File size: 74,675 Bytes
1ef1de7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 | """Portfolio-level claim review across multiple benchmark experiments."""
from __future__ import annotations
import json
import math
from dataclasses import dataclass, field
from fnmatch import fnmatch
from pathlib import Path
from typing import Any
from .errors import ClaimReviewError
from .manifest import ManifestEntry, load_manifest, sha256_file
from .models import ModelCard, audit_model_card, component_baseline_names, inspect_routed_calibration_collapse, model_admission_status, trained_recognizer_provenance_status
from .validation import gate_has_quantitative_evidence, sample_key
@dataclass(slots=True)
class RequiredValidationGate:
baseline: str
metric: str
slices: list[str] = field(default_factory=list)
@classmethod
def from_dict(cls, data: dict[str, Any]) -> "RequiredValidationGate":
baseline = str(data.get("baseline") or "")
metric = str(data.get("metric") or "")
if not baseline:
raise ClaimReviewError("required_validation_gates baseline is required")
if not metric:
raise ClaimReviewError("required_validation_gates metric is required")
return cls(baseline=baseline, metric=metric, slices=_string_list(data.get("slices")))
def to_dict(self) -> dict[str, Any]:
return {"baseline": self.baseline, "metric": self.metric, "slices": self.slices}
@dataclass(slots=True)
class RequiredExperiment:
name: str
datasets: list[str] = field(default_factory=list)
eval_manifest_patterns: list[str] = field(default_factory=list)
require_validation_status: str | None = None
required_validation_config: str | None = None
require_baselines: list[str] = field(default_factory=list)
required_metrics: list[str] = field(default_factory=list)
required_metric_slices: dict[str, list[str]] = field(default_factory=dict)
required_output_artifacts: list[str] = field(default_factory=list)
required_output_artifact_slices: dict[str, list[str]] = field(default_factory=dict)
required_validation_gates: list[RequiredValidationGate] = field(default_factory=list)
require_no_errors: bool = True
required_slices: list[str] = field(default_factory=list)
require_candidate_fallback_metrics: bool = False
max_candidate_fallback_trigger_rate: float | None = None
max_candidate_fallback_failed_rate: float | None = None
@classmethod
def from_dict(cls, data: dict[str, Any]) -> "RequiredExperiment":
name = str(data.get("name") or "")
if not name:
raise ClaimReviewError("required_experiment.name is required")
datasets = _string_list(data.get("datasets"))
eval_manifest_patterns = _string_list(data.get("eval_manifest_patterns"))
if not datasets and not eval_manifest_patterns:
raise ClaimReviewError(f"required experiment {name!r} must define datasets or eval_manifest_patterns")
require_validation_status = data.get("require_validation_status")
if require_validation_status is not None:
require_validation_status = str(require_validation_status)
required_validation_config = data.get("required_validation_config")
if required_validation_config is not None:
required_validation_config = str(required_validation_config)
require_baselines = _string_list(data.get("require_baselines"))
required_metrics = _string_list(data.get("required_metrics"))
required_metric_slices = _metric_slice_requirements(data.get("required_metric_slices"))
required_output_artifacts = _string_list(data.get("required_output_artifacts"))
required_output_artifact_slices = _metric_slice_requirements(data.get("required_output_artifact_slices"))
required_validation_gates = _required_validation_gates(data.get("required_validation_gates"))
_validate_required_validation_gates(
name,
require_validation_status,
require_baselines,
required_metrics,
required_validation_gates,
)
required_slices = _string_list(data.get("required_slices"))
require_candidate_fallback_metrics = bool(data.get("require_candidate_fallback_metrics", False))
max_candidate_fallback_trigger_rate = _optional_rate(data.get("max_candidate_fallback_trigger_rate"), "max_candidate_fallback_trigger_rate")
max_candidate_fallback_failed_rate = _optional_rate(data.get("max_candidate_fallback_failed_rate"), "max_candidate_fallback_failed_rate")
return cls(
name=name,
datasets=datasets,
eval_manifest_patterns=eval_manifest_patterns,
require_validation_status=require_validation_status,
required_validation_config=required_validation_config,
require_baselines=require_baselines,
required_metrics=required_metrics,
required_metric_slices=required_metric_slices,
required_output_artifacts=required_output_artifacts,
required_output_artifact_slices=required_output_artifact_slices,
