@@ -186,6 +186,157 @@ def parse(self, raw: str) -> dict[str, object]:
         return data
 
 
+_HALLUCINATION_REQUIRED_KEYS: frozenset[str] = frozenset(
+    {"response_begin", "response_end", "response_text", "faithfulness", "explanation"}
+)
+_HALLUCINATION_FAITHFULNESS_LABELS: tuple[str, ...] = (
+    "faithful",
+    "partial",
+    "unfaithful",
+    "NA",
+)
+
+
+class _HallucinationDetectionContract(IOContract):
+    """Validate and deduplicate `hallucination_detection` adapter output.
+
+    Unlike other contracts in this module (which return parsed data unchanged),
+    `parse()` drops duplicate records that share `(response_begin, response_end)`
+    and agree on `faithfulness` (keeping the first record). Duplicate spans that
+    disagree on `faithfulness`, or distinct spans whose half-open intervals
+    `[response_begin, response_end)` overlap, raise `AdapterSchemaMismatchError`.
+
+    The `faithfulness` value set (`"faithful"`, `"partial"`, `"unfaithful"`,
+    `"NA"`) is closed on purpose (fail-closed): on Granite Switch, `io.yaml`
+    comes from the served model's `io_configs/` while this contract comes from
+    Mellea's registry, so a Switch release adding a new label raises
+    `AdapterSchemaMismatchError` until Mellea is upgraded.
+    """
+
+    def build_prompt(self, **_kwargs: object) -> Component:
+        raise NotImplementedError(_BUILD_PROMPT_NOT_IMPLEMENTED)
+
+    def parse(self, raw: str) -> dict[str, object]:
+        """Parse, validate, and deduplicate `hallucination_detection` output.
+
+        Args:
+            raw (str): Raw JSON string from the model.
+
+        Returns:
+            dict[str, object]: `{"items": <list of validated, deduplicated dicts>}`.
+
+        Raises:
+            ValueError: When `raw` is not valid JSON, is not a JSON array, or
+                contains a non-object element.
+            AdapterSchemaMismatchError: When any record is missing a required key,
+                has an invalid `faithfulness` value, has non-`int` (or `bool`) or
+                inverted span offsets (`response_begin > response_end`), has a
+                non-`str` `explanation`, conflicts on `faithfulness` for a repeated
+                `(response_begin, response_end)` span, or overlaps another span's
+                half-open interval `[response_begin, response_end)`.
+        """
+        data = json.loads(raw)
+        if not isinstance(data, list):
+            raise ValueError(
+                f"Adapter 'hallucination_detection' output must be a JSON array, "
+                f"got {type(data).__name__}."
+            )
+
+        seen_spans: dict[tuple[int, int], str] = {}
+        accepted_spans: list[tuple[int, int, frozenset[str]]] = []
+        deduplicated: list[dict[str, object]] = []
+
+        for item in data:
+            if not isinstance(item, dict):
+                raise ValueError(
+                    f"Adapter 'hallucination_detection' output array must contain "
+                    f"only JSON objects, got a {type(item).__name__} element."
+                )
+            observed = frozenset(item.keys())
+            missing = _HALLUCINATION_REQUIRED_KEYS - observed
+            if missing:
+                raise AdapterSchemaMismatchError(
+                    "hallucination_detection", observed, _HALLUCINATION_REQUIRED_KEYS
+                )
+
+            faithfulness = item["faithfulness"]
+            if (
+                not isinstance(faithfulness, str)
+                or faithfulness not in _HALLUCINATION_FAITHFULNESS_LABELS
+            ):
+                valid_labels = ", ".join(
+                    repr(lbl) for lbl in _HALLUCINATION_FAITHFULNESS_LABELS
+                )
+                raise AdapterSchemaMismatchError(
+                    "hallucination_detection",
+                    observed,
+                    _HALLUCINATION_REQUIRED_KEYS,
+                    f"invalid `faithfulness` value {faithfulness!r}; expected one of "
+                    f"{valid_labels} (upgrade `mellea` if the upstream adapter added "
+                    f"a new label)",
+                )
+
+            begin = item["response_begin"]
+            end = item["response_end"]
+            if (
+                isinstance(begin, bool)
+                or not isinstance(begin, int)
+                or isinstance(end, bool)
+                or not isinstance(end, int)
+            ):
+                raise AdapterSchemaMismatchError(
+                    "hallucination_detection",
+                    observed,
+                    _HALLUCINATION_REQUIRED_KEYS,
+                    "`response_begin` and `response_end` must be `int` (not `bool`)",
+                )
+            if begin > end:
+                raise AdapterSchemaMismatchError(
+                    "hallucination_detection",
+                    observed,
+                    _HALLUCINATION_REQUIRED_KEYS,
+                    f"`response_begin` ({begin}) must be <= `response_end` ({end})",
+                )
+
+            explanation = item["explanation"]
+            if not isinstance(explanation, str):
+                raise AdapterSchemaMismatchError(
+                    "hallucination_detection",
+                    observed,
+                    _HALLUCINATION_REQUIRED_KEYS,
+                    f"`explanation` must be a `str`, got {type(explanation).__name__}",
+                )
+
+            span = (begin, end)
+            if span in seen_spans:
+                if seen_spans[span] != faithfulness:
+                    raise AdapterSchemaMismatchError(
+                        "hallucination_detection",
+                        observed,
+                        _HALLUCINATION_REQUIRED_KEYS,
+                        f"conflicting `faithfulness` values ({seen_spans[span]!r} vs "
+                        f"{faithfulness!r}) for duplicate span {span}",
+                    )
+                continue
+
+            seen_spans[span] = faithfulness
+            accepted_spans.append((begin, end, observed))
+            deduplicated.append(item)
+
+        for i, (a_begin, a_end, _) in enumerate(accepted_spans):
+            for b_begin, b_end, b_observed in accepted_spans[i + 1 :]:
+                if max(a_begin, b_begin) < min(a_end, b_end):
+                    raise AdapterSchemaMismatchError(
+                        "hallucination_detection",
+                        b_observed,
+                        _HALLUCINATION_REQUIRED_KEYS,
+                        f"overlapping response spans [{a_begin}, {a_end}) and "
+                        f"[{b_begin}, {b_end})",
+                    )
+
+        return {"items": deduplicated}
+
+
 _INTRINSIC_IO_CONTRACTS: dict[str, IOContract] = {
     "answerability": _DictContract("answerability", frozenset({"answerability"})),
     "query_rewrite": _DictContract("query_rewrite", frozenset({"rewritten_question"})),
@@ -206,18 +357,7 @@ def parse(self, raw: str) -> dict[str, object]:
             }
         ),
     ),
-    "hallucination_detection": _ListContract(
-        "hallucination_detection",
-        frozenset(
-            {
-                "response_begin",
-                "response_end",
-                "response_text",
-                "faithfulness",
-                "explanation",
-            }
-        ),
-    ),
+    "hallucination_detection": _HallucinationDetectionContract(),
     "policy-guardrails": _PolicyGuardrailsContract(),
     "guardian-core": _GuardianCheckContract(),
     "factuality-detection": _DictContract("factuality-detection", frozenset({"score"})),
 
