File size: 19,705 Bytes
fa82a9d | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 | #!/usr/bin/env python3
"""Audit Objects temporal confirmation without treating abstention as absence."""
from __future__ import annotations
import argparse
from collections import Counter, defaultdict
import hashlib
import itertools
import json
from pathlib import Path
import sys
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT / 'src'))
from vbench_prompts_compile import annotation_jobs as J, records as R
from vbench_prompts_compile.experiments import cluster_summary, text_key
from vbench_prompts_compile.objects_repair import video_scores
from score_matrix import score_one, paired_rows, native_entity_codec
from cache_matrix_transforms import evidence_id
from analyze_object_controls import matched_contrasts
from render_matrix_tables import table, ci_text
def sha(path):
return hashlib.sha256(Path(path).read_bytes()).hexdigest()
def indexed(rows, key):
result = {}
for row in rows:
identity = row[key]
if identity in result:
raise ValueError('duplicate ' + key + ': ' + str(identity))
result[identity] = row
return result
def regress(previous_path, current_path):
unchanged, before, after = Counter(), [], []
total = 0
with Path(previous_path).open() as old, Path(current_path).open() as new:
for a, b in itertools.zip_longest(old, new):
if a is None or b is None:
raise ValueError('changed matrix denominator')
a, b = json.loads(a), json.loads(b)
if (a['sample_id'], a['scheme']) != (b['sample_id'], b['scheme']):
raise ValueError('changed observation identity/order')
total += 1
if a['task'] != 'objects' or a['scheme'] == 'Origin':
if a != b:
raise ValueError('untargeted record changed: ' + a['sample_id'])
unchanged[a['task'] + '/' + a['scheme']] += 1
if a['task'] == 'objects':
before.append(a)
after.append(b)
return {'total_rows': total, 'untouched_rows_exact': dict(unchanged)}, before, after
def frame_truth(target, other):
"""Three-state conjunction of independently stored visibility judgments."""
if target not in {'visible', 'not_visible', 'uncertain'} or other not in {'visible', 'not_visible', 'uncertain'}:
raise ValueError('invalid visibility label')
if target == other == 'visible':
return 'both_visible'
if 'not_visible' in (target, other):
return 'at_least_one_not_visible'
return 'uncertain'
def legacy_result(entities, evidence):
frames = evidence.get('frame_labels', [])
states = ['missing' if i >= len(frames) or frames[i] is None else
'supported' if all(name in frames[i] for name in entities) else 'absent' for i in range(16)]
scores = [float(v == 'supported') for v in states]
return {'score': sum(scores)/16, 'coverage': 1-states.count('missing')/16,
'abstention': 0., 'frame_scores': scores,
'objects_resolution': {'frame_states': states}}
def summarize_metrics(rows, keys):
return {key: cluster_summary(rows, key) for key in keys}
def main(argv=None):
p = argparse.ArgumentParser(description=__doc__)
for name in ('matrix', 'cache', 'transforms', 'visibility', 'predictions', 'metadata',
'previous-paired', 'paired', 'score-report', 'out'):
p.add_argument('--' + name, required=True)
args = p.parse_args(argv)
regression, before, after = regress(args.previous_paired, args.paired)
matrix = indexed([r for r in J.read_jsonl(Path(args.matrix)) if r['task'] == 'objects'], 'sample_id')
caches = indexed(J.read_jsonl(Path(args.cache)), 'relative_path')
transforms = indexed(J.read_jsonl(Path(args.transforms)), 'evidence_id')
reviews = indexed(J.read_jsonl(Path(args.visibility)), 'evidence_id')
predictions = {text_key('objects', r['prompt']): r['target'] for r in
indexed(J.read_jsonl(Path(args.predictions)), 'request_id').values()}
codec = native_entity_codec(json.loads(Path(args.metadata).read_text()))
report = json.loads(Path(args.score_report).read_text())
if not report.get('execution_complete') or report.get('objects_backend') != 'repair-v2':
raise ValueError('complete Objects v2 execution required')
