diff --git a/.github/scripts/benchmark.py b/.github/scripts/benchmark.py
new file mode 100755
index 0000000..520d72a
--- /dev/null
+++ b/.github/scripts/benchmark.py
@@ -0,0 +1,227 @@
+#!/usr/bin/env python3
+"""Self-contained PR benchmark — lives entirely in .github/scripts, nothing in the package.
+
+Subcommands:
+  run --out FILE                 generate seeded synthetic inputs, time every cp_measure get_*  -> JSON
+  compare --base F --head F [--md F]   diff two run JSONs into a head-vs-main timing table
+
+Run once per environment (PR head, main) on the SAME seeded inputs (pure-numpy generation is
+deterministic), then compare. The driver installs each env; this script only needs cp_measure
+importable plus numpy.
+"""
+
+import os
+
+for _v in (
+    "OMP_NUM_THREADS",
+    "OPENBLAS_NUM_THREADS",
+    "MKL_NUM_THREADS",
+    "NUMEXPR_NUM_THREADS",
+    "VECLIB_MAXIMUM_THREADS",
+    "NUMBA_NUM_THREADS",
+):
+    os.environ.setdefault(_v, "1")
+
+import argparse  # noqa: E402
+import json  # noqa: E402
+import signal  # noqa: E402
+import statistics  # noqa: E402
+from contextlib import contextmanager  # noqa: E402
+from pathlib import Path  # noqa: E402
+from time import perf_counter  # noqa: E402
+
+import numpy  # noqa: E402
+
+MATRIX = {"sizes": (256, 512, 1024), "counts": (16, 64, 256), "seeds": (0, 1, 2)}
+BLOBS_PER_CHANNEL = 5
+WARMUP, REPS, TIMEOUT = 1, 3, 120.0
+AFFECTED = 1.05  # report a function if any cell moves by this factor either way (faster or slower)
+
+
+# --- synthetic generator: n ellipses on a grid + random Gaussian blobs per channel --------------
+def generate(size: int, n: int, seed: int = 0):
+    rng = numpy.random.default_rng(seed)
+    yy, xx = numpy.mgrid[0:size, 0:size]
+    labels = numpy.zeros((size, size), numpy.int32)
+    if n:
+        cols = int(numpy.ceil(numpy.sqrt(n)))
+        rows = int(numpy.ceil(n / cols))
+        a, b = 0.35 * size / rows, 0.35 * size / cols
+        for k in range(n):
+            r, c = divmod(k, cols)
+            cy, cx = (r + 0.5) * size / rows, (c + 0.5) * size / cols
+            labels[((yy - cy) / a) ** 2 + ((xx - cx) / b) ** 2 <= 1] = k + 1
+    channels = []
+    for _ in range(2):  # ch0 → core features, ch0+ch1 → colocalisation
+        img = numpy.zeros((size, size))
+        for _ in range(BLOBS_PER_CHANNEL):
+            cy, cx = rng.uniform(0, size, 2)
+            s = rng.uniform(size / 10, size / 5)
+            img += numpy.exp(-((yy - cy) ** 2 + (xx - cx) ** 2) / (2 * s * s))
+        channels.append(img.astype(numpy.float32))
+    return labels, numpy.stack(channels)
+
+
+# --- timing -------------------------------------------------------------------------------------
+class _Timeout(Exception):
+    pass
+
+
+def _raise_timeout(*_):
+    raise _Timeout()
+
+
+@contextmanager
+def _time_limit(seconds: float):
+    signal.signal(signal.SIGALRM, _raise_timeout)
+    signal.setitimer(signal.ITIMER_REAL, seconds)
+    try:
+        yield
+    finally:
+        signal.setitimer(signal.ITIMER_REAL, 0)
+
+
+def _norm01(img):
+    img = img.astype("float64")
+    lo, hi = float(img.min()), float(img.max())
+    return (img - lo) / (hi - lo) if hi > lo else img - lo
+
+
+def _functions():
+    from cp_measure import bulk
+
+    out = []
+    for arity, reg in (
+        (1, bulk.get_core_measurements()),
+        (2, bulk.get_correlation_measurements()),
+    ):
+        for name, fn in reg.items():
+            out.append((name, fn, arity))
+    return out
+
+
+def _time(fn, args) -> dict:
+    try:
+        with _time_limit(TIMEOUT):
+            for _ in range(WARMUP):
+                fn(*args)
+            reps = []
+            for _ in range(REPS):
+                t = perf_counter()
+                fn(*args)
+                reps.append(perf_counter() - t)
+    except _Timeout:
+        return {"status": "timeout"}
+    except Exception as exc:
+        return {"status": "error", "error": f"{type(exc).__name__}: {exc}"[:200]}
+    return {"status": "ok", "reps": reps}
+
+
+def run(out_path: str):
+    funcs = _functions()
+    cells, results = [], {name: {} for name, _, _ in funcs}
+    for size in MATRIX["sizes"]:
+        for n in MATRIX["counts"]:
+            for seed in MATRIX["seeds"]:
