diff --git a/CHANGELOG.md b/CHANGELOG.md
index b9712467..3f76b220 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -17,7 +17,7 @@ All notable user-facing changes are tracked here. Public statements in release n
 - Strict typed Python surface: `py.typed`, `MatchStatus`, `Literal` policies, `TypedDict` summaries, a typed `TargetRow`, and mypy strict and pyright gates (`make typecheck`).
 - Example and snapshot testing flow: executable doctests, Rust `expect-test`-style snapshots, static API contract tests, and a typed quickstart (`make examples-test`).
 - The whole `dotmatch`/`quickdna` package now passes the strict typed gate, with Hypothesis invariants and a Python 3.9 runtime compatibility check (`make compat39`).
-- Scientific reliability gate: scalar oracle equivalence across scan/indexed/exact/Hamming policies, stream count conservation, lossless integer count round-trips above float precision, gzip/plain FASTQ equivalence, seeded panel determinism, FASTQ/count/AssaySpec fuzzing, byte-deterministic assignment/panel/feature-matrix artifacts, indexed-vs-scan agreement, panel safety-vs-distance checks, unique-only feature-matrix counts, and mutation-sensitivity cases (`make scientific-tests`). See `docs/scientific-reliability.md`.
+- Scientific reliability gate: scalar oracle equivalence across scan/indexed/exact/Hamming policies, stream count conservation, lossless integer count round-trips above float precision, gzip/plain FASTQ equivalence, seeded panel determinism, FASTQ/count/AssaySpec fuzzing, byte-deterministic assignment/panel/feature-matrix artifacts, indexed-vs-scan agreement, panel safety-vs-distance checks, unique-only feature-matrix counts, sensitivity artifact determinism, and mutation-sensitivity cases (`make scientific-tests`). See `docs/scientific-reliability.md`.
 
 ### Changed
 - Assignment result records (`MatchResult`, `AssignmentStats`, `PosteriorAssignment`, `FastqRecord`, `StreamAssignment`) are immutable `NamedTuple`s. Field access and keyword construction are unchanged. They are now tuple-compatible, so equality with a plain tuple is true; `dataclasses.asdict`, `fields`, and `replace` no longer apply.
diff --git a/docs/scientific-reliability.md b/docs/scientific-reliability.md
index ed700bfa..c1c03fbb 100644
--- a/docs/scientific-reliability.md
+++ b/docs/scientific-reliability.md
@@ -26,7 +26,8 @@ protects that contract; it does not by itself prove biological correctness.
   agreement, repeated-identifier agreement, gzip/plain equivalence, and target
   ID uniqueness are checked before assignment.
 - **Determinism.** Seeded panel design and feature-matrix artifacts are
-  reproducible for the same inputs; assignment TSVs are byte-identical.
+  reproducible for the same inputs; assignment TSVs are byte-identical, and
+  sensitivity artifacts are byte-identical apart from the recorded wall clock.
 - **Boundary fuzzing.** Generated valid and corrupted FASTQ, count TSV, and
   AssaySpec TOML inputs must either parse into typed records or raise a
   documented input error, never an unexpected exception.
diff --git a/python/tests/test_scientific_invariants.py b/python/tests/test_scientific_invariants.py
index fbfeec0d..001c622d 100644
--- a/python/tests/test_scientific_invariants.py
+++ b/python/tests/test_scientific_invariants.py
@@ -20,6 +20,7 @@
     design_panel,
     hamming_distance,
 )
+from dotmatch.sensitivity import run_sensitivity
 from hypothesis import given, settings
 from hypothesis import strategies as st
 
@@ -322,3 +323,39 @@ def test_feature_matrix_counts_only_unique_assignments(
             expected[key] = expected.get(key, 0) + 1
     observed = {(row["cell_barcode"], row["target_id"]): int(row["count"]) for row in rows}
     assert observed == expected
+
+
+def test_sensitivity_artifacts_are_deterministic_except_wall_clock(tmp_path: Path) -> None:
+    targets = tmp_path / "targets.tsv"
+    targets.write_text("target_id\ttarget_seq\nA\tACGTACGT\nB\tTTTTTTTT\n", encoding="utf-8")
+    reads = tmp_path / "reads.fastq"
+    reads.write_text(
+        "@r1\nACGTACGT\n+\nIIIIIIII\n"
+        "@r2\nACGTACGA\n+\nIIIIIIII\n"
+        "@r3\nTTTTTTTT\n+\nIIIIIIII\n",
+        encoding="ascii",
+    )
+    first = run_sensitivity(
+        targets=targets,
+        reads=reads,
+        target_start=0,
+        target_length=8,
+        out_dir=tmp_path / "first",
+        batch_size=2,
+    )
+    second = run_sensitivity(
+        targets=targets,
+        reads=reads,
+        target_start=0,
+        target_length=8,
+        out_dir=tmp_path / "second",
+        batch_size=2,
+    )
+
+    first.pop("execution", {}).pop("seconds", None)
+    second.pop("execution", {}).pop("seconds", None)
+    assert first == second
+    for name in sorted(path.name for path in (tmp_path / "first").iterdir()):
+        if name == "summary.json":
+            continue
+        assert (tmp_path / "first" / name).read_bytes() == (tmp_path / "second" / name).read_bytes(), name