diff --git a/CHANGELOG.md b/CHANGELOG.md
index 8600206..6612b48 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,11 @@ All notable changes follow Keep a Changelog and Semantic Versioning.
 
 ## [Unreleased]
 
+### Fixed
+
+- Invalidate Bloom filter membership caches after insertion so earlier misses cannot hide
+  newly added entries. Validate filter parameters and allocate at least one bit.
+
 ## [1.34.0] - 2026-09-30
 
 ### Added
diff --git a/HANDOVER.md b/HANDOVER.md
index 091c165..955ccf4 100644
--- a/HANDOVER.md
+++ b/HANDOVER.md
@@ -399,3 +399,28 @@ verification unless they are intentionally retained as a user-approved release a
 - Push only when explicitly requested.
 - A version is published only after its matching tag, successful release workflow, PyPI artifact,
   and GitHub release exist. Until then, keep changes under `[Unreleased]`.
+
+
+## Bloom filter correctness review (2026-10-02)
+
+The mutation regression was reproduced before fixing it: a cached negative lookup
+remained negative after insertion. Small filters could also allocate zero bits.
+The change clears membership cache entries after insertion, validates parameters,
+and rounds the bit count upward. No dependency or detection threshold changes.
+
+Observed verification on Python 3.14.2 with the frozen lockfile:
+
+- Focused Bloom and gazetteer tests: 11 passed.
+- `uv run --frozen pre-commit run --all-files`: passed.
+- `uv run --frozen mypy`: passed, 155 source files.
+- `uv run --frozen mkdocs build --strict`: passed.
+- `uv build` and `uv run --frozen python scripts/verify_release.py`: passed.
+- `uv run --frozen pytest --ignore=tests/unit/backends/test_onnx.py
+  --ignore=tests/unit/backends/test_bio_decoder.py --no-cov`: 493 passed,
+  1 skipped because Tesseract was unavailable.
+- The full suite terminated with exit 137 during ONNX tests. Full coverage and the
+  cross-platform matrix remain unverified locally; the partial run is not a
+  substitute for either gate.
+
+Reusable lesson assessment: mutation must invalidate cached negative membership
+results. The regression test records that invariant in its owning implementation.
diff --git a/src/pseudonymize/memory/bloom.py b/src/pseudonymize/memory/bloom.py
index d394e61..a790fa5 100644
--- a/src/pseudonymize/memory/bloom.py
+++ b/src/pseudonymize/memory/bloom.py
@@ -8,6 +8,10 @@ class BloomFilter:
     __slots__ = ("_bit_array", "_contains_cached", "_hash_count", "_size")
 
     def __init__(self, capacity: int, error_rate: float = 0.001) -> None:
+        if capacity <= 0:
+            raise ValueError("capacity must be positive")
+        if not 0 < error_rate < 1:
+            raise ValueError("error_rate must be between 0 and 1")
         self._size = self._get_size(capacity, error_rate)
         self._hash_count = self._get_hash_count(self._size, capacity)
         self._bit_array = bytearray((self._size + 7) // 8)
@@ -41,6 +45,9 @@ def add(self, item: str) -> None:
             digest = int.from_bytes(h.digest(), "big") % self._size
             self._bit_array[digest // 8] |= 1 << (digest % 8)
 
+        # Cached misses become stale whenever the bit array changes.
+        self._contains_cached.cache_clear()
+
     def _raw_contains(self, item: str) -> bool:
         item_bytes = item.encode("utf-8")
         base = hashlib.sha256(item_bytes)
@@ -58,7 +65,7 @@ def __contains__(self, item: str) -> bool:
     @staticmethod
     def _get_size(n: int, p: float) -> int:
         m = -(n * math.log(p)) / (math.log(2) ** 2)
-        return int(m)
+        return max(1, math.ceil(m))
 
     @staticmethod
     def _get_hash_count(m: int, n: int) -> int:
diff --git a/tests/unit/test_bloom.py b/tests/unit/test_bloom.py
index ace3c63..3e45ad7 100644
--- a/tests/unit/test_bloom.py
+++ b/tests/unit/test_bloom.py
@@ -10,3 +10,39 @@ def test_bloom_filter() -> None:
 
     assert "Jonathan" not in bf
     assert "Doe" not in bf
+
+
+def test_cached_miss_is_invalidated_after_add() -> None:
+    bf = BloomFilter(100)
+    for word in ("new-entry", "Élodie", "東京", ""):
+        assert word not in bf
+        bf.add(word)
+        assert word in bf
+        bf.add(word)
+        assert word in bf
+
+
+def test_mutation_preserves_previous_members() -> None:
+    bf = BloomFilter.from_words(iter(["alpha", "beta"]))
+    assert "alpha" in bf
+    assert "gamma" not in bf
+    bf.add("gamma")
+    assert all(word in bf for word in ("alpha", "beta", "gamma"))
+
+
+def test_small_filter_has_at_least_one_bit() -> None:
+    bf = BloomFilter(1, error_rate=0.99)
+    assert "entry" not in bf
+    bf.add("entry")
+    assert "entry" in bf
+
+
+def test_invalid_filter_parameters() -> None:
+    import pytest
+
+    for capacity in (0, -1):
+        with pytest.raises(ValueError, match="capacity must be positive"):
+            BloomFilter(capacity)
+    for rate in (0.0, -0.1, 1.0, 2.0, float("nan"), float("inf")):
+        with pytest.raises(ValueError, match="error_rate must be between 0 and 1"):
+            BloomFilter(10, rate)