required_validation_gates=required_validation_gates,
require_no_errors=bool(data.get("require_no_errors", True)),
required_slices=required_slices,
require_candidate_fallback_metrics=require_candidate_fallback_metrics,
max_candidate_fallback_trigger_rate=max_candidate_fallback_trigger_rate,
max_candidate_fallback_failed_rate=max_candidate_fallback_failed_rate,
)
def to_dict(self) -> dict[str, Any]:
return {
"name": self.name,
"datasets": self.datasets,
"eval_manifest_patterns": self.eval_manifest_patterns,
"require_validation_status": self.require_validation_status,
"required_validation_config": self.required_validation_config,
"require_baselines": self.require_baselines,
"required_metrics": self.required_metrics,
"required_metric_slices": self.required_metric_slices,
"required_output_artifacts": self.required_output_artifacts,
"required_output_artifact_slices": self.required_output_artifact_slices,
"required_validation_gates": [gate.to_dict() for gate in self.required_validation_gates],
"require_no_errors": self.require_no_errors,
"required_slices": self.required_slices,
"require_candidate_fallback_metrics": self.require_candidate_fallback_metrics,
"max_candidate_fallback_trigger_rate": self.max_candidate_fallback_trigger_rate,
"max_candidate_fallback_failed_rate": self.max_candidate_fallback_failed_rate,
}
@dataclass(slots=True)
class ClaimReviewConfig:
required_experiments: list[RequiredExperiment]
require_shared_candidate_model: bool = True
require_fallback_model_present: bool = False
require_candidate_fallback_provenance: bool = False
require_shared_fallback_model: bool = True
require_model_audit_passed: bool = True
require_fallback_model_audit_passed: bool = True
require_trained_recognizer: bool = False
require_model_admission: bool = False
@classmethod
def from_path(cls, path: str | Path | None) -> "ClaimReviewConfig":
if path is None:
return default_claim_review_config()
try:
payload = json.loads(Path(path).read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
raise ClaimReviewError(f"invalid claim review config JSON {path}: {exc}") from exc
if not isinstance(payload, dict):
raise ClaimReviewError("claim review config must be a JSON object")
required = payload.get("required_experiments", [])
if not isinstance(required, list):
raise ClaimReviewError("required_experiments must be a list")
experiments = [RequiredExperiment.from_dict(item) for item in required]
if not experiments:
raise ClaimReviewError("claim review config must define at least one required experiment")
return cls(
required_experiments=experiments,
require_shared_candidate_model=bool(payload.get("require_shared_candidate_model", True)),
require_fallback_model_present=bool(payload.get("require_fallback_model_present", False)),
require_candidate_fallback_provenance=bool(payload.get("require_candidate_fallback_provenance", False)),
require_shared_fallback_model=bool(payload.get("require_shared_fallback_model", True)),
require_model_audit_passed=bool(payload.get("require_model_audit_passed", True)),
require_fallback_model_audit_passed=bool(payload.get("require_fallback_model_audit_passed", True)),
require_trained_recognizer=bool(payload.get("require_trained_recognizer", False)),
require_model_admission=bool(payload.get("require_model_admission", False)),
)
def to_dict(self) -> dict[str, Any]:
return {
"required_experiments": [item.to_dict() for item in self.required_experiments],
"require_shared_candidate_model": self.require_shared_candidate_model,
"require_fallback_model_present": self.require_fallback_model_present,
"require_candidate_fallback_provenance": self.require_candidate_fallback_provenance,
"require_shared_fallback_model": self.require_shared_fallback_model,
"require_model_audit_passed": self.require_model_audit_passed,
"require_fallback_model_audit_passed": self.require_fallback_model_audit_passed,
"require_trained_recognizer": self.require_trained_recognizer,
"require_model_admission": self.require_model_admission,
}
@dataclass(slots=True)
class ClaimReviewCheck:
name: str
status: str
summary: str
details: list[str] = field(default_factory=list)
def to_dict(self) -> dict[str, Any]:
return {
"name": self.name,
"status": self.status,
"summary": self.summary,
"details": self.details,
}
@dataclass(slots=True)
class ExperimentEvidence:
experiment_path: str
experiment_id: str
eval_manifest: str
eval_manifest_sha256: str | None
recorded_eval_manifest_sha256: str | None
eval_entries: list[ManifestEntry]
datasets: list[str]
slice_counts: dict[str, int]
baselines: list[str]
benchmark_report: str
benchmark_rows: list[dict[str, Any]]