# This audit must be attached to exactly the inputs/code of the scored run.
for path, digest in report['input_sha256'].items():
if sha(path) != digest:
raise ValueError('score input changed: ' + path)
for path, digest in report['scorer_source_sha256'].items():
if sha(ROOT / path) != digest:
raise ValueError('scorer source changed: ' + path)
for key, review in reviews.items():
digest = hashlib.sha256(json.dumps(transforms.get(key), sort_keys=True).encode()).hexdigest()
if digest != review.get('transform_record_sha256'):
raise ValueError('visibility review does not match transformed evidence')
expected_ids = set(matrix)
old_by_scheme = {s: indexed([r for r in before if r['scheme'] == s], 'sample_id')
for s in ('Origin', 'Repair-rule', 'Repair-model')}
new_by_scheme = {s: indexed([r for r in after if r['scheme'] == s], 'sample_id') for s in old_by_scheme}
if any(set(rows) != expected_ids for rows in [*old_by_scheme.values(), *new_by_scheme.values()]):
raise ValueError('Objects matrix not fully paired')
legacy_rows = paired_rows(list(matrix.values()), {'objects': caches}, transforms,
predictions, None, codec, reviews, objects_backend='legacy-official')
if legacy_rows != before:
raise ValueError('full legacy Objects rows no longer reproduce previous run')
regression['legacy_full_rows_exact'] = len(legacy_rows)
del legacy_rows
checked = Counter()
for sample_id, row in matrix.items():
evidence = caches[row['relative_path']] if row['evidence_kind'] == 'original' else transforms[evidence_id(row)]
for scheme in ('Repair-rule', 'Repair-model'):
old, new = old_by_scheme[scheme][sample_id], new_by_scheme[scheme][sample_id]
replay = score_one(row, scheme, evidence, predictions, None, codec, objects_backend='legacy-official')
for field, dest in [('score', 'transformed_score'), ('coverage', 'transformed_coverage'),
('abstention', 'transformed_abstention'), ('frame_scores', 'transformed_frame_scores'),
('target', 'transformed_target'), ('backend_target', 'transformed_backend_target')]:
if replay.get(field) != old[dest]:
raise ValueError('legacy Objects replay mismatch: ' + sample_id)
if old['original_backend_target'] != new['original_backend_target'] or old['transformed_backend_target'] != new['transformed_backend_target']:
raise ValueError('object requirement changed')
if any(a > b for a, b in zip(new['transformed_frame_scores'], old['transformed_frame_scores'])):
raise ValueError('temporal confirmation invented a positive')
if old['original_coverage'] != new['original_coverage'] or old['transformed_coverage'] != new['transformed_coverage']:
raise ValueError('changed backend evidence coverage')
checked['legacy_replayed_rows'] += 1
if row['transform'] == 'identity' or row.get('level') == 0:
if new['original_score'] != new['transformed_score'] or new['original_objects_resolution'] != new['transformed_objects_resolution']:
raise ValueError('zero/identity is not exact')
checked['identity_or_zero_exact_rows'] += 1
rule, model = new_by_scheme['Repair-rule'][sample_id], new_by_scheme['Repair-model'][sample_id]
for field in ('original_score', 'transformed_score', 'original_abstention', 'transformed_abstention',
'original_backend_target', 'transformed_backend_target', 'original_objects_resolution', 'transformed_objects_resolution'):
if rule[field] != model[field]:
raise ValueError('rule/model differ; separate ablation runs required')
regression.update(checked)
variants = ('legacy', 'adjacent_labels', 'adjacent_boxes')
scored = {v: {} for v in variants}
for sample_id, row in matrix.items():
evidence = caches[row['relative_path']] if row['evidence_kind'] == 'original' else transforms[evidence_id(row)]
entities = new_by_scheme['Repair-model'][sample_id]['transformed_backend_target']['entities']
for variant in variants:
value = legacy_result(entities, evidence) if variant == 'legacy' else video_scores(
entities, evidence.get('frame_detections', []), confirmation=variant.removeprefix('adjacent_'))
scored[variant][sample_id] = value
actual = new_by_scheme['Repair-model'][sample_id]
if scored['adjacent_boxes'][sample_id]['objects_resolution'] != actual['transformed_objects_resolution']:
raise ValueError('live scorer audit replay mismatch: ' + sample_id)
originals = {r['relative_path']: r for r in matrix.values() if r['transform'] == 'identity'}
endpoints = [r for r in matrix.values() if r.get('occlusion_control') == 'target' and r.get('level') == 1]
