diff --git a/.github/workflows/bench-pr.yml b/.github/workflows/bench-pr.yml
index cfab4f7fd3b..ff272a39bb0 100644
--- a/.github/workflows/bench-pr.yml
+++ b/.github/workflows/bench-pr.yml
@@ -72,7 +72,18 @@ jobs:
           profiling_frequency: 199
           extra_args: "--off-cpu-threshold=0.03"  # Personally tuned by @brancz
 
+      - name: Run ${{ matrix.benchmark.name }} benchmark (per-combination)
+        if: matrix.benchmark.id == 'random-access-bench'
+        shell: bash
+        env:
+          RUST_BACKTRACE: full
+          VORTEX_EXPERIMENTAL_PATCHED_ARRAY: "1"
+          FLAT_LAYOUT_INLINE_ARRAY_NODE: "1"
+        run: |
+          python3 scripts/random-access-split.py
+
       - name: Run ${{ matrix.benchmark.name }} benchmark
+        if: matrix.benchmark.id != 'random-access-bench'
         shell: bash
         env:
           RUST_BACKTRACE: full
diff --git a/.github/workflows/bench.yml b/.github/workflows/bench.yml
index 839db948add..146f0946dc0 100644
--- a/.github/workflows/bench.yml
+++ b/.github/workflows/bench.yml
@@ -85,7 +85,18 @@ jobs:
           profiling_frequency: 199
           extra_args: "--off-cpu-threshold=0.03"  # Personally tuned by @brancz
 
+      - name: Run ${{ matrix.benchmark.name }} benchmark (per-combination)
+        if: matrix.benchmark.id == 'random-access-bench'
+        shell: bash
+        env:
+          RUST_BACKTRACE: full
+          VORTEX_EXPERIMENTAL_PATCHED_ARRAY: "1"
+          FLAT_LAYOUT_INLINE_ARRAY_NODE: "1"
+        run: |
+          python3 scripts/random-access-split.py --v3
+
       - name: Run ${{ matrix.benchmark.name }} benchmark
+        if: matrix.benchmark.id != 'random-access-bench'
         shell: bash
         env:
           RUST_BACKTRACE: full
diff --git a/scripts/random-access-split.py b/scripts/random-access-split.py
new file mode 100755
index 00000000000..92b44aee7e0
--- /dev/null
+++ b/scripts/random-access-split.py
@@ -0,0 +1,102 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright the Vortex contributors
+
+"""
+Run random-access-bench once per (dataset, format, pattern, open-mode)
+then merge the per-combination outputs
+"""
+
+import argparse
+import glob
+import json
+import subprocess
+from collections.abc import Callable
+from pathlib import Path
+
+SCRIPT_DIR = Path(__file__).resolve().parent
+BINARY = "target/release_debug/random-access-bench"
+PARTS_DIR = Path("parts")
+
+DATASETS = ["taxi", "feature-vectors", "nested-lists", "nested-structs"]
+FORMATS = ["parquet", "lance", "vortex"]
+PATTERNS = ["correlated", "uniform"]
+OPEN_MODES = ["cached", "reopen"]
+
+
+def run_combinations(emit_v3: bool) -> None:
+    PARTS_DIR.mkdir(parents=True, exist_ok=True)
+    i = 0
+    for dataset in DATASETS:
+        for fmt in FORMATS:
+            for pattern in PATTERNS:
+                for open_mode in OPEN_MODES:
+                    args = [
+                        "bash",
+                        str(SCRIPT_DIR / "bench-taskset.sh"),
+                        BINARY,
+                        "--datasets",
+                        dataset,
+                        "--formats",
+                        fmt,
+                        "--patterns",
+                        pattern,
+                        "--open-mode",
+                        open_mode,
+                        "-d",
+                        "gh-json",
+                        "-o",
+                        str(PARTS_DIR / f"{i}.gh.json"),
+                    ]
+                    if emit_v3:
+                        args += ["--gh-json-v3", str(PARTS_DIR / f"{i}.v3.jsonl")]
+                    print("+", " ".join(args), flush=True)
+                    subprocess.run(args, check=True)
+                    i += 1
+
+
+"""
+This function exists only because of taxi-legacy.
+
+Every taxi invocation re-emits the pattern-less legacy taxi rows, so we need
+the merge to drop the duplicates. Otherwise we could just merge JSONL lines.
+"""
+
+
+def merge(pattern: str, key: Callable[[dict], object], out_path: str) -> None:
+    seen: set[object] = set()
+    lines: list[str] = []
+    for path in sorted(glob.glob(pattern)):
+        with open(path, encoding="utf-8") as handle:
+            for line in handle:
+                line = line.strip()
+                if not line:
+                    continue
+                identity = key(json.loads(line))
+                if identity in seen:
+                    continue
+                seen.add(identity)
+                lines.append(line)
+    Path(out_path).write_text("".join(line + "\n" for line in lines), encoding="utf-8")
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--v3",
+        action="store_true",
+        help="merge --gh-json-v3 records into results.v3.jsonl",
+    )
+    args = parser.parse_args()
+
+    run_combinations(args.v3)
+    merge(f"{PARTS_DIR}/*.gh.json", lambda record: record["name"], "results.json")
+    if args.v3:
+        merge(
+            f"{PARTS_DIR}/*.v3.jsonl",
+            lambda record: (record["kind"], record["dataset"], record["format"]),
+            "results.v3.jsonl",
+        )
+
+
+if __name__ == "__main__":
+    main()