Implemented the spoken_word LLM review module

tests/test_llm_validators.py

@@ -6,10 +6,11 @@ from audita.core.errors import AuditaLLMError
 from audita.core.schemas import parse_glossary_yaml, parse_transcript_json
 from audita.framework.models import CorrectionProposal, ModuleRunSpec
 from audita.validators.base import ValidationContext
-from audita.validators.llm import MeaningReversalValidator, SpokenFormPlausibilityValidator
+from audita.validators.llm import MeaningReversalValidator, SpokenFormPlausibilityValidator, SpokenWordValidator
 from audita.validators.prompts import (
     build_meaning_reversal_messages,
     build_spoken_form_plausibility_messages,
+    build_spoken_word_messages,
 )
 import audita.validators.llm as llm_module
 
@@ -204,6 +205,110 @@ def test_meaning_reversal_validator_rejects_reversal_and_approves_nonreversal(tm
     assert "The figure became invisible in the doorway." in prompt_text
 
 
+def test_spoken_word_validator_approves_cleanup_and_rejects_rewrite(tmp_path):
+    transcript = parse_transcript_json(
+        """
+        [
+          {"id": 1, "speaker": "A", "start": 0.0, "end": 1.0, "text": "I, uh, I think we should go."},
+          {"id": 2, "speaker": "A", "start": 1.0, "end": 2.0, "text": "We should maybe proceed carefully."}
+        ]
+        """
+    )
+    proposals = [
+        CorrectionProposal(
+            proposal_index=0,
+            module_instance="spoken_word",
+            module_key="spoken_word",
+            id=1,
+            original_text="I, uh, I think",
+            corrected_text="I think",
+            confidence=0.95,
+        ),
+        CorrectionProposal(
+            proposal_index=1,
+            module_instance="spoken_word",
+            module_key="spoken_word",
+            id=2,
+            original_text="maybe proceed carefully",
+            corrected_text="go now",
+            confidence=0.95,
+        ),
+    ]
+    client = FakeStructuredLLMClient(
+        [
+            {
+                "validations": [
+                    {
+                        "correction_index": 0,
+                        "approved": True,
+                        "confidence": 0.97,
+                        "reason": "Reasonable dysfluency cleanup that preserves meaning.",
+                    },
+                    {
+                        "correction_index": 1,
+                        "approved": False,
+                        "confidence": 0.99,
+                        "reason": "This changes the substance of the segment rather than cleaning a dysfluency.",
+                    },
+                ]
+            }
+        ]
+    )
+
+    result = SpokenWordValidator("spoken_word_review").validate(
+        _context(proposals=proposals, transcript=transcript, llm_client=client, tmp_path=tmp_path)
+    )
+
+    assert [(decision.proposal_index, decision.approved) for decision in result.decisions] == [
+        (0, True),
+        (1, False),
+    ]
+    prompt_text = client.calls[0]["messages"][1]["content"]
+    assert "I think we should go." in prompt_text
+    assert "go now" in prompt_text
+
+
+def test_spoken_word_validator_allows_punctuation_cleanup_tied_to_dysfluency(tmp_path):
+    transcript = parse_transcript_json(
+        """
+        [
+          {"id": 1, "speaker": "A", "start": 0.0, "end": 1.0, "text": "Well ... I think we should go."}
+        ]
+        """
+    )
+    proposals = [
+        CorrectionProposal(
+            proposal_index=0,
+            module_instance="spoken_word",
+            module_key="spoken_word",
+            id=1,
+            original_text="Well ... ",
+            corrected_text="",
+            confidence=0.95,
+        )
+    ]
+    client = FakeStructuredLLMClient(
+        [
+            {
+                "validations": [
+                    {
+                        "correction_index": 0,
+                        "approved": True,
+                        "confidence": 0.95,
+                        "reason": "Removes a hesitation artifact without changing substantive meaning.",
+                    }
+                ]
+            }
+        ]
+    )
+
+    result = SpokenWordValidator("spoken_word_review").validate(
+        _context(proposals=proposals, transcript=transcript, llm_client=client, tmp_path=tmp_path)
+    )
+
+    assert [(decision.proposal_index, decision.approved) for decision in result.decisions] == [(0, True)]
+
+
 @pytest.mark.parametrize(
     ("validator", "payload", "message_fragment"),
     [
@@ -246,6 +351,20 @@ def test_meaning_reversal_validator_rejects_reversal_and_approves_nonreversal(tm
             },
             "unknown correction_index",
         ),
+        (
+            SpokenWordValidator("spoken_word_review"),
+            {
+                "validations": [
+                    {
+                        "correction_index": 99,
+                        "approved": True,
+                        "confidence": 0.9,
+                        "reason": "unknown",
+                    }
+                ]
+            },
+            "unknown correction_index",
+        ),
     ],
 )
 def test_llm_validators_reject_bad_correction_indexes(tmp_path, validator, payload, message_fragment):
@@ -356,6 +475,27 @@ def test_meaning_reversal_prompt_emphasizes_antonyms_and_segment_context():
     assert "original_segment_text" in messages[1]["content"]
 
 
+def test_spoken_word_prompt_emphasizes_dysfluency_cleanup():
+    messages = build_spoken_word_messages(
+        [
+            {
+                "correction_index": 0,
+                "id": 1,
+                "original_segment_text": "I, uh, I think we should go.",
+                "corrected_segment_text": "I think we should go.",
+                "original_text": "I, uh, I think",
+                "corrected_text": "I think",
+            }
+        ]
+    )
+
+    combined = messages[0]["content"] + messages[1]["content"]
+    assert "dysfluencies" in combined
+    assert "punctuation" in combined
+    assert "substantive meaning" in combined
+    assert "original_segment_text" in messages[1]["content"]
+
+
 def test_llm_validators_use_shared_token_batching_helper(monkeypatch, tmp_path):
     transcript = parse_transcript_json(
         """