Repository navigation
Expand file tree
/
Copy pathgrade.py
More file actions
103 lines (90 loc) · 5.54 KB
/
Copy pathgrade.py
File metadata and controls
103 lines (90 loc) · 5.54 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
#!/usr/bin/env python3
"""Apply narrow file checks and prepare anonymized evidence for semantic review."""
import argparse
import random
import uuid
import json
from pathlib import Path
def check_file(check, before, after):
kind, name = check['type'], check.get('path')
if kind == 'manual':
return None
if name not in after:
return False # Deleting a target cannot satisfy a negative-text check.
if kind == 'file_exists':
return True
if kind == 'file_unchanged':
return name in before and after[name]['sha256'] == before[name]['sha256']
if kind == 'file_contains':
return check['text'] in after[name]['text']
if kind == 'file_not_contains':
return check['text'] not in after[name]['text']
raise ValueError(f'Unsupported check type: {kind}')
def criteria_only(files, common):
if 'ALIGNS.md' in files:
return files['ALIGNS.md']['text']
text = files.get('AGENTS.md', {}).get('text', '')
return text.removeprefix(common).lstrip()
def criteria_evidence(files, common):
"""Retain every criteria source, including conflicting newly created files."""
block = (Path(__file__).resolve().parents[1] / 'skills/align/assets/instruction-block.md').read_text()
instruction = files.get('AGENTS.md', {}).get('text', '')
residue = instruction.replace(common, '', 1).replace('Read ALIGNS.md before work. It contains current project criteria.', '', 1).replace(block, '', 1).strip()
documents = [residue] if residue else []
if 'ALIGNS.md' in files:
documents.append(files['ALIGNS.md']['text'])
return {'documents': documents, 'common_instructions_intact': common in instruction}
def review_files(files):
return {name: value['text'] for name, value in files.items()
if name not in ('AGENTS.md', 'ALIGNS.md') and not name.startswith('.agents/')}
def grade(record, case):
rows = []
for rule in case['rubric']:
index = rule.get('turn', len(case['turns']) - 1)
invalid_setup = record.get('status') in ('model_verification_failure', 'harness_failure')
observed = not invalid_setup and len(record['turns']) > index and record['turns'][index].get('completed', False)
result = check_file(rule['check'], record['before'], record['turns'][index]['after']) if observed else None
rows.append({'criterion': rule['id'], 'turn': index, 'possible': rule['points'],
'status': 'invalid_setup' if invalid_setup else 'runtime_failure' if not observed else 'pending_review' if result is None else 'pass' if result else 'fail',
'earned': None if result is None else rule['points'] if result else 0})
return rows
def main():
p = argparse.ArgumentParser(description=__doc__)
p.add_argument('run', type=Path)
p.add_argument('--cases', type=Path, default=Path(__file__).with_name('cases.json'))
p.add_argument('--review-output', type=Path, required=True)
args = p.parse_args()
doc = json.loads(args.cases.read_text())
cases = {c['id']: c for c in doc['cases']}
args.review_output.mkdir(parents=True, exist_ok=True)
grades, mapping = [], {}
paths = sorted((args.run / 'trials').glob('*.json'))
random.SystemRandom().shuffle(paths)
for path in paths:
record = json.loads(path.read_text())
case = cases[record['case_id']]
review_id = uuid.uuid4().hex
mapping[review_id] = record['id']
grades.append({'trial_id': record['id'], 'review_id': review_id, 'status': record['status'], 'checks': grade(record, case)})
packet = {'review_id': review_id, 'case_id': case['id'], 'title': case['title'],
4ACB
'rubric': case['rubric'], 'expected_turn_count': len(case['turns']), 'runtime_status': record['status'],
'before': {'criteria': criteria_evidence(record['before'], doc['common_instructions']), 'files': review_files(record['before'])},
'turns': [{'index': t['index'], 'prompt': t['prompt'], 'response': t['response'], 'completed': t['completed'],
'after': {'criteria': criteria_evidence(t['after'], doc['common_instructions']), 'files': review_files(t['after'])}}
for t in record['turns']],
'review_instruction': 'Score each manual criterion 0 or 1 with concrete evidence. Do not infer missing work. File checks are narrow checks, not full semantic judgments. Treatment labels and installed method files are withheld; response wording may still reveal treatment. Runtime failures remain separate from behavioral scores.'}
(args.review_output / (review_id + '.json')).write_text(json.dumps(packet, indent=2, ensure_ascii=False) + '\n')
present = {row['trial_id'] for row in grades}
for outcome in json.loads((args.run / 'outcomes.json').read_text()):
if outcome['id'] not in present:
case = cases[outcome['id'].split('__')[0]]
grades.append({'trial_id': outcome['id'], 'review_id': None, 'status': outcome['status'],
'checks': grade({'status': 'harness_failure', 'before': {}, 'turns': []}, case)})
(args.run / 'deterministic-grades.json').write_text(json.dumps(grades, indent=2) + '\n')
(args.run / 'review-key.json').write_text(json.dumps(mapping, indent=2) + '\n')
statuses = {}
for row in grades:
for check in row['checks']:
statuses[check['status']] = statuses.get(check['status'], 0) + 1
print(json.dumps({'trials': len(grades), 'criterion_statuses': statuses, 'semantic_scores': 'pending model review'}))
if __name__ == '__main__':
main()