diff --git a/bench/README.md b/bench/README.md
new file mode 100644
index 0000000..c331840
--- /dev/null
+++ b/bench/README.md
@@ -0,0 +1,36 @@
+# Nightly binding benchmarks (Refs project-minigraf/minigraf#776 P16)
+
+Run from this worktree after building the local binding. No released package
+is loaded as a fallback. Build on Linux:
+
+```sh
+python3 -m venv target/bench-venv
+source target/bench-venv/bin/activate
+pip install "maturin>=1.5,<2"
+export CARGO_TARGET_DIR="$PWD/target/p16"
+cargo build --bin uniffi-bindgen
+maturin develop --release
+target/bench-venv/bin/python bench/nightly.py --test
+```
+
+`--test` invokes each operation exactly once, checks fixed answers and reads
+no timer. Omit it only in GitHub Actions to collect median latency (nanoseconds)
+in `binding-results.json`. Do not run timed cases on a developer laptop or
+on the occupied benchmark machine.
+
+IDs use `binding/python/` and the six suffixes
+`prepared_execute/1k`, `cursor_batch_1/1k`, `cursor_batch_256/1k`,
+`set_roundtrip/2`, `predicate/1k`, `aggregate/1k`.
+
+Fixtures contain 1,000 deterministic integer facts and one two-element set.
+Setup, prepare, registration and correctness checks are outside measurement;
+foreign result decoding and cursor close are included. Set binding containers
+are created per call. Callbacks are implemented in this host language.
+
+Routing, build provenance, dedicated Bencher testbeds, provisional 5–15 minute
+allowance plus setup per job, and manual coverage dispositions are documented
+in [the core binding benchmark guide](https://github.com/project-minigraf/minigraf/blob/v3/docs/BINDING_BENCHMARKS.md).
+The core `bindings` matrix builds this shim against the measured core SHA
+using a temporary Cargo patch; this repository's production dependency pin is
+unchanged. Land this PR before the core workflow PRs. Phase B performs the
+serial workflow dispatch and upload/history validation.
diff --git a/bench/nightly.py b/bench/nightly.py
new file mode 100644
index 0000000..4ddf6a5
--- /dev/null
+++ b/bench/nightly.py
@@ -0,0 +1,106 @@
+"""#776 P16: real Python/UniFFI crossings; --test never reads a timer.
+
+Fixture/prepare/registration and correctness checks are outside measurement.
+Each sample averages 20 foreign API round trips, including JSON decode.
+"""
+import argparse
+import json
+import statistics
+import time
+from pathlib import Path
+
+from minigraf import Accumulator, Aggregate, MiniGrafDb, Predicate, Value
+
+N = 1000
+
+
+def rows(raw):
+    return json.loads(raw)["results"]
+
+
+def summary(answer):
+    return [len(answer), sum(row[0] for row in answer)]
+
+
+class Even(Predicate):
+    def test(self, value):
+        return value.value % 2 == 0
+
+
+class TotalAcc(Accumulator):
+    def __init__(self):
+        self.total = 0
+
+    def step(self, value):
+        self.total += value.value
+
+    def finish(self, count):
+        return Value.INT64(self.total)
+
+
+class Total(Aggregate):
+    def accumulator(self):
+        return TotalAcc()
+
+
+def cases():
+    db = MiniGrafDb.open_in_memory()
+    db.execute("(transact [" + " ".join(f"[:e{i} :n {i}]" for i in range(N)) + "])")
+    db.execute('(transact [[:tags :tags #{1 "x"}]])')
+    db.register_predicate("bench-even?", Even())
+    db.register_aggregate("bench-total", Total())
+    scan = db.prepare("(query [:find ?n :where [?e :n ?n] [(>= ?n $min)]])")
+    point = db.prepare("(query [:find ?n :where [?e :n $n] [?e :n ?n]])")
+    sets = db.prepare("(query [:find ?s :where [:tags :tags $s] [:tags :tags ?s]])")
+    predicate = db.prepare("(query [:find ?n :where [?e :n ?n] [(bench-even? ?n)]])")
+    aggregate = db.prepare("(query [:find (bench-total ?n) :where [?e :n ?n]])")
+
+    def drain(batch_size):
+        answer = []
+        with scan.query(min=0) as cursor:
+            while (batch := cursor.next_batch(batch_size)) is not None:
+                answer.extend(json.loads(batch))
+        return answer
+
+    return [
+        ("prepared_execute/1k", lambda: rows(point.execute(n=500)), [[500]]),
+        ("cursor_batch_1/1k", lambda: drain(1), [N, N * (N - 1) // 2]),
+        ("cursor_batch_256/1k", lambda: drain(256), [N, N * (N - 1) // 2]),
+        ("set_roundtrip/2", lambda: rows(sets.execute(s={1, "x"})), [[["x", 1]]]),
+        ("predicate/1k", lambda: rows(predicate.execute()), [500, 249500]),
+        ("aggregate/1k", lambda: rows(aggregate.execute()), [[499500]]),
+    ]
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--test", action="store_true")
+    parser.add_argument("--output", default="binding-results.json")
+    args = parser.parse_args()
+    metrics = {}
+    for name, operation, expected in cases():
+        name = "binding/python/" + name
+        check = summary if "cursor_batch" in name or "/predicate/" in name else lambda x: x
+        samples = []
+        if not args.test:
+            for _ in range(3):
+                assert check(operation()) == expected, f"incorrect result: {name}"
+        for _ in range(1 if args.test else 10):
+            if args.test:
+                actual = operation()
+            else:
+                start = time.perf_counter_ns()
+                for _ in range(20):
+                    actual = operation()
+                samples.append((time.perf_counter_ns() - start) / 20)
+            assert check(actual) == expected, f"incorrect result: {name}"
+        print("ok: " + name, flush=True)
+        if samples:
+            metrics[name] = {"latency": {"value": statistics.median(samples)}}
+    # Test mode writes no fake timing data.
+    if not args.test:
+        Path(args.output).write_text(json.dumps(metrics, indent=2) + "\n")
+
+
+if __name__ == "__main__":
+    main()