diff --git a/CHANGES.rst b/CHANGES.rst
index 9d3149f3..6c4b1c55 100644
--- a/CHANGES.rst
+++ b/CHANGES.rst
@@ -2,6 +2,15 @@
 Changelog
 =========
 
+development version
+-------------------
+
+* :issue:`892`: Fixed: An absolute number of allowed errors (``-e N`` with N >= 1, or
+  ``max_errors=N``) is now honoured for every adapter length. Previously,
+  floating-point rounding reduced the number of allowed errors by one for some
+  adapter lengths: for example, ``-e 1`` allowed no errors at all for a 49 nt
+  adapter and ``-e 2`` allowed only one error for a 98 nt adapter.
+
 v5.2 (2025-10-23)
 -----------------
 
diff --git a/src/cutadapt/_align.pyx b/src/cutadapt/_align.pyx
index 1ce926f0..3268b570 100644
--- a/src/cutadapt/_align.pyx
+++ b/src/cutadapt/_align.pyx
@@ -18,6 +18,11 @@ DEF MISMATCH_SCORE = -1
 DEF INSERTION_SCORE = -2
 DEF DELETION_SCORE = -2
 
+# Products such as 49 * (1 / 49) evaluate to 0.9999999999999999 in floating
+# point, which would turn "one allowed error" (-e 1) into "no errors". The
+# tolerance is far below the resolution of any meaningful error rate.
+DEF ERROR_RATE_TOLERANCE = 1e-9
+
 
 # structure for a DP matrix entry
 ctypedef struct _Entry:
@@ -340,7 +345,7 @@ cdef class Aligner:
         cdef int i, j
 
         # maximum no. of errors
-        cdef int k = <int> (max_error_rate * m)
+        cdef int k = <int> (max_error_rate * m + ERROR_RATE_TOLERANCE)
 
         # Determine largest and smallest column we need to compute
         cdef int max_n = n
@@ -510,7 +515,7 @@ cdef class Aligner:
                             cur_effective_length = self.effective_length
                     is_acceptable = (
                         length >= self._min_overlap
-                        and cost <= cur_effective_length * max_error_rate
+                        and cost <= cur_effective_length * max_error_rate + ERROR_RATE_TOLERANCE
                     )
                     best_length = m + min(best.origin, 0)
 
@@ -556,7 +561,7 @@ cdef class Aligner:
 
                 is_acceptable = (
                     length >= self._min_overlap
-                    and cost <= cur_effective_length * max_error_rate
+                    and cost <= cur_effective_length * max_error_rate + ERROR_RATE_TOLERANCE
                 )
                 best_length = best.ref_stop + min(best.origin, 0)
 
@@ -630,7 +635,7 @@ cdef class PrefixComparer:
                 raise ValueError("Cannot have only N wildcards in the sequence")
         if not (0 <= max_error_rate <= 1.):
             raise ValueError("max_error_rate must be between 0 and 1")
-        self.max_k = int(max_error_rate * self.effective_length)
+        self.max_k = int(max_error_rate * self.effective_length + ERROR_RATE_TOLERANCE)
         if min_overlap < 1:
             raise ValueError("min_overlap must be at least 1")
         self.min_overlap = min_overlap
diff --git a/src/cutadapt/adapters.py b/src/cutadapt/adapters.py
index 3bda8e4f..4c277763 100644
--- a/src/cutadapt/adapters.py
+++ b/src/cutadapt/adapters.py
@@ -20,6 +20,7 @@
     SuffixComparer,
     edit_environment,
     hamming_sphere,
+    max_errors_for_length,
 )
 from .kmer_heuristic import create_positions_and_kmers, kmer_probability_analysis
 
@@ -1375,7 +1376,7 @@ def _accept(cls, adapter: SingleAdapter, prefix: bool):
             raise ValueError("Wildcards in the read not supported")
         if adapter.adapter_wildcards:
             raise ValueError("Wildcards in the adapter not supported")
-        k = int(len(adapter) * adapter.max_error_rate)
+        k = max_errors_for_length(len(adapter), adapter.max_error_rate)
         if k > 3:
             raise ValueError("Error rate too high")
 
@@ -1397,7 +1398,7 @@ def _make_index(self) -> Tuple[List[int], "AdapterIndexDict", int]:
         start_time = time.time()
         max_k = max(
             (
-                int(adapter.max_error_rate * len(adapter.sequence))
+                max_errors_for_length(len(adapter.sequence), adapter.max_error_rate)
                 for adapter in self._adapters
                 if adapter.indels
             ),
@@ -1415,7 +1416,7 @@ def _make_index(self) -> Tuple[List[int], "AdapterIndexDict", int]:
         ambiguous = {}
         for adapter in self._adapters:
             sequence = adapter.sequence
-            k = int(adapter.max_error_rate * len(sequence))
+            k = max_errors_for_length(len(sequence), adapter.max_error_rate)
 
             if adapter.indels:
                 for s, errors, matches in edit_environment(sequence, k):
diff --git a/src/cutadapt/align.py b/src/cutadapt/align.py
index eaa5cbc5..6a9b4b84 100644
--- a/src/cutadapt/align.py
+++ b/src/cutadapt/align.py
@@ -7,6 +7,7 @@
     "hamming_environment",
     "edit_environment",
     "edit_distance",
+    "max_errors_for_length",
 ]
 
 from enum import IntFlag
@@ -20,6 +21,20 @@
     edit_environment,
 )
 
