eric/audita
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 2 Wiki Activity

Implemented a validation step to confirm that grammatical changes do not change substantive meaning

Browse Source
This commit is contained in:
Eric Rakestraw
2026-04-22 15:58:05 -05:00
parent 87c47517a0
commit d7788da1f4
15 changed files with 849 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

239
tests/test_pipeline.py
View File

@@ -1,23 +1,49 @@
import json
import pytest
from audita.config import AuditaConfig
from audita.errors import AuditaError
from audita.pipeline import process_transcript
from audita.schemas import CorrectionCandidate, CorrectionSet, parse_glossary_yaml, parse_source_transcript_json
from audita.schemas import (
CorrectionCandidate,
CorrectionSet,
GrammarValidationDecision,
GrammarValidationSet,
parse_glossary_yaml,
parse_source_transcript_json,
)
class FakeLLMClient:
def __init__(self, responses):
def __init__(self, responses, validation_responses=None):
self.responses = list(responses)
self.validation_responses = list(validation_responses or [])
self.calls = 0
self.validation_calls = 0
self.messages = []
self.validation_messages = []
def create_corrections(self, messages, config):
self.calls += 1
self.messages.append(messages)
return self.responses.pop(0)
def create_grammar_validations(self, messages, config):
self.validation_calls += 1
self.validation_messages.append(messages)
if not self.validation_responses:
raise AssertionError("Unexpected grammar validation request.")
return self.validation_responses.pop(0)
def _config(tmp_path, glossary_max_llm_passes=3, grammar_max_llm_passes=3):
def _config(
tmp_path,
glossary_max_llm_passes=3,
grammar_max_llm_passes=3,
grammar_validation_enabled=False,
grammar_validation_confidence_threshold=0.8,
):
return AuditaConfig(
api_key="key",
max_section_tokens=16000,
@@ -26,6 +52,8 @@ def _config(tmp_path, glossary_max_llm_passes=3, grammar_max_llm_passes=3):
max_retries=3,
glossary_max_llm_passes=glossary_max_llm_passes,
grammar_max_llm_passes=grammar_max_llm_passes,
grammar_validation_enabled=grammar_validation_enabled,
grammar_validation_confidence_threshold=grammar_validation_confidence_threshold,
work_dir=tmp_path / "work",
)
@@ -541,6 +569,8 @@ def test_pipeline_writes_stage_metadata_for_unresolved_retries(tmp_path):
assert metadata["grammar_max_llm_passes"] == 3
assert metadata["glossary_confidence_threshold"] == 0.8
assert metadata["grammar_confidence_threshold"] == 0.8
assert metadata["grammar_validation_enabled"] is False
assert metadata["grammar_validation_confidence_threshold"] == 0.8
assert [item["stage"] for item in metadata["stages"]] == ["glossary", "grammar"]
assert [item["pass_number"] for item in metadata["stages"][0]["passes"]] == [1, 2]
assert metadata["stages"][0]["passes"][0]["retry_segment_count"] == 1
@@ -589,6 +619,209 @@ def test_grammar_stage_runs_after_glossary_and_sees_corrected_text(tmp_path):
assert revised[0].text == "I ask Chauntea."
def test_grammar_validation_rejects_semantic_change(tmp_path):
transcript = parse_source_transcript_json(
"""
[
{"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "He became visible."}
]
"""
)
grammar_correction = CorrectionCandidate(
id=1,
original_text="visible",
corrected_text="invisible",
confidence=0.95,
)
validation = GrammarValidationDecision(
correction_index=0,
is_meaning_preserving=False,
confidence=0.99,
reason="This reverses visible to invisible.",
)
fake_client = FakeLLMClient(
[
CorrectionSet(corrections=[]),
CorrectionSet(corrections=[grammar_correction]),
],
validation_responses=[GrammarValidationSet(validations=[validation])],
)
revised = process_transcript(
transcript,
_glossary(),
_config(tmp_path, grammar_validation_enabled=True),
llm_client=fake_client,
)
assert revised[0].text == "He became visible."
assert fake_client.validation_calls == 1
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"))
skipped = diagnostics["skipped_corrections"][0]
assert skipped["stage"] == "grammar"
assert skipped["reason"] == "grammar validation rejected semantic change"