+                labels, channels = generate(size, n, seed)
+                imgs = (_norm01(channels[0]), _norm01(channels[1]))
+                key = f"s{size}_n{n}_seed{seed}"
+                cells.append({"key": key, "size": size, "n_objects": n})
+                for name, fn, arity in funcs:
+                    args = (
+                        (labels, imgs[0]) if arity == 1 else (imgs[0], imgs[1], labels)
+                    )
+                    results[name][key] = _time(fn, args)
+    Path(out_path).write_text(
+        json.dumps({"cells": cells, "results": results}, indent=2)
+    )
+
+
+# --- compare ------------------------------------------------------------------------------------
+def _median_ms(results_for_fn: dict, keys: list[str]):
+    """Median (ms) over all ok rep times in a cell's seeds, or None."""
+    times = [
+        t
+        for k in keys
+        if results_for_fn.get(k, {}).get("status") == "ok"
+        for t in results_for_fn[k]["reps"]
+    ]
+    return statistics.median(times) * 1e3 if times else None
+
+
+def compare(base: dict, head: dict, commit: str = "") -> str:
+    groups: dict[tuple, list[str]] = {}
+    for e in head["cells"]:
+        groups.setdefault((e["size"], e["n_objects"]), []).append(e["key"])
+    sizes = sorted({s for s, _ in groups})
+    counts = sorted({n for _, n in groups})
+    br, hr = base["results"], head["results"]
+
+    ref = f"`{commit[:7]}`" if commit else "PR head"
+    out = [
+        f"### Benchmark — {ref} vs `main`",
+        "",
+        f"`speedup = main/head` (>1 faster, <1 slower) · median per cell · "
+        f"showing functions that moved ≥{AFFECTED:.2f}× either way",
+    ]
+
+    lo = 1.0 / AFFECTED  # a cell at or below this is a regression worth reporting
+    affected = []  # (function, {(size, count): speedup})
+    for fn in sorted(hr):
+        grid, speedups = {}, []
+        for size in sizes:
+            for n in counts:
+                m = _median_ms(br.get(fn, {}), groups.get((size, n), []))
+                h = _median_ms(hr[fn], groups.get((size, n), []))
+                grid[(size, n)] = (m / h) if (m and h) else None
+                if grid[(size, n)]:
+                    speedups.append(grid[(size, n)])
+        if speedups and (max(speedups) >= AFFECTED or min(speedups) <= lo):
+            affected.append((fn, grid))
+
+    if not affected:
+        out += ["", f"_No function moved by ≥{AFFECTED:.2f}× (faster or slower)._"]
+        return "\n".join(out)
+
+    for fn, grid in affected:
+        out += [
+            "",
+            f"#### `{fn}`",
+            "",
+            "| size \\ objects | " + " | ".join(str(n) for n in counts) + " |",
+            "|---" + "|--:" * len(counts) + "|",
+        ]
+        for size in sizes:
+            row = [
+                (f"{grid[(size, n)]:.2f}×" if grid.get((size, n)) else "—")
+                for n in counts
+            ]
+            out.append(f"| **{size}** | " + " | ".join(row) + " |")
+    return "\n".join(out)
+
+
+def main(argv=None) -> int:
+    p = argparse.ArgumentParser(description=__doc__)
+    sub = p.add_subparsers(dest="cmd", required=True)
+    r = sub.add_parser("run")
+    r.add_argument("--out", required=True)
+    c = sub.add_parser("compare")
+    c.add_argument("--base", required=True)
+    c.add_argument("--head", required=True)
+    c.add_argument("--commit", default="")
+    c.add_argument("--md")
+    a = p.parse_args(argv)
+    if a.cmd == "run":
+        run(a.out)
+    else:
+        md = compare(
+            json.loads(Path(a.base).read_text()),
+            json.loads(Path(a.head).read_text()),
+            a.commit,
+        )
+        (Path(a.md).write_text(md) if a.md else print(md))
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/.github/scripts/run_benchmark.sh b/.github/scripts/run_benchmark.sh
new file mode 100755
index 0000000..fb101ae
--- /dev/null
+++ b/.github/scripts/run_benchmark.sh
@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# Install the PR head and main into separate venvs and run benchmark.py (from this checkout) in
+# each, then compare. Each run regenerates the same seeded inputs, so nothing is shared on disk.
