Updated the LLM prompt to enforce acoustic/phoenetic similarity requirements for proposed corrections

src/audita/prompts.py

@@ -15,14 +15,22 @@ def build_glossary_correction_messages(section: TranscriptSection, glossary: Glo
     system = (
         "You are Audita, a careful transcript correction assistant. "
         "Identify only transcription errors that are strongly supported by the glossary. "
+        "A valid correction must be acoustically plausible: the original transcript text "
+        "should sound similar to the proposed correction when spoken aloud. "
         "Do not make generic grammar, spelling, capitalization, or style edits. "
+        "Do not substitute an unrelated glossary term just because it could fit the topic. "
         "Do not rewrite unchanged transcript segments. "
         "Preserve speaker names, timestamps, and meaning."
     )
     user = (
         "Review this transcript section and return only corrections that should be applied.\n\n"
         "Rules:\n"
-        "- Correct domain-specific names, aliases, jargon, deities, locations, NPCs, players, and similar terms when the glossary supports the correction.\n"
+        "- Correct domain-specific names, aliases, jargon, deities, locations, NPCs, players, and similar terms only when both the glossary and surrounding transcript context support the correction.\n"
+        "- The correction must be likely to fix a transcription error: the original words should be phonetically or acoustically similar to the corrected words in spoken English.\n"
+        "- Appropriate example: correcting \"gestures\" to \"Jesters\" can be valid if \"Jesters\" appears in the glossary and nearby context supports that inference.\n"
+        "- Inappropriate example: correcting \"Lyra\" to \"Jesters\" should be omitted because those words are not similar in spoken English, even if \"Jesters\" appears in the glossary.\n"
+        "- Do not replace one clear glossary term, character name, location, or ordinary word with a different glossary term unless it is a plausible mishearing.\n"
+        "- Assign high confidence only when the correction is supported by glossary evidence, local context, and spoken-word similarity; otherwise omit the correction.\n"
         "- Use the exact segment_index, speaker, start, end, and original_text from the input segment.\n"
         "- corrected_text must contain the full corrected text for that segment.\n"
         "- confidence must be between 0.0 and 1.0.\n"
@@ -31,4 +39,3 @@ def build_glossary_correction_messages(section: TranscriptSection, glossary: Glo
         f"Transcript section:\n{section_json}"
     )
     return [{"role": "system", "content": system}, {"role": "user", "content": user}]
-

tests/test_prompts.py

@@ -0,0 +1,34 @@
+from audita.chunking import chunk_transcript
+from audita.prompts import build_glossary_correction_messages
+from audita.schemas import parse_glossary_yaml, parse_transcript_json
+
+
+def test_prompt_requires_acoustically_plausible_transcription_errors():
+    transcript = parse_transcript_json(
+        """
+        [
+          {"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "The gestures are nearby."}
+        ]
+        """
+    )
+    glossary = parse_glossary_yaml(
+        """
+        glossary:
+          - name: "Jesters"
+            category: faction
+            summary: "The Jesters are a local faction."
+          - name: "Lyra"
+            category: npc
+            summary: "Lyra is a hostile NPC."
+        """
+    )
+    section = chunk_transcript(transcript, max_section_tokens=16000)[0]
+
+    messages = build_glossary_correction_messages(section, glossary)
+    prompt_text = "\n".join(message["content"] for message in messages)
+
+    assert "acoustically plausible" in prompt_text
+    assert "phonetically or acoustically similar" in prompt_text
+    assert '"gestures" to "Jesters"' in prompt_text
+    assert '"Lyra" to "Jesters"' in prompt_text
+    assert "should be omitted" in prompt_text