ))}
+
+      {!isViewOnly && <StartForm accessToken={accessToken} />}
+
+      <PreviousJobs accessToken={accessToken} jobs={listed} />
+    </div>
+  );
+};
+
+export default ShadowEvalSection;
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useShadowEval.test.ts b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useShadowEval.test.ts
new file mode 100644
index 00000000000..0a2c900bb7b
--- /dev/null
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useShadowEval.test.ts
@@ -0,0 +1,15 @@
+import { describe, expect, it, vi } from "vitest";
+
+vi.mock("@/lib/http/api", () => ({ $api: { useQuery: vi.fn() }, fetchClient: { POST: vi.fn() } }));
+vi.mock("@/components/molecules/notifications_manager", () => ({ default: { fromBackend: vi.fn() } }));
+
+import { shadowEvalPollMs } from "./useShadowEval";
+
+describe("shadowEvalPollMs", () => {
+  it("keeps polling while the job is active or its status is not yet known", () => {
+    expect(shadowEvalPollMs("pending")).toBe(15_000);
+    expect(shadowEvalPollMs("running")).toBe(15_000);
+    expect(shadowEvalPollMs(undefined)).toBe(15_000);
+    expect(shadowEvalPollMs("completed")).toBe(false);
+  });
+});
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useShadowEval.ts b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useShadowEval.ts
new file mode 100644
index 00000000000..10b10b6007a
--- /dev/null
+++ b/ui/litellm-dashboard/src/app/(dashboard)/cost-optimization/_components/useShadowEval.ts
@@ -0,0 +1,67 @@
+import { useMutation, useQueryClient, type QueryClient } from "@tanstack/react-query";
+
+import NotificationsManager from "@/components/molecules/notifications_manager";
+import { $api, fetchClient } from "@/lib/http/api";
+
+import type { components } from "@/lib/http/schema";
+
+export type ShadowEvalJob = components["schemas"]["GetShadowEvalJobResponse"];
+export type ShadowEvalSlice = components["schemas"]["ShadowEvalSlice"];
+export type StartShadowEvalRequest = components["schemas"]["StartShadowEvalRequest"];
+
+const LIST_PATH = "/auto_router/shadow_eval" as const;
+const DETAIL_PATH = "/auto_router/shadow_eval/{job_id}" as const;
+
+const ACTIVE_POLL_MS = 15_000;
+
+export const shadowEvalPollMs = (status: ShadowEvalJob["status"] | undefined): number | false =>
+  status === "pending" || status === "running" || status === undefined ? ACTIVE_POLL_MS : false;
+
+const invalidateShadowEval = (queryClient: QueryClient) =>
+  Promise.all([
+    queryClient.invalidateQueries({ queryKey: ["get", LIST_PATH] }),
+    queryClient.invalidateQueries({ queryKey: ["get", DETAIL_PATH] }),
+  ]);
+
+export const useShadowEvalJobs = (accessToken: string | null) =>
+  $api.useQuery("get", LIST_PATH, {}, { enabled: Boolean(accessToken), retry: 1 });
+
+export const useShadowEvalJob = (accessToken: string | null, jobId: string | null) =>
+  $api.useQuery(
+    "get",
+    DETAIL_PATH,
+    { params: { path: { job_id: jobId ?? "" } } },
+    {
+      enabled: Boolean(accessToken) && Boolean(jobId),
+      retry: 1,
+      refetchInterval: (query) => shadowEvalPollMs(query.state.data?.status),
+    },
+  );
+
+export const useStartShadowEval = () => {
+  const queryClient = useQueryClient();
+  return useMutation({
+    mutationFn: async (body: StartShadowEvalRequest) => {
+      const { data } = await fetchClient.POST("/auto_router/shadow_eval/start", { body });
+      return data;
+    },
+    onSuccess: () => invalidateShadowEval(queryClient),
+    onError: (error: unknown) =>
+      NotificationsManager.fromBackend(error instanceof Error ? error.message : "Failed to start shadow eval"),
+  });
+};
+
+export const useStopShadowEval = () => {
+  const queryClient = useQueryClient();
+  return useMutation({
+    mutationFn: async (jobId: string) => {
+      const { data } = await fetchClient.POST("/auto_router/shadow_eval/{job_id}/stop", {
+        params: { path: { job_id: jobId } },
+      });
+      return data;
+    },
+    onSuccess: () => invalidateShadowEval(queryClient),
+    onError: (error: unknown) =>
+      NotificationsManager.fromBackend(error instanceof Error ? error.message : "Failed to stop shadow eval"),
+  });
+};
diff --git a/ui/litellm-dashboard/src/app/(dashboard)/hooks/keys/useKeys.ts b/ui/litellm-dashboard/src/app/(dashboard)/hooks/keys/useKeys.ts
index 198058803eb..1881c6e4605 100644
--- a/ui/litellm-dashboard/src/app/(dashboard)/hooks/keys/useKeys.ts
+++ b/ui/litellm-dashboard/src/app/(dashboard)/hooks/keys/useKeys.ts
@@ -1,4 +1,4 @@
-import { keepPreviousData, useQuery, UseQueryResult } from "@tanstack/react-query";
+import { keepPreviousData, useInfiniteQuery, useQuery, UseQueryResult } from "@tanstack/react-query";
 import { createQueryKeys } from "../common/queryKeysFactory";
 import { getProxyBaseUrl, getGlobalLitellmHeaderName, deriveErrorMessage, handleError } from "@/components/networking";
 import { KeyResponse } from "@/components/key_team_helpers/key_list";
@@ -113,6 +113,21 @@ export const useKeys = (
   });
 };
 
+const infiniteKeyKeys = createQueryKeys("infiniteKeys");
+
+export const useInfiniteKeys = (pageSize: number, options: KeyListCallOptions = {}) => {
+  const { accessToken } = useAuthorized();
+
+  return useInfiniteQuery<KeysResponse>({
+    queryKey: infiniteKeyKeys.list({ limit: pageSize, ...options }),
+    queryFn: async ({ pageParam }) => await keyListCall(accessToken!, pageParam as number, pageSize, options),
+    initialPageParam: 1,
+    getNextPageParam: (lastPage) =>
+      lastPage.current_page < lastPage.total_pages ? lastPage.current_page + 1 : undefined,
+    enabled: Boolean(accessToken),
+  });
+};
+
 export const deletedKeyKeys = createQueryKeys("deletedKeys");
 export const useDeletedKeys = (
   page: number,