Bugfixes in the glossary protection guard: allow changes toward protected glossary terms, but deny changes away from protected terms

This commit is contained in:
2026-04-22 11:04:35 -05:00
parent 9afb5c5a4e
commit 183f8ba31d
5 changed files with 124 additions and 7 deletions

View File

@@ -358,6 +358,90 @@ def test_grammar_stage_cannot_reverse_glossary_protected_term(tmp_path):
assert "protected glossary term" in diagnostics["skipped_corrections"][0]["reason"]
def test_grammar_stage_can_correct_toward_protected_term(tmp_path):
transcript = parse_source_transcript_json(
"""
[
{"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Pawpaw's just worn out."}
]
"""
)
glossary = parse_glossary_yaml(
"""
glossary:
- name: "Popov"
category: npc
summary: "Popov is an allied NPC."
"""
)
grammar_correction = CorrectionCandidate(
id=1,
original_text="Pawpaw's",
corrected_text="Popov's",
confidence=0.95,
)
fake_client = FakeLLMClient(
[
CorrectionSet(corrections=[]),
CorrectionSet(corrections=[grammar_correction]),
]
)
revised = process_transcript(
transcript,
glossary,
_config(tmp_path),
llm_client=fake_client,
)
assert revised[0].text == "Popov's just worn out."
assert list((tmp_path / "work").iterdir()) == []
def test_grammar_stage_cannot_change_away_from_protected_term(tmp_path):
transcript = parse_source_transcript_json(
"""
[
{"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Popov's just worn out."}
]
"""
)
glossary = parse_glossary_yaml(
"""
glossary:
- name: "Popov"
category: npc
summary: "Popov is an allied NPC."
"""
)
grammar_correction = CorrectionCandidate(
id=1,
original_text="Popov's",
corrected_text="Pawpaw's",
confidence=0.95,
)
fake_client = FakeLLMClient(
[
CorrectionSet(corrections=[]),
CorrectionSet(corrections=[grammar_correction]),
]
)
revised = process_transcript(
transcript,
glossary,
_config(tmp_path, grammar_max_llm_passes=1),
llm_client=fake_client,
)
assert revised[0].text == "Popov's just worn out."
run_dirs = list((tmp_path / "work").iterdir())
assert len(run_dirs) == 1
diagnostics = json.loads((run_dirs[0] / "skipped-corrections.json").read_text(encoding="utf-8"))
assert diagnostics["skipped_corrections"][0]["stage"] == "grammar"
assert "protected glossary term" in diagnostics["skipped_corrections"][0]["reason"]
def test_grammar_stage_retries_repeated_span_and_applies_unique_retry(tmp_path):
transcript = parse_source_transcript_json(
"""