@@ -24,6 +24,21 @@ adheres to [Semantic Versioning](https://semver.org/).
   is now rejected instead of being coerced to `123` and `120.5`. US and Indian
   lakh formats are unchanged. **Default-output change:** a European-format
   currency string in a money column now cleans to its correct magnitude.
+- A value the caller has explicitly declared permitted is no longer discarded
+  as a null marker. `fd.clean`'s `normalize_sentinels` step applied the
+  built-in sentinel set unconditionally, so `"NA"` in an ISO-3166 country
+  column (Namibia) and `"None"` in a brand column became missing even when
+  `allowed_values` for that column listed them. `fd.validate_fields` has
+  honoured the opposite rule since the `TestAllowedValuesBeatNullMarkers`
+  regression — "when the schema literally allows a value, it is a value, not a
+  missing marker" — so the same declaration was respected by one public API
+  and ignored by another. The only previous escapes were protecting the column
+  outright, which disables every other repair, or `normalize_sentinels=False`,
+  which is global. A column's declared `allowed_values` (whether passed through
+  `semantic_context` or compiled from a `context=` policy) now removes those
+  tokens from that column's sentinel set, matched casefolded and trimmed. The
+  exemption is scoped to the declaring column, and a column with no declaration
+  is unchanged — `"NA"` with no vocabulary is still read as missing.
 - `engine="duckdb"` no longer silently changes temporal values on the fully
   native path (`strategy="conservative"`, `fix_dtypes=False`). A nanosecond
   `timedelta64[ns]` column was truncated to DuckDB's microsecond `INTERVAL`
 
@@ -162,16 +162,40 @@ def _text_columns(df: pd.DataFrame) -> list:
     ]
 
 
+def declared_vocabulary(config: CleanConfig, column: object) -> frozenset[str]:
+    """Casefolded ``allowed_values`` the caller declared for *column*.
+
+    An explicit vocabulary outranks a generic null marker: ``"NA"`` is Namibia
+    in an ISO-3166 column and ``"None"`` is a real brand. ``fieldcheck`` has
+    applied that rule since the TestAllowedValuesBeatNullMarkers regression;
+    this is the same rule for the cleaning path.
+    """
+    context = getattr(config, "semantic_context", None)
+    if not isinstance(context, dict):
+        return frozenset()
+    columns = context.get("columns")
+    if not isinstance(columns, dict):
+        return frozenset()
+    hints = columns.get(str(column))
+    if not isinstance(hints, dict):
+        return frozenset()
+    allowed = hints.get("allowed_values")
+    if not allowed or isinstance(allowed, (str, bytes)):
+        return frozenset()
+    try:
+        return frozenset(str(v).casefold().strip() for v in allowed)
+    except TypeError:
+        return frozenset()
+
+
 def clean_strings(df: pd.DataFrame, config: CleanConfig, report: CleanReport) -> pd.DataFrame:
     """Apply whitespace stripping and sentinel→missing to text-capable columns.
 
