diff --git a/internal-packages/run-engine/src/run-queue/fairQueueSelectionStrategy.ts b/internal-packages/run-engine/src/run-queue/fairQueueSelectionStrategy.ts
index 0e2205e413..87a2a2becf 100644
--- a/internal-packages/run-engine/src/run-queue/fairQueueSelectionStrategy.ts
+++ b/internal-packages/run-engine/src/run-queue/fairQueueSelectionStrategy.ts
@@ -209,7 +209,7 @@ export class FairQueueSelectionStrategy implements RunQueueSelectionStrategy {
   }
 
   #weightedShuffle(weightedItems: WeightedEnv[]): string[] {
-    const totalWeight = weightedItems.reduce((sum, item) => sum + item.weight, 0);
+    let totalWeight = weightedItems.reduce((sum, item) => sum + item.weight, 0);
     const result: string[] = [];
     const items = [...weightedItems];
 
@@ -224,8 +224,11 @@ export class FairQueueSelectionStrategy implements RunQueueSelectionStrategy {
       }
       index = Math.max(0, index - 1);
 
-      // Add selected item to result and remove from items
+      // Add selected item to result and remove from items. Decrement totalWeight
+      // so the next draw is scaled to the remaining items; otherwise random
+      // routinely overshoots the shrinking set and the tail item is over-picked.
       result.push(items[index].envId);
+      totalWeight -= items[index].weight;
       items.splice(index, 1);
     }
 
diff --git a/internal-packages/run-engine/src/run-queue/tests/fairQueueSelectionStrategy.test.ts b/internal-packages/run-engine/src/run-queue/tests/fairQueueSelectionStrategy.test.ts
index fa315d3d8a..bbaa62016f 100644
--- a/internal-packages/run-engine/src/run-queue/tests/fairQueueSelectionStrategy.test.ts
+++ b/internal-packages/run-engine/src/run-queue/tests/fairQueueSelectionStrategy.test.ts
@@ -1203,6 +1203,90 @@ describe("FairDequeuingStrategy", () => {
     expect(queuesByEnv["env-1"]).toBeDefined();
     expect(queuesByEnv["env-1"].length).toBe(2);
   });
+
+  redisTest(
+    "weighted env shuffle stays uniform across every position for equal-weight envs",
+    async ({ redisOptions: redis }) => {
+      const keyProducer = new RunQueueFullKeyProducer();
+      // Biases must be non-zero so env ordering goes through the weighted shuffle
+      // path rather than the plain shuffle short-circuit.
+      const strategy = new FairQueueSelectionStrategy({
+        redis,
+        keys: keyProducer,
+        defaultEnvConcurrencyLimit: 10,
+        parentQueueLimit: 100,
+        seed: "weighted-shuffle-seed",
+        biases: {
+          concurrencyLimitBias: 0.75,
+          availableCapacityBias: 0.3,
+          queueAgeRandomization: 0,
+        },
+      });
+
+      const now = Date.now();
+
+      // Four envs, each its own org, one queue each. Identical concurrency limit
+      // and current usage means identical weights, so a correct weighted shuffle
+      // should land each env in each position equally often. Insertion order is
+      // alphabetical by org, which is the order the tail-overshoot bug skews by.
+      const envIds = ["env-1", "env-2", "env-3", "env-4"];
+      for (let i = 0; i < envIds.length; i++) {
+        const orgId = `org-${i + 1}`;
+        const projectId = `proj-${i + 1}`;
+        const envId = envIds[i];
+
+        await setupQueue({
+          redis,
+          keyProducer,
+          parentQueue: "parent-queue",
+          score: now - 1000,
+          queueId: `queue-${envId}`,
+          orgId,
+          projectId,
+          envId,
+        });
+
+        await setupConcurrency({
+          redis,
+          keyProducer,
+          env: { envId, projectId, orgId, currentConcurrency: 5, limit: 10 },
+        });
+      }
+
+      const iterations = 2000;
+      // positionCounts[position][envId] = times envId landed in that position
+      const positionCounts: Array<Record<string, number>> = envIds.map(() =>
+        Object.fromEntries(envIds.map((envId) => [envId, 0]))
+      );
+
+      for (let i = 0; i < iterations; i++) {
+        const envResult = await strategy.distributeFairQueuesFromParentQueue(
+          "parent-queue",
+          `consumer-${i % 3}`
+        );
+        const result = flattenResults(envResult);
+        expect(result).toHaveLength(envIds.length);
+
+        result.forEach((queueId, position) => {
+          const envId = keyProducer.envIdFromQueue(queueId);
+          positionCounts[position][envId]++;
+        });
+      }
+
+      // For equal-weight envs the share at every position should be ~1/N. The
+      // tail-overshoot bug leaves position 0 fair but skews later positions hard
+      // (one env far below, the tail env far above). Assert each share stays
+      // within 40% of uniform at every position, which the bug violates.
+      const expectedShare = 100 / envIds.length;
+      for (let position = 0; position < envIds.length; position++) {
+        for (const envId of envIds) {
+          const share = (positionCounts[position][envId] / iterations) * 100;
+          expect(share).toBeGreaterThan(expectedShare * 0.6);
+          expect(share).toBeLessThan(expectedShare * 1.4);
+        }
+      }
+    }
+  );
 });
 
 // Helper function to flatten results for counting