diff --git a/sorts/benchmark_sorts.py b/sorts/benchmark_sorts.py
new file mode 100644
index 000000000000..72b6b487e1a9
--- /dev/null
+++ b/sorts/benchmark_sorts.py
@@ -0,0 +1,145 @@
+#!/usr/bin/env -S uv run --script
+
+"""
+Benchmark several sorting algorithms on the same random datasets.
+
+This is a *reference* benchmark, not a rigorous one: it times each algorithm on a
+few shared, randomly generated integer datasets and prints a small comparison
+table.  It exists so that visitors can see the practical cost of the different
+strategies in this directory side by side, without embedding timing code inside
+the individual algorithm modules (which keeps those files clean, import-cheap,
+and focused on being readable reference implementations).
+
+Run it from the repository root:
+
+    python -m sorts.benchmark_sorts
+
+The individual algorithms are imported from their own modules, so this file never
+re-implements a sort.
+"""
+
+import random
+import sys
+from collections.abc import Callable, Sequence
+from itertools import pairwise
+from timeit import timeit
+from typing import Protocol
+
+from sorts.bubble_sort import bubble_sort_iterative
+from sorts.cocktail_shaker_sort import cocktail_shaker_sort
+from sorts.comb_sort import comb_sort
+from sorts.gnome_sort import gnome_sort
+from sorts.heap_sort import heap_sort
+from sorts.insertion_sort import insertion_sort
+from sorts.merge_sort import merge_sort
+from sorts.quick_sort import quick_sort
+from sorts.selection_sort import selection_sort
+from sorts.shell_sort import shell_sort
+from sorts.tim_sort import tim_sort
+
+# name -> callable.  Every callable accepts a list and returns the sorted list.
+SORTS: dict[str, Callable[[list[int]], Sequence[int]]] = {
+    "bubble_sort": bubble_sort_iterative,
+    "cocktail_shaker_sort": cocktail_shaker_sort,
+    "comb_sort": comb_sort,
+    "gnome_sort": gnome_sort,
+    "heap_sort": heap_sort,
+    "insertion_sort": insertion_sort,
+    "merge_sort": merge_sort,
+    "quick_sort": quick_sort,
+    "selection_sort": selection_sort,
+    "shell_sort": shell_sort,
+    "tim_sort": tim_sort,
+}
+
+
+def is_sorted(collection: Sequence[int]) -> bool:
+    """
+    Return True if every element is less than or equal to the next one.
+
+    >>> is_sorted([1, 2, 2, 3])
+    True
+    >>> is_sorted([1, 3, 2])
+    False
+    >>> is_sorted([])
+    True
+    """
+    return all(a <= b for a, b in pairwise(collection))
+
+
+def all_sorts_agree(data: list[int]) -> bool:
+    """
+    Return True if every algorithm in ``SORTS`` sorts ``data`` correctly.
+
+    Each algorithm is given a fresh copy of the data (some sort in place), and its
+    result is checked against Python's built-in ``sorted`` as the ground truth.
+
+    >>> all_sorts_agree([5, 1, 4.2, 2, 8.5, 0, 2])
+    True
+    >>> all_sorts_agree([])
+    True
+    >>> all_sorts_agree([42])
+    True
+    >>> all_sorts_agree(list(range(5, -6, -1)))
+    True
+    >>> all_sorts_agree(list("Python"))
+    True
+    """
+    expected = sorted(data)
+    return all(list(sort_fn(data.copy())) == expected for sort_fn in SORTS.values())
+
+
+class Comparable(Protocol):
+    def __lt__(self, other: object, /) -> bool: ...
+
+
+def benchmark[T: Comparable](data: list[T], number: int = 1) -> dict[str, float]:
+    """
+    Time every algorithm in ``SORTS`` on a copy of ``data``.
+
+    Returns a mapping of algorithm name to the elapsed seconds for ``number``
+    repetitions.  Each timed call receives its own fresh copy so in-place sorts do
+    not hand an already-sorted list to the next repetition.
+
+    >>> benchmark([])
+    Traceback (most recent call last):
+        ...
+    ValueError: Please provide a non-empty dataset
+    >>> benchmark([1], number=0)
+    Traceback (most recent call last):
+        ...
+    ValueError: Number of repetitions must be positive
+    """
+    if not data:
+        raise ValueError("Please provide a non-empty dataset")
+    if number <= 0:
+        raise ValueError("Number of repetitions must be positive")
+    timings: dict[str, float] = {}
+    for name, sort_fn in SORTS.items():
+        timings[name] = timeit(lambda fn=sort_fn: fn(data.copy()), number=number)
+    return timings
+
+
+def main() -> None:
+    # A couple of the imported algorithms (e.g. tim_sort) merge recursively, so
+    # give them headroom to sort the largest dataset without hitting the limit.
+    sys.setrecursionlimit(10_000)
+    sizes = (100, 1_000, 3_000)
+    random.seed(0)
+    datasets = {size: [random.randint(0, size) for _ in range(size)] for size in sizes}
+
+    header = "algorithm".ljust(22) + "".join(f"{size:>12}" for size in sizes)
+    print(header)
+    print("-" * len(header))
+
+    per_size = {size: benchmark(data) for size, data in datasets.items()}
+    for name in SORTS:
+        row = name.ljust(22)
+        row += "".join(f"{per_size[size][name]:>12.4f}" for size in sizes)
+        print(row)
+
+    print("\nseconds per sort (lower is better); dataset = uniform random ints")
+
+
+if __name__ == "__main__":
+    main()