@@ -715,11 +715,17 @@ def _transform(self, value: Any, path: tuple, prepare_output: dict) -> dict:
         document_ids = prepare_output.get("document_ids")
         message_indices = prepare_output.get("message_indices")
 
-        if not isinstance(value, int):
+        if isinstance(value, bool) or not isinstance(value, int):
             raise TypeError(
                 f"Expected integer sentence number at path {path}, but "
                 f"found non-integer value {value} of type {type(value)}"
             )
+        num_sentences = len(begins)
+        if value < 0 or value >= num_sentences:
+            raise ValueError(
+                f"Sentence number {value} at path {path} is out of range for "
+                f"{num_sentences} sentence(s)"
+            )
         sentence_num = value
 
         result = {}
 
@@ -142,7 +142,11 @@ def find_context_attributions(
 
     Raises:
         ValueError: When the model output is not valid JSON, is not a
-            JSON array, or contains a non-object element.
+            JSON array, contains a non-object element, or contains a
+            sentence index that is negative or out of range for the decoded
+            source sentences.
+        TypeError: When sentence decoding encounters a non-integer (or
+            `bool`) sentence index in the model output.
         AdapterSchemaMismatchError: When any record in the output is missing a
             required field.
         AloraActivationError: When the resolved adapter is an aLoRA whose
 