if len(originals) != len(caches) or len(endpoints) != len(caches) or len(reviews) != len(caches):
raise ValueError('incomplete source or endpoint denominator')
impact, endpoint_stats, visibility_stats, residuals = [], [], [], []
for variant in variants:
base_values, retained_frames = [], []
for row in originals.values():
a, b = scored['legacy'][row['sample_id']], scored[variant][row['sample_id']]
base_values.append({'family': row['family'], 'score': b['score'], 'delta_from_legacy': b['score']-a['score'],
'abstention': b['abstention'], 'positive_video_retained': b['score'] > 0, 'legacy_positive_video': a['score'] > 0})
retained_frames.extend({'family': row['family'], 'retained': new > 0}
for old, new in zip(a['frame_scores'], b['frame_scores']) if old > 0)
impact.append({'variant': variant, **summarize_metrics(base_values, ('score', 'delta_from_legacy', 'abstention')),
'original_positive_frame_retention': cluster_summary(retained_frames, 'retained'),
'original_positive_video_retention': cluster_summary([r for r in base_values if r['legacy_positive_video']], 'positive_video_retained')})
invisible, full_plan, isolated = [], [], []
visibility_rows = defaultdict(list)
for row in endpoints:
key = evidence_id(row)
review = reviews[key]
a = scored[variant][originals[row['relative_path']]['sample_id']]
b = scored[variant][row['sample_id']]
states = b['objects_resolution']['frame_states']
old_endpoint = scored['legacy'][row['sample_id']]
item = {'family': row['family'], 'positive_video': b['score'] > 0,
'no_positive': b['coverage'] == 1 and b['score'] == 0,
'all_decisive_negative': all(s == 'absent' for s in states),
'positive_frame_fraction': b['score'], 'unconfirmed_frame_fraction': b['abstention'],
'positive_video_delta': int(b['score'] > 0)-int(old_endpoint['score'] > 0),
'positive_frame_fraction_delta': b['score']-old_endpoint['score']}
known = bool(review.get('endpoint_invisible_verified'))
full_plan.append({'family': row['family'], 'verified_no_positive': known and item['no_positive'],
'verified_all_decisive_negative': known and item['all_decisive_negative']})
if known:
invisible.append(item)
if variant == 'adjacent_boxes' and (b['score'] > 0 or b['abstention'] > 0):
residuals.append({'relative_path': row['relative_path'], 'family': row['family'],
'legacy': scored['legacy'][row['sample_id']]['score'], 'repair': b['score'],
'abstention': b['abstention'], 'frame_states': states})
old_base = scored['legacy'][originals[row['relative_path']]['sample_id']]
if review.get('removal_verified') and old_base['score'] > 0:
isolated.append({'family': row['family'], 'drop': a['score']-b['score'] >= .03,
'decisive_drop': a['score']-b['score'] >= .03 and a['abstention'] == b['abstention'] == 0,
'baseline_retained': a['score'] > 0})
labels = review.get('labels')
if labels:
for side, value in [('before', a), ('after', b)]:
truths = [frame_truth(x, y) for x, y in zip(labels[side+'_target'], labels[side+'_other'])]
if len(truths) != 16:
raise ValueError('visibility frame slots changed')
for truth, state in zip(truths, value['objects_resolution']['frame_states']):
visibility_rows[side, truth].append({'family': row['family'], 'positive': state == 'supported',
'decisive_negative': state == 'absent', 'unconfirmed': state == 'unconfirmed', 'missing': state == 'missing'})
endpoint_stats.append({'variant': variant, 'remaining_positive_videos': sum(r['positive_video'] for r in invisible),
'no_positive_videos': sum(r['no_positive'] for r in invisible),
'all_decisive_negative_videos': sum(r['all_decisive_negative'] for r in invisible),
'invisible': summarize_metrics(invisible, ('positive_video', 'no_positive', 'all_decisive_negative', 'positive_frame_fraction', 'unconfirmed_frame_fraction', 'positive_video_delta', 'positive_frame_fraction_delta')),
'full_plan': summarize_metrics(full_plan, ('verified_no_positive', 'verified_all_decisive_negative')),
'same_legacy_positive_isolated_cohort': summarize_metrics(isolated, ('drop', 'decisive_drop', 'baseline_retained'))})
visibility_stats.extend({'variant': variant, 'side': key[0], 'truth': key[1],
**summarize_metrics(rows, ('positive', 'decisive_negative', 'unconfirmed', 'missing'))}
for key, rows in sorted(visibility_rows.items()))
controls = [{'scheme': scheme, 'level': level,
**summarize_metrics(rows, ('original', 'target', 'background', 'target_minus_background'))}
for (scheme, level), rows in sorted(matched_contrasts(after).items())]
out = Path(args.out)
