diff --git a/src/queue/ai-review-orchestration.ts b/src/queue/ai-review-orchestration.ts
index e3951f898..165dcc908 100644
--- a/src/queue/ai-review-orchestration.ts
+++ b/src/queue/ai-review-orchestration.ts
@@ -26,7 +26,7 @@ import {
 import { buildPullRequestAdvisory } from "../rules/advisory";
 import { recordAuditEvent, getDecryptedRepositoryAiKey, getRepository, listCheckSummaries, listPullRequestFiles } from "../db/repositories";
 import { registerHeldLock, unregisterHeldLock } from "./held-lock-registry";
-import { recordRoutingShadow } from "../services/reviewer-routing";
+import { recordRoutingShadow, routingShadowMetrics } from "../services/reviewer-routing";
 import { judgmentAgreementMetrics, scoreJudgmentAgreement } from "../review/judgment-agreement";
 import { persistDecisionReplayPrompt } from "../review/decision-replay";
 import { createInstallationToken } from "../github/app";
@@ -992,11 +992,23 @@ export async function runAiReviewForAdvisory(
     // preferred for this repo (audit metadata only; the recap aggregates it). Same best-effort discipline
     // as the votes above: internally fail-safe, zero AI spend, and a no-signal review records nothing.
     if (result.reviewerVotes.length >= 2) {
-      await recordRoutingShadow(env, {
+      const shadow = await recordRoutingShadow(env, {
         repoFullName: args.repoFullName,
         prNumber: args.pr.number,
         actualProviders: result.reviewerVotes.map((vote) => vote.reviewer),
       });
+      // #10265: the same shadow decision onto the review's own AI trace, so how decisively the evidence
+      // favours a reviewer can be read against what that review cost. Reuses the decision `recordRoutingShadow`
+      // already returned rather than re-reading the track records. A null decision (no density / tie / read
+      // error) emits nothing, and the capture is a no-op with PostHog off or no ambient trace.
+      if (shadow) {
+        for (const metric of routingShadowMetrics(shadow)) {
+          capturePostHogAiMetric({
+            ...metric,
+            context: { repo: args.repoFullName, pullNumber: args.pr.number, agent: shadow.preferredProvider },
+          });
+        }
+      }
     }
     const findings: AdvisoryFinding[] = [];
     // #9124: the model/prompt commitments an AI-judgment finding carries — computed once, shared by both the
diff --git a/src/services/reviewer-routing.ts b/src/services/reviewer-routing.ts
index 15d224a59..9786bcb2d 100644
--- a/src/services/reviewer-routing.ts
+++ b/src/services/reviewer-routing.ts
@@ -42,6 +42,35 @@ export type RoutingShadowDecision = {
   basis: Array<{ provider: string; decided: number; precision: number }>;
 };
 
+/**
+ * PURE: the shadow decision as PostHog `$ai_metric` measurements (#10265), joined by the caller to the
+ * review's own AI trace so the routing evidence reads against the cost and model of the run that produced it.
+ *
+ * Deliberately NOT "did the shadow agree with what ran": `computeWouldHaveRouted` picks its leader OUT of
+ * `actualProviders`, so a preferred provider is by construction always one that ran, and an agreement metric
+ * would be a constant 1. What varies — and what the shadow actually measures — is how DECISIVELY the evidence
+ * favours the leader, so that is what is reported:
+ *
+ * - `routing_shadow_precision_margin` — leader precision minus runner-up. Strictly > 0 (a tie yields no
+ *   decision at all), and a margin near zero says the preference is real but weak.
+ * - `routing_shadow_preferred_precision` — the leader's own precision, without which the margin cannot be
+ *   read (0.02 between 0.97 and 0.95 is a different claim than between 0.05 and 0.03).
+ *
+ * `basis` arrives in `actualProviders` order rather than ranked, so the top two are taken by sorting here.
+ * A decision always carries at least two entries — the null guards upstream (lone reviewer, missing/floor-shy
+ * row, tie) mean this is only ever reached with a strict leader over a runner-up.