@@ -218,7 +218,11 @@ def find_citations(
 
     Raises:
         ValueError: When the model output is not valid JSON, is not a
-            JSON array, or contains a non-object element.
+            JSON array, contains a non-object element, or contains a
+            sentence index that is negative or out of range for the decoded
+            source sentences.
+        TypeError: When sentence decoding encounters a non-integer (or
+            `bool`) sentence index in the model output.
         AdapterSchemaMismatchError: When any record in the output is missing a
             required field.
         AloraActivationError: When the resolved adapter is an aLoRA whose
@@ -254,15 +258,30 @@ def flag_hallucinated_content(
     user question are faithful to the retrieved document snippets. Sentences that do
     not align with the retrieved snippets are flagged as potential hallucinations.
 
-    The `faithfulness` field in each record is a string label (e.g.
-    `"faithful"`, `"hallucinated"`); coercion to a boolean is the caller's
-    responsibility.
+    The `faithfulness` field in each record is one of the four string labels
+    defined by the adapter's `io.yaml` schema:
+    - `"faithful"` — the sentence is fully supported by the retrieved documents.
+    - `"partial"` — the sentence is only partially supported by the retrieved
+      documents.
+    - `"unfaithful"` — the sentence is not supported by (or contradicts) the
+      retrieved documents.
+    - `"NA"` — faithfulness does not apply (for example, the sentence makes no
+      factual claim that needs grounding in the documents).
+
+    Coercion to a boolean is the caller's responsibility.
 
     Output contract — each record must contain: `response_begin`,
     `response_end`, `response_text`, `faithfulness`, `explanation`.  A
-    record missing any of these keys raises
-    :class:`~mellea.backends.adapters.AdapterSchemaMismatchError`; extra optional
-    keys do not raise (forward-compatible).
+    record missing any of these keys raises `AdapterSchemaMismatchError`;
+    extra optional keys do not raise (forward-compatible).  In addition, the
+    contract validates that `faithfulness` is one of `"faithful"`, `"partial"`,
+    `"unfaithful"`, or `"NA"`; `response_begin` and `response_end` are `int`
+    (not `bool`) with `response_begin <= response_end`; and `explanation` is a
+    `str`.  Duplicate records for the same `(response_begin, response_end)`
+    span that agree on `faithfulness` are deduplicated (keeping the first
+    record), whereas duplicate spans with conflicting `faithfulness` values or
+    distinct spans whose half-open intervals `[response_begin, response_end)`
+    overlap raise `AdapterSchemaMismatchError`.
 
     Args:
         response: The assistant's response to the user's question in the last turn
@@ -284,9 +303,17 @@ def flag_hallucinated_content(
 
     Raises:
         ValueError: When the model output is not valid JSON, is not a
-            JSON array, or contains a non-object element.
+            JSON array, contains a non-object element, or contains a
+            sentence index that is negative or out of range for the decoded
+            response sentences.
+        TypeError: When sentence decoding encounters a non-integer (or
+            `bool`) sentence index in the model output.
         AdapterSchemaMismatchError: When any record in the output is missing a
-            required field.
+            required field, has an invalid `faithfulness` label, has non-`int`
+            (or `bool`) or inverted span offsets (`response_begin > response_end`),
+            has a non-`str` `explanation`, has conflicting `faithfulness`
+            labels for the same `(response_begin, response_end)` span, or has
+            overlapping `[response_begin, response_end)` spans.
         AloraActivationError: When the resolved adapter is an aLoRA whose
             declared activation sequence is absent from the assembled prompt.
     """