+# Usage: run_benchmark.sh <out-dir> <head-commit-sha>
+set -euo pipefail
+
+OUT="${1:-bench-out}"
+COMMIT="${2:-}"
+HEAD_DIR="$(pwd)"
+WORK="$(mktemp -d)"
+BENCH="$HEAD_DIR/.github/scripts/benchmark.py"
+mkdir -p "$OUT"
+trap 'git worktree remove --force "$WORK/main" 2>/dev/null || true; rm -rf "$WORK"' EXIT
+
+# six is a centrosome runtime dep not declared in its metadata; install it into the bench venvs only.
+echo "::group::PR head env"
+uv venv "$WORK/venv-head"
+uv pip install --python "$WORK/venv-head/bin/python" -e "$HEAD_DIR" six
+"$WORK/venv-head/bin/python" "$BENCH" run --out "$OUT/head.json"
+echo "::endgroup::"
+
+echo "::group::main env"
+git fetch --no-tags --depth=1 origin main
+git worktree add --detach "$WORK/main" origin/main
+uv venv "$WORK/venv-main"
+uv pip install --python "$WORK/venv-main/bin/python" -e "$WORK/main" six
+"$WORK/venv-main/bin/python" "$BENCH" run --out "$OUT/main.json"
+echo "::endgroup::"
+
+"$WORK/venv-head/bin/python" "$BENCH" compare \
+  --base "$OUT/main.json" --head "$OUT/head.json" --commit "$COMMIT" --md "$OUT/table.md"
+cat "$OUT/table.md"
diff --git a/.github/workflows/benchmark.yml b/.github/workflows/benchmark.yml
new file mode 100644
index 0000000..9c64cb9
--- /dev/null
+++ b/.github/workflows/benchmark.yml
@@ -0,0 +1,45 @@
+name: benchmark
+
+# Runs on every commit to a PR: times every public get_* on the PR head vs main and posts a
+# sticky comment with the timing table. Self-contained — workflow, tooling and generator all live
+# on the PR branch (pull_request runs the workflow from the head), so nothing is needed on main.
+
+on:
+  pull_request:
+
+permissions:
+  pull-requests: write
+
+concurrency:
+  group: bench-${{ github.event.pull_request.number }}
+  cancel-in-progress: true
+
+jobs:
+  benchmark:
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - uses: astral-sh/setup-uv@v5
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+      - name: Benchmark head vs main
+        run: bash .github/scripts/run_benchmark.sh bench-out "${{ github.event.pull_request.head.sha }}"
+      - name: Post sticky comment
+        env:
+          GH_TOKEN: ${{ github.token }}
+          REPO: ${{ github.repository }}
+          PR: ${{ github.event.pull_request.number }}
+        run: |
+          set -euo pipefail
+          { echo '<!-- cp-bench -->'; echo; cat bench-out/table.md; } > body.md
+          CID="$(gh api "repos/$REPO/issues/$PR/comments?per_page=100" \
+                 --jq 'map(select(.body | startswith("<!-- cp-bench -->")))[0].id // empty')"
+          if [ -n "$CID" ]; then
+            gh api -X PATCH "repos/$REPO/issues/comments/$CID" -F body=@body.md >/dev/null
+          else
+            gh api -X POST "repos/$REPO/issues/$PR/comments" -F body=@body.md >/dev/null
+          fi
diff --git a/.gitignore b/.gitignore
index 099eb8c..2f3d1d4 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,3 +1,6 @@
 __pycache__/
 /.agent-shell/
 .pre-commit-config.yaml
+
+# local benchmark output
+bench-out/