+ */
+export function routingShadowMetrics(decision: RoutingShadowDecision): Array<{ name: string; value: number }> {
+  const ranked = [...decision.basis].sort((a, b) => b.precision - a.precision);
+  const leader = ranked[0];
+  const runnerUp = ranked[1];
+  if (!leader || !runnerUp) return [];
+  return [
+    { name: "routing_shadow_precision_margin", value: leader.precision - runnerUp.precision },
+    { name: "routing_shadow_preferred_precision", value: leader.precision },
+  ];
+}
+
 /**
  * PURE: what would evidence-weighted routing have preferred for this repo, given the current track
  * records and the providers the review ACTUALLY used? Null — record nothing — unless EVERY actual
diff --git a/test/unit/reviewer-routing.test.ts b/test/unit/reviewer-routing.test.ts
index 6c167671a..bc5b13e0c 100644
--- a/test/unit/reviewer-routing.test.ts
+++ b/test/unit/reviewer-routing.test.ts
@@ -10,6 +10,7 @@ import {
   REVIEWER_VOTE_EVENT_TYPE,
   REVIEWER_VOTE_SCAN_LIMIT,
   ROUTING_MIN_DECIDED,
+  routingShadowMetrics,
 } from "../../src/services/reviewer-routing";
 import { buildRoutingRecapSection } from "../../src/services/maintainer-recap-routing";
 import { formatMaintainerRecap, runMaintainerRecap } from "../../src/services/maintainer-recap";
@@ -68,6 +69,37 @@ describe("computeWouldHaveRouted (#8229 stage 1)", () => {
   });
 });
 
+describe("routingShadowMetrics (#10265)", () => {
+  const PAIR = ["claude-code", "codex"];
+
+  it("reports the leader's margin over the runner-up AND the leader's own precision", () => {
+    const decision = computeWouldHaveRouted([record("claude-code", { precision: 0.9 }), record("codex", { precision: 0.6 })], REPO, PAIR)!;
+    expect(routingShadowMetrics(decision)).toEqual([
+      { name: "routing_shadow_precision_margin", value: 0.9 - 0.6 },
+      { name: "routing_shadow_preferred_precision", value: 0.9 },
+    ]);
+  });
+
+  it("ranks the basis itself rather than trusting its order -- basis arrives in actualProviders order, not by precision", () => {
+    // codex leads, but claude-code is first in `basis` because that is the order the review ran them in.
+    const decision = computeWouldHaveRouted([record("claude-code", { precision: 0.4 }), record("codex", { precision: 0.95 })], REPO, PAIR)!;
+    expect(decision.basis[0]?.provider).toBe("claude-code"); // unranked, as the decision contract states
+    expect(decision.preferredProvider).toBe("codex");
+    expect(routingShadowMetrics(decision)).toEqual([
+      { name: "routing_shadow_precision_margin", value: 0.95 - 0.4 },
+      { name: "routing_shadow_preferred_precision", value: 0.95 },
+    ]);
+  });
+
+  it("reports nothing without a leader AND a runner-up to compare -- a margin needs two", () => {
+    // Unreachable through computeWouldHaveRouted (its lone-reviewer guard fires first), so constructed
+    // directly: the helper must not invent a margin from a one-sided basis if a caller ever hands it one.