assert skipped["validation_confidence"] == 0.99
assert skipped["validation_reason"] == "This reverses visible to invisible."
metadata = json.loads((run_dirs[0] / "metadata.json").read_text(encoding="utf-8"))
grammar_pass = metadata["stages"][1]["passes"][0]
assert grammar_pass["validation_candidate_count"] == 1
assert grammar_pass["validation_approved_count"] == 0
assert grammar_pass["validation_rejected_count"] == 1
assert grammar_pass["validation_bypassed_count"] == 0
def test_grammar_validation_accepts_meaning_preserving_fix(tmp_path):
transcript = parse_source_transcript_json(
"""
[
{"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "Keep in bind."}
]
"""
)
grammar_correction = CorrectionCandidate(
id=1,
original_text="bind",
corrected_text="mind",
confidence=0.95,
)
validation = GrammarValidationDecision(
correction_index=0,
is_meaning_preserving=True,
confidence=0.95,
reason="This fixes the phrase keep in mind.",
)
fake_client = FakeLLMClient(
[
CorrectionSet(corrections=[]),
CorrectionSet(corrections=[grammar_correction]),
],
validation_responses=[GrammarValidationSet(validations=[validation])],
)
revised = process_transcript(
transcript,
_glossary(),
_config(tmp_path, grammar_validation_enabled=True),
llm_client=fake_client,
)
assert revised[0].text == "Keep in mind."
assert fake_client.validation_calls == 1
assert list((tmp_path / "work").iterdir()) == []
def test_grammar_validation_bypasses_protected_vocabulary_correction(tmp_path):
transcript = parse_source_transcript_json(
"""
[
{"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "The gestures arrived."}
]
"""
)
glossary = parse_glossary_yaml(
"""
glossary:
- name: "Jesters"
category: faction
summary: "The Jesters are a faction."
"""
)
grammar_correction = CorrectionCandidate(
id=1,
original_text="gestures",
corrected_text="Jesters",
confidence=0.95,
)
fake_client = FakeLLMClient(
[
CorrectionSet(corrections=[]),
CorrectionSet(corrections=[grammar_correction]),
]
)
revised = process_transcript(
transcript,
glossary,
_config(tmp_path, grammar_validation_enabled=True),
llm_client=fake_client,
)
assert revised[0].text == "The Jesters arrived."
assert fake_client.validation_calls == 0
assert list((tmp_path / "work").iterdir()) == []
def test_disabled_grammar_validation_preserves_current_behavior(tmp_path):
transcript = parse_source_transcript_json(
"""
[
{"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "He became visible."}
]
"""
)
grammar_correction = CorrectionCandidate(
id=1,
original_text="visible",
corrected_text="invisible",
confidence=0.95,
)
fake_client = FakeLLMClient(
[
CorrectionSet(corrections=[]),
CorrectionSet(corrections=[grammar_correction]),
]
)
revised = process_transcript(
transcript,
_glossary(),
_config(tmp_path, grammar_validation_enabled=False),
llm_client=fake_client,
)
assert revised[0].text == "He became invisible."
assert fake_client.validation_calls == 0
assert list((tmp_path / "work").iterdir()) == []
def test_missing_grammar_validation_decision_fails_and_preserves_diagnostics(tmp_path):
transcript = parse_source_transcript_json(
"""
[
{"speaker": "Eric", "start": 0.0, "end": 1.0, "text": "He became visible."}
]
"""
)
grammar_correction = CorrectionCandidate(
id=1,
original_text="visible",
corrected_text="invisible",
confidence=0.95,
)
fake_client = FakeLLMClient(
[
CorrectionSet(corrections=[]),
CorrectionSet(corrections=[grammar_correction]),
],
validation_responses=[GrammarValidationSet(validations=[])],
)
with pytest.raises(AuditaError):
process_transcript(
transcript,
_glossary(),
_config(tmp_path, grammar_validation_enabled=True),
llm_client=fake_client,
)
run_dirs = list((tmp_path / "work").iterdir())
assert len(run_dirs) == 1
assert (run_dirs[0] / "grammar" / "pass-0001" / "validation-prompt-0000.json").exists()
assert (run_dirs[0] / "grammar" / "pass-0001" / "validation-response-0000.json").exists()
def test_grammar_stage_cannot_reverse_glossary_protected_term(tmp_path):
transcript = parse_source_transcript_json(
"""