diff --git a/src/renderer/src/lib/browser-history-match-budget.ts b/src/renderer/src/lib/browser-history-match-budget.ts
index 2f6468948d7..79e84e2c8ba 100644
--- a/src/renderer/src/lib/browser-history-match-budget.ts
+++ b/src/renderer/src/lib/browser-history-match-budget.ts
@@ -4,18 +4,18 @@
  * These are ceilings for catching order-of-magnitude regressions, not targets —
  * the measured numbers on a developer machine sit far under each one.
  *
- * The ceilings are asserted against the *fastest* sample of a batch, never the
- * slowest: a vitest worker sharing cores with the rest of the suite gets
- * preempted mid-measurement, so the slowest sample measures the machine while
- * the fastest still approximates the matcher.
+ * The ceilings are asserted against the *lower quartile* of a batch, never the
+ * slowest sample: a vitest worker sharing cores with the rest of the suite gets
+ * preempted mid-measurement, so the slow tail measures the machine. A quartile
+ * rather than the minimum keeps one lucky sample from carrying a regressed batch.
  *
  * Raising any value requires a fresh measurement recorded in the PR.
  */
 export const BROWSER_HISTORY_MATCH_BUDGET = {
   /** Entries prepared in one omnibox open. Tracks MAX_BROWSER_HISTORY_ENTRIES. */
   candidateCount: 200,
-  /** Milliseconds to prepare the whole corpus once (cold open), fastest sample. Measured 0.08 ms. */
+  /** Milliseconds to prepare the whole corpus once (cold open), lower quartile. Measured 0.07 ms. */
   prepareMs: 2,
-  /** Milliseconds to match the prepared corpus against one query, fastest sample. Measured 0.03 ms. */
+  /** Milliseconds to match the prepared corpus against one query, lower quartile. Measured 0.03 ms. */
   matchMs: 2
 } as const
diff --git a/src/renderer/src/lib/browser-history-match.performance.test.ts b/src/renderer/src/lib/browser-history-match.performance.test.ts
index 0ec2925530d..f40ed3e29b1 100644
--- a/src/renderer/src/lib/browser-history-match.performance.test.ts
+++ b/src/renderer/src/lib/browser-history-match.performance.test.ts
@@ -25,13 +25,19 @@ const entries = Array.from({ length: candidateCount }, (_, index) => makeEntry(i
 const WORST_QUERY = 'observability rollout'
 
 /**
- * Why the fastest sample and not p95: this runs in a vitest worker competing for
- * cores with the rest of the suite, so a slow sample records a preemption rather
- * than the matcher. Measured across idle and full-suite runs, the fastest sample
- * moved 0.05 ms -> 0.08 ms while p95 of the same batches swung 0.09 ms -> 0.50 ms.
+ * Why the lower quartile and not p95 or the minimum: this runs in a vitest worker
+ * competing for cores with the rest of the suite, so a slow sample records a
+ * preemption rather than the matcher, while a single fastest sample would pass
+ * with the other 19 runs over budget. Requiring a quarter of the batch inside the
+ * ceiling still discards the preempted tail: measured for preparation, the lower
+ * quartile moved 0.07 ms -> 0.15 ms between idle and 4x core oversubscription,
+ * while p95 of those same batches swung 0.10 ms -> 0.81 ms and the slowest sample
+ * reached 6.9 ms.
  */
-function fastestSample(samples: readonly number[]): number {
-  return Math.min(...samples)
+function lowerQuartileSample(samples: readonly number[]): number {
+  const sorted = [...samples].sort((a, b) => a - b)
+  // Nearest-rank p25: ceil(n*0.25)-1, so 20 samples pick the 5th fastest.
+  return sorted[Math.max(0, Math.ceil(sorted.length * 0.25) - 1)]
 }
 
 describe('browser history match performance budget', () => {
@@ -51,7 +57,7 @@ describe('browser history match performance budget', () => {
       prepareBrowserHistoryEntries(entries.slice())
       samples.push(performance.now() - start)
     }
-    expect(fastestSample(samples)).toBeLessThan(prepareMs)
+    expect(lowerQuartileSample(samples)).toBeLessThan(prepareMs)
   })
 
   it('matches one query against the prepared corpus within budget', () => {
@@ -67,6 +73,6 @@ describe('browser history match performance budget', () => {
       run()
       samples.push(performance.now() - start)
     }
-    expect(fastestSample(samples)).toBeLessThan(matchMs)
+    expect(lowerQuartileSample(samples)).toBeLessThan(matchMs)
   })
 })