     Categorical columns with text categories are repaired too; they keep their
     categorical dtype (see :func:`normalize_categorical`).
     """
     if not (
-        config.strip_whitespace
-        or config.normalize_sentinels
-        or config.string_case is not None
+        config.strip_whitespace or config.normalize_sentinels or config.string_case is not None
     ):
         return df
     sentinels = active_sentinels(config)
@@ -182,20 +206,34 @@ def clean_strings(df: pd.DataFrame, config: CleanConfig, report: CleanReport) ->
         if str(col) in protected:
             continue  # context-protected columns must stay byte-identical
         s = df[col]
+        # A token the caller declared as a permitted value is a value, not a
+        # missing marker, so it is removed from this column's sentinel set.
+        column_sentinels = sentinels - declared_vocabulary(config, col)
         normalize = (
             normalize_categorical if isinstance(s.dtype, pd.CategoricalDtype) else normalize_text
         )
-        normalized, n_stripped, n_sentinels, n_case = normalize(s, config, sentinels)
+        normalized, n_stripped, n_sentinels, n_case = normalize(s, config, column_sentinels)
         if n_stripped:
-            report.add("strip_whitespace", "trimmed surrounding whitespace",
-                       column=str(col), count=n_stripped)
+            report.add(
+                "strip_whitespace",
+                "trimmed surrounding whitespace",
+                column=str(col),
+                count=n_stripped,
+            )
         if n_sentinels:
-            report.add("normalize_sentinels",
-                       'replaced sentinel strings ("N/A", "-", "", …) with missing',
-                       column=str(col), count=n_sentinels)
+            report.add(
+                "normalize_sentinels",
+                'replaced sentinel strings ("N/A", "-", "", …) with missing',
+                column=str(col),
+                count=n_sentinels,
+            )
         if n_case:
-            report.add("normalize_case", f"converted text to {config.string_case}",
-                       column=str(col), count=n_case)
+            report.add(
+                "normalize_case",
+                f"converted text to {config.string_case}",
+                column=str(col),
+                count=n_case,
+            )
         if n_stripped or n_sentinels or n_case:
             df[col] = normalized
     return df
@@ -0,0 +1,135 @@
+"""A declared vocabulary outranks a generic null marker in fd.clean (FD2-001).
+
+``fieldcheck`` has applied this rule since the ``TestAllowedValuesBeatNullMarkers``
+regression: "'NA' may be Namibia: when the schema literally allows a value, it
+is a value, not a missing marker" (``fieldcheck.py:466``). ``fd.clean`` did not.
+``normalize_sentinels`` applied ``DEFAULT_SENTINELS`` unconditionally, so a
+caller who had explicitly declared ``NA`` as permitted still lost it, and the
+only escapes were protecting the column outright (which disables every other
+repair) or turning sentinel handling off globally for every column.
+
+The gauntlet fixture states the intended behaviour directly: "without a
+vocabulary containing 'NA', the null-marker reading wins; with allowed_values
+that includes 'NA' the value survives".
+"""
+
+from __future__ import annotations
+
+import pandas as pd
+import pytest
+
+import freshdata as fd
+
+ISO = ["US", "GB", "FR", "DE", "JP", "CA", "AU", "BR", "NA"]
+
+
+def _frame():
+    return pd.DataFrame({"cust": [f"c{i}" for i in range(9)], "country": list(ISO)})
+
+
+def test_declared_allowed_values_keep_a_sentinel_looking_value():
+    """'NA' is Namibia when the caller says the column allows it."""
+    out = fd.clean(
+        _frame(),
+        verbose=False,
+        semantic_context={"columns": {"country": {"allowed_values": ISO}}},
+    )
+    assert out["country"].iloc[8] == "NA"
+
+
+def test_a_compiled_context_policy_reaches_the_sentinel_step():
+    """The documented natural-language route must work as well as the dict."""
+    out = fd.clean(
+        _frame(),
+        verbose=False,
+        context="Allowed country values are US, GB, FR, DE, JP, CA, AU, BR, NA.",
+    )
+    assert out["country"].iloc[8] == "NA"
+
+
+def test_without_a_declaration_the_null_marker_reading_still_wins():
+    """Unchanged default. This is the gauntlet 'sentinel_collision' gold label."""
+    out = fd.clean(_frame(), verbose=False)
+    assert pd.isna(out["country"].iloc[8])
+
+
+def test_a_vocabulary_that_excludes_na_still_nulls_it():
+    """Mirrors fieldcheck's test_na_outside_vocabulary_is_still_a_null_marker."""
+    out = fd.clean(
+        _frame(),
+        verbose=False,
+        semantic_context={"columns": {"country": {"allowed_values": ["US", "GB", "FR"]}}},
+    )
+    assert pd.isna(out["country"].iloc[8])
+
+
+def test_the_exemption_is_scoped_to_the_declaring_column():
+    """Declaring NA for one column must not rescue it everywhere."""
+    df = pd.DataFrame(
+        {
+            "country": ["US", "NA", "DE", "FR", "GB", "JP", "CA", "AU", "BR"],
+            "note": ["a", "NA", "c", "d", "e", "f", "g", "h", "i"],
+        }
+    )
+    out = fd.clean(
+        df,
+        verbose=False,
+        semantic_context={"columns": {"country": {"allowed_values": ISO}}},
+    )
+    assert out["country"].iloc[1] == "NA"
+    assert pd.isna(out["note"].iloc[1])
+
+
+@pytest.mark.parametrize("declared", ["na", "Na", " NA "])
+def test_vocabulary_matching_is_casefolded_and_trimmed(declared):
+    """Consistent with fieldcheck and with extra_sentinels normalisation."""
+    out = fd.clean(
+        _frame(),
+        verbose=False,
+        semantic_context={"columns": {"country": {"allowed_values": ["US", declared]}}},
+    )
+    assert out["country"].iloc[8] == "NA"
+
+
+@pytest.mark.parametrize("allowed", [None, [], "NA", 42, {"nested": "dict"}])
+def test_a_malformed_vocabulary_is_ignored_not_fatal(allowed):
+    """A bad hint must not crash cleaning, and must not grant an exemption.
+
+    A bare string is rejected deliberately: iterating it would treat 'N' and
+    'A' as separate permitted values.
+    """
+    out = fd.clean(
+        _frame(),
+        verbose=False,
+        semantic_context={"columns": {"country": {"allowed_values": allowed}}},
+    )
+    assert pd.isna(out["country"].iloc[8])
+
+
+def test_the_none_brand_case_behaves_the_same_way():
+    """'None' is a real brand; the same declaration rescues it."""
+    df = pd.DataFrame(
+        {
+            "k": [f"r{i}" for i in range(9)],
+            "brand": [
+                "Acme",
+                "Globex",
+                "Initech",
+                "Stark",
+                "Wayne",
+                "Hooli",
+                "Umbrella",
+                "Vandelay",
+                "None",
+            ],
+        }
+    )
+    plain = fd.clean(df, verbose=False)
+    assert pd.isna(plain["brand"].iloc[8])
+
+    declared = fd.clean(
+        df,
+        verbose=False,
+        semantic_context={"columns": {"brand": {"allowed_values": [*df["brand"].tolist()]}}},
+    )
+    assert declared["brand"].iloc[8] == "None"