diff --git a/src/services/ai-review.ts b/src/services/ai-review.ts
index aa1510c013..c38b1f0a6e 100644
--- a/src/services/ai-review.ts
+++ b/src/services/ai-review.ts
@@ -1729,12 +1729,21 @@ export async function runGittensoryAiReview(
   const dual = combine !== "single" && (!configured || configured.length > 1);
   const freeAiCalls =
     (input.mode === "block" ? (dual ? 2 : 1) : 0) + (input.providerKey ? 0 : 1);
+  // Consensus disagreements may spend extra free calls on the order-swapped tie-break judge. Reserve the
+  // worst-case retry budget up front so the daily limiter remains a hard cap even when judge output is unstable.
+  const tieBreakAiCalls =
+    input.mode === "block" && dual && combine === "consensus"
+      ? 2 * 3 * (primaryFallback && primaryFallback !== primary.model ? 2 : 1)
+      : 0;
   // Estimate against the EFFECTIVE system prompt (`system`) so grounding's extra context is billed against the
   // budget. Flag-OFF, `system === REVIEW_SYSTEM_PROMPT`, so the estimate is byte-identical to today.
   const estimatedNeurons =
-    freeAiCalls === 0
+    (freeAiCalls === 0
       ? 0
-      : estimateNeurons(system.length + user.length, maxTokens, freeAiCalls);
+      : estimateNeurons(system.length + user.length, maxTokens, freeAiCalls)) +
+    (tieBreakAiCalls === 0
+      ? 0
+      : estimateNeurons(system.length + user.length, 512, tieBreakAiCalls));
   // FAIL-SAFE default (#budget-no-starve): the daily neuron budget is a runaway-LOOP backstop, not a normal-
   // operation gate. An absent/empty/non-numeric env var must default HIGH (the clamp max), never to a tiny value
   // that silently starves every dual-AI review into quota_exceeded — that exact misconfig (the deployed worker
diff --git a/test/unit/ai-review.test.ts b/test/unit/ai-review.test.ts
index 927a0da54c..ff774818c5 100644
--- a/test/unit/ai-review.test.ts
+++ b/test/unit/ai-review.test.ts
@@ -142,6 +142,41 @@ describe("runGittensoryAiReview gating", () => {
     expect(run).not.toHaveBeenCalled();
   });
 
+  it("reserves consensus tie-break judge retries in the shared daily neuron budget", async () => {
+    const run = vi.fn();
+    const env = createTestEnv({
+      AI: { run } as unknown as Ai,
+      AI_SUMMARIES_ENABLED: "true",
+      AI_PUBLIC_COMMENTS_ENABLED: "true",
+      AI_DAILY_NEURON_BUDGET: "600",
+    });
+    await expect(
+      runGittensoryAiReview(env, { ...baseInput, mode: "block" }),
+    ).resolves.toMatchObject({
+      status: "quota_exceeded",
+    });
+    expect(run).not.toHaveBeenCalled();
+  });
+
+  it("still reserves the tie-break budget (at the x1 fallback multiplier) when a configured reviewer has no distinct fallback model", async () => {
+    // A self-host pair with no explicit `fallback` reuses its own model (primaryFallback === primary.model),
+    // so the worst-case tie-break reservation must fall back to the x1 multiplier instead of x2 -- still
+    // non-zero, still enforced against the shared daily budget.
+    const run = vi.fn();
+    const env = createTestEnv({
+      AI: { run } as unknown as Ai,
+      AI_SUMMARIES_ENABLED: "true",
+      AI_PUBLIC_COMMENTS_ENABLED: "true",
+      AI_DAILY_NEURON_BUDGET: "600",
+    });
+    await expect(
+      runGittensoryAiReview(env, { ...baseInput, mode: "block", reviewers: [{ model: "claude-code" }, { model: "codex" }] }),
+    ).resolves.toMatchObject({
+      status: "quota_exceeded",
+    });
+    expect(run).not.toHaveBeenCalled();
+  });
+
   it("clamps a non-numeric AI_MAX_OUTPUT_TOKENS back to the default", async () => {
     const run = vi.fn(async () => ({ response: reviewJson() }));
     const env = createTestEnv({