+# Products such as 49 * (1 / 49) evaluate to 0.9999999999999999 in floating
+# point, which would turn "one allowed error" (-e 1) into "no errors". The
+# tolerance is far below the resolution of any meaningful error rate. The same
+# constant is used in _align.pyx.
+ERROR_RATE_TOLERANCE = 1e-9
+
+
+def max_errors_for_length(length: int, max_error_rate: float) -> int:
+    """
+    Return the number of errors allowed in an alignment covering *length*
+    characters of the adapter at the given maximum error rate.
+    """
+    return int(length * max_error_rate + ERROR_RATE_TOLERANCE)
+
 
 class EndSkip(IntFlag):
     """
diff --git a/src/cutadapt/kmer_heuristic.py b/src/cutadapt/kmer_heuristic.py
index b25b1899..c54ca915 100644
--- a/src/cutadapt/kmer_heuristic.py
+++ b/src/cutadapt/kmer_heuristic.py
@@ -2,6 +2,8 @@
 from typing import List, Optional, Set, Tuple
 from collections import defaultdict
 
+from .align import max_errors_for_length
+
 
 def kmer_chunks(sequence: str, chunks: int) -> Set[str]:
     """
@@ -92,7 +94,7 @@ def create_back_overlap_searchsets(
     max_error = 0
     search_sets: List[SearchSet] = []
     for i in range(adapter_length + 1):
-        if int(i * error_rate) > max_error:
+        if max_errors_for_length(i, error_rate) > max_error:
             error_lengths.append((max_error, i - 1))
             max_error += 1
     error_lengths.append((max_error, adapter_length))
@@ -139,7 +141,7 @@ def create_positions_and_kmers(
     This function returns the positions and the accompanying words while also
     taking into account partial overlap for back and front adapters.
     """
-    max_errors = int(len(adapter) * error_rate)
+    max_errors = max_errors_for_length(len(adapter), error_rate)
     search_sets = []
     if back_adapter:
         search_sets.extend(
diff --git a/tests/test_adapters.py b/tests/test_adapters.py
index f24ecdfd..e3581bd9 100644
--- a/tests/test_adapters.py
+++ b/tests/test_adapters.py
@@ -703,3 +703,32 @@ def test_very_long_adapter_issue_749():
     assert match.astart == 0
     assert match.astop == 20
     assert match.errors == 0
+
+
+def test_back_adapter_absolute_errors_floating_point_boundary():
+    # 49 * (1 / 49) is slightly less than 1 in floating point;
+    # one error must nevertheless be allowed
+    sequence = "CAGATTTTCATATTATGCAGAAAATCTACTTCGCCTGATACGAGTCGGT"
+    assert len(sequence) == 49
+    mutated = sequence[:20] + "T" + sequence[21:]
+    adapter = BackAdapter(sequence, max_errors=1)
+    match = adapter.match_to("GGGG" + mutated + "GGGG")
+    assert match is not None
+    assert match.errors == 1
+    assert (match.rstart, match.rstop) == (4, 53)
+
+
+def test_indexed_prefix_adapters_absolute_errors_floating_point_boundary():
+    sequence1 = "CAGATTTTCATATTATGCAGAAAATCTACTTCGCCTGATACGAGTCGGT"
+    sequence2 = "TGCCGTAATCGAGGCTAGTTCACCGATAGGTCCTAAGCTTGACGATCAA"
+    assert len(sequence1) == len(sequence2) == 49
+    adapters = [
+        PrefixAdapter(sequence1, max_errors=1, indels=False),
+        PrefixAdapter(sequence2, max_errors=1, indels=False),
+    ]
+    ma = IndexedPrefixAdapters(adapters)
+    mutated = sequence1[:20] + "T" + sequence1[21:]
+    match = ma.match_to(mutated + "GGGG")
+    assert match is not None
+    assert match.adapter is adapters[0]
+    assert match.errors == 1
diff --git a/tests/test_align.py b/tests/test_align.py
index 505107cd..7b9fbe38 100644
--- a/tests/test_align.py
+++ b/tests/test_align.py
@@ -10,6 +10,7 @@
     hamming_sphere,
     edit_environment,
     edit_distance,
+    max_errors_for_length,
     naive_edit_environment,
     slow_edit_environment,
     py_edit_environment,
@@ -495,3 +496,30 @@ def test_edit_environment(k, s, environment_func):
             assert m == score
             assert m <= len(s), (s, t, dist)
             assert m <= len(t), (s, t, dist)
+
+
+def test_max_errors_for_length_floating_point_boundary():
+    # 49 * (1 / 49) == 0.9999999999999999
+    assert max_errors_for_length(49, 1 / 49) == 1
+    assert max_errors_for_length(98, 2 / 98) == 2
+    assert max_errors_for_length(47, 3 / 47) == 3
+    assert max_errors_for_length(48, 1 / 49) == 0
+    assert max_errors_for_length(10, 0.1) == 1
+    assert max_errors_for_length(9, 0.1) == 0
+
+
+def test_absolute_errors_floating_point_boundary():
+    # An adapter of length 49 with max_errors=1 gets max_error_rate 1/49,
+    # and 49 * (1 / 49) is slightly less than 1 in floating point
+    reference = "CAGATTTTCATATTATGCAGAAAATCTACTTCGCCTGATACGAGTCGGT"
+    assert len(reference) == 49
+    mutated = reference[:20] + "T" + reference[21:]
+    assert mutated != reference
+    aligner = Aligner(reference, 1 / 49, flags=Where.BACK)
+    result = aligner.locate("GGGG" + mutated + "GGGG")
+    assert result is not None
+    assert result[5] == 1  # errors
+    comparer = PrefixComparer(reference, 1 / 49)
+    result = comparer.locate(mutated + "GGGG")
+    assert result is not None
+    assert result[5] == 1