out.mkdir(parents=True, exist_ok=True)
R.write_jsonl(out/'residual-endpoints.jsonl', residuals)
summary = {'version': 'objects-repair-v2', 'regression': regression, 'original_impact': impact,
'endpoints': endpoint_stats, 'visibility': visibility_stats, 'matched_controls': controls,
'planned_videos': len(caches), 'planned_conditions': len(matrix),
'visibility_status': dict(Counter(r['status'] for r in reviews.values())),
'summary': [r for r in report['summary'] if r['task'] == 'objects'],
'input_sha256': {v: sha(v) for k, v in vars(args).items() if k != 'out'},
'source_sha256': {n: sha(ROOT/n) for n in ['scripts/evaluate_objects_repair.py', 'scripts/score_matrix.py',
'scripts/render_matrix_tables.py', 'scripts/analyze_object_controls.py',
'src/vbench_prompts_compile/objects_repair.py', 'src/vbench_prompts_compile/experiments.py']},
'protocol_sha256': sha(ROOT/'docs/plans/15-objects-temporal-repair.md'),
'notes': ['Post-hoc engineering repair, not an independent blind evaluation.',
'Same 0.5 IoU rule on originals and every control; no threshold sweep or new training/inference.',
'Visibility model consensus is used for evaluation only, never by the scorer; not human gold.',
'Unconfirmed predictions are abstentions scored zero, not verified absence; report both no-positive and decisive-negative endpoints.',
'Original positive frame retention measures agreement with the old detector, not recall; visibility-stratified rates are separate.',
'The same legacy-positive isolated-removal cohort is retained for comparing variants.',
'2000 source-family bootstrap draws, seed 20260919; frames are not independent observations.',
'Sustained detector hallucinations and brief/fast-moving true objects remain limitations.']}
rows = [{'Variant': r['variant'], 'Videos': r['score']['n'], 'Mean score (CI)': ci_text(r['score']),
'Delta from legacy (CI)': ci_text(r['delta_from_legacy']), 'Abstain (CI)': ci_text(r['abstention']),
'Old positive frames retained (CI)': ci_text(r['original_positive_frame_retention']),
'Old positive videos retained (CI)': ci_text(r['original_positive_video_retention'])} for r in impact]
table(out, 'original-impact', rows, list(rows[0]))
rows = [{'Variant': r['variant'], 'Invisible videos': r['invisible']['positive_video']['n'],
'Still positive': r['remaining_positive_videos'], 'No positive': r['no_positive_videos'],
'All decisive negative': r['all_decisive_negative_videos'],
'Positive video delta (CI)': ci_text(r['invisible']['positive_video_delta']),
'Positive frames (CI)': ci_text(r['invisible']['positive_frame_fraction']),
'Unconfirmed frames (CI)': ci_text(r['invisible']['unconfirmed_frame_fraction']),
'Same isolated cohort drop (CI)': ci_text(r['same_legacy_positive_isolated_cohort']['drop']),
'Same isolated cohort decisive drop (CI)': ci_text(r['same_legacy_positive_isolated_cohort']['decisive_drop'])}
for r in endpoint_stats]
table(out, 'endpoints', rows, list(rows[0]))
rows = [{'Variant': r['variant'], 'Side': r['side'], 'Model visibility': r['truth'],
'Frames': r['positive']['n'], 'Families': r['positive']['families'],
'Positive rate (CI)': ci_text(r['positive']), 'Decisive negative rate (CI)': ci_text(r['decisive_negative']),
'Unconfirmed rate (CI)': ci_text(r['unconfirmed'])} for r in visibility_stats]
table(out, 'visibility', rows, list(rows[0]))
rows = [{'Scheme': r['scheme'], 'Level': r['level'], 'Complete videos': r['target']['n'],
'Target (CI)': ci_text(r['target']), 'Background (CI)': ci_text(r['background']),
'Target minus background (CI)': ci_text(r['target_minus_background'])} for r in controls]
table(out, 'controls', rows, list(rows[0]))
rows = [{'Scheme': r['scheme'], 'Transform': r['transform'], 'N': r['eligible'],
'Original': r['original_score']['estimate'], 'Transformed': r['transformed_score']['estimate'],
'Delta (CI)': ci_text(r['paired_delta']), 'Coverage': r['transformed_coverage']['estimate'],
'Abstain': r['transformed_abstention']['estimate']} for r in summary['summary']]
table(out, 'scores', rows, list(rows[0]))
summary['output_sha256'] = {p.name: sha(p) for p in sorted(out.iterdir()) if p.suffix in {'.md', '.csv', '.tex', '.jsonl'}}
J.atomic_json(out/'report.json', summary)
print(json.dumps({'regression': regression, 'remaining_positive_endpoints':
{r['variant']: r['remaining_positive_videos'] for r in endpoint_stats}, 'out': str(out)}, indent=2))
return 0
if __name__ == '__main__':
raise SystemExit(main())
|