+    const base = { repoFullName: REPO, preferredProvider: "claude-code", actualProviders: PAIR };
+    expect(routingShadowMetrics({ ...base, basis: [] })).toEqual([]);
+    expect(routingShadowMetrics({ ...base, basis: [{ provider: "claude-code", decided: 12, precision: 0.9 }] })).toEqual([]);
+  });
+});
+
 describe("loadLiveProviderTrackRecords (#8229 stage 1 read path)", () => {
   it("joins live reviewer_vote rows to the labeled corpus; corrupt/malformed vote rows are never evidence", async () => {
     const env = createTestEnv();
diff --git a/test/unit/reviewer-vote-capture.test.ts b/test/unit/reviewer-vote-capture.test.ts
index baccd3fbb..5576b99c3 100644
--- a/test/unit/reviewer-vote-capture.test.ts
+++ b/test/unit/reviewer-vote-capture.test.ts
@@ -1,6 +1,8 @@
 import { describe, expect, it, vi } from "vitest";
 import { runAiReviewForAdvisory } from "../../src/queue/processors";
 import * as repositories from "../../src/db/repositories";
+import * as posthogModule from "../../src/selfhost/posthog";
+import * as routingModule from "../../src/services/reviewer-routing";
 import { createTestEnv } from "../helpers/d1";
 import type { Advisory, RepositorySettings } from "../../src/types";
 
@@ -105,6 +107,7 @@ describe("routing shadow orchestration hook (#8229 stage 1)", () => {
   it("a dual ok review invokes the shadow (recording nothing on a sparse corpus); the review outcome is untouched", async () => {
     const seen: string[] = [];
     const env = voteEnv(seen);
+    const metricSpy = vi.spyOn(posthogModule, "capturePostHogAiMetric").mockImplementation(() => undefined);
     const result = await runAiReviewForAdvisory(env, {
       mode: "live",
       settings: { aiReviewMode: "block" } as RepositorySettings,
@@ -121,5 +124,42 @@ describe("routing shadow orchestration hook (#8229 stage 1)", () => {
     // The votes themselves persisted — the hook runs strictly after and independently of them.
     const votes = await env.DB.prepare("SELECT COUNT(*) AS n FROM audit_events WHERE event_type = 'reviewer_vote' AND target_key = ?").bind(`${REPO}#21`).first<{ n: number }>();
     expect(votes?.n).toBe(2);
+    // #10265: no decision ⇒ no metric. An orphan margin would join to a review the shadow never ranked.
+    expect(metricSpy.mock.calls.filter((call) => call[0].name.startsWith("routing_shadow_"))).toEqual([]);
+    vi.restoreAllMocks();
+  });
+
+  it("#10265: a shadow decision emits the margin and the leader's precision onto the review's trace", async () => {
+    const seen: string[] = [];
+    const env = voteEnv(seen);
+    const metricSpy = vi.spyOn(posthogModule, "capturePostHogAiMetric").mockImplementation(() => undefined);
+    // The corpus math that PRODUCES a decision is pinned in reviewer-routing.test.ts; this test owns the
+    // wiring only, so the hook is stubbed with a decision whose basis is deliberately NOT precision-ordered.
+    vi.spyOn(routingModule, "recordRoutingShadow").mockResolvedValue({
+      repoFullName: REPO,
+      preferredProvider: "codex",
+      actualProviders: ["claude-code", "codex"],
+      basis: [
+        { provider: "claude-code", decided: 12, precision: 0.5 },
+        { provider: "codex", decided: 14, precision: 0.8 },
+      ],
+    });
+    const result = await runAiReviewForAdvisory(env, {
+      mode: "live",
+      settings: { aiReviewMode: "block" } as RepositorySettings,
+      repoFullName: REPO,
+      pr: { number: 22, title: "Add helper", body: "Adds a helper." },
+      author: "alice",
+      confirmedContributor: true,
+      advisory: advisory(22),
+    });
+    expect(result).toBeDefined();
+    const shadowMetrics = metricSpy.mock.calls.map((call) => call[0]).filter((event) => event.name.startsWith("routing_shadow_"));
+    expect(shadowMetrics).toEqual([
+      // The margin is ranked from the basis, not read off its (unordered) first entry.
+      { name: "routing_shadow_precision_margin", value: 0.8 - 0.5, context: { repo: REPO, pullNumber: 22, agent: "codex" } },
+      { name: "routing_shadow_preferred_precision", value: 0.8, context: { repo: REPO, pullNumber: 22, agent: "codex" } },
+    ]);
+    vi.restoreAllMocks();
   });
 });