benchmark_status_counts: dict[str, int]
validation_status: str | None
validation_report: str | None
validation_config: str | None
validation_config_sha256: str | None
train_manifests: list[dict[str, str | None]]
candidate_model_config: str | None
candidate_model_config_sha256: str | None
fallback_model_config: str | None
fallback_model_config_sha256: str | None
def to_dict(self) -> dict[str, Any]:
return {
"experiment_path": self.experiment_path,
"experiment_id": self.experiment_id,
"eval_manifest": self.eval_manifest,
"eval_manifest_sha256": self.eval_manifest_sha256,
"recorded_eval_manifest_sha256": self.recorded_eval_manifest_sha256,
"datasets": self.datasets,
"slice_counts": self.slice_counts,
"baselines": self.baselines,
"benchmark_report": self.benchmark_report,
"benchmark_status_counts": self.benchmark_status_counts,
"validation_status": self.validation_status,
"validation_report": self.validation_report,
"validation_config": self.validation_config,
"validation_config_sha256": self.validation_config_sha256,
"train_manifests": self.train_manifests,
"candidate_model_config": self.candidate_model_config,
"candidate_model_config_sha256": self.candidate_model_config_sha256,
"fallback_model_config": self.fallback_model_config,
"fallback_model_config_sha256": self.fallback_model_config_sha256,
}
@dataclass(slots=True)
class ClaimReviewReport:
passed: bool
config: dict[str, Any]
model_config: str | None
experiments: list[ExperimentEvidence]
checks: list[ClaimReviewCheck]
def to_dict(self) -> dict[str, Any]:
return {
"passed": self.passed,
"config": self.config,
"model_config": self.model_config,
"experiments": [item.to_dict() for item in self.experiments],
"checks": [item.to_dict() for item in self.checks],
}
def review_claim(
experiment_reports: list[str | Path],
output_dir: str | Path,
*,
config_path: str | Path | None = None,
model_config: str | Path | None = None,
) -> ClaimReviewReport:
if not experiment_reports:
raise ClaimReviewError("claim review requires at least one experiment report")
config = ClaimReviewConfig.from_path(config_path)
evidence = [_load_experiment_evidence(path) for path in experiment_reports]
checks: list[ClaimReviewCheck] = []
checks.append(_shared_candidate_model_check(evidence, model_config, config))
checks.append(_fallback_model_check(evidence, config))
checks.append(_candidate_fallback_provenance_check(evidence, model_config, config))
checks.append(_trained_recognizer_check(evidence, model_config, config))
checks.append(_model_admission_review_check(evidence, model_config, config))
checks.append(_candidate_calibration_review_check(evidence, model_config))
checks.append(_candidate_baseline_equivalence_check(evidence, model_config))
checks.extend(_required_experiment_checks(evidence, config))
report = ClaimReviewReport(
passed=all(check.status != "fail" for check in checks),
config=config.to_dict(),
model_config=str(model_config) if model_config else None,
experiments=evidence,
checks=checks,
)
out = Path(output_dir)
out.mkdir(parents=True, exist_ok=True)
_write_claim_review(report, out)
return report
def default_claim_review_config() -> ClaimReviewConfig:
return ClaimReviewConfig(
required_experiments=[
RequiredExperiment(
name="internal-nepali-pack",
datasets=["ocrtech-eval-pack"],
require_validation_status="validated",
require_baselines=["candidate", "tesseract", "stock-paddle", "glm-ocr", "paddleocr-vl"],
required_metrics=[
"cer",
"wer",
"table_cell_f1",
"reading_order_pair_accuracy",
"figure_detection_f1",
"figure_caption_cer",
],
required_slices=["nepali", "english", "table", "reading_order", "figure", "scan"],
)
]
)
def _load_experiment_evidence(path: str | Path) -> ExperimentEvidence:
experiment_path = Path(path)
_require_existing_file(experiment_path, "experiment report")
try:
payload = json.loads(experiment_path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
raise ClaimReviewError(f"invalid experiment JSON {experiment_path}: {exc}") from exc
if not isinstance(payload, dict):
raise ClaimReviewError(f"experiment report must be a JSON object: {experiment_path}")
eval_manifest = str(payload.get("eval_manifest") or "")
if not eval_manifest:
raise ClaimReviewError(f"experiment report missing eval_manifest: {experiment_path}")
_require_existing_file(Path(eval_manifest), f"eval_manifest in {experiment_path}")
eval_entries = load_manifest(eval_manifest)
eval_manifest_sha256 = sha256_file(Path(eval_manifest)) if Path(eval_manifest).exists() else None
datasets = sorted({entry.dataset for entry in eval_entries})
slice_counts = _slice_counts(eval_entries)
benchmark_report = str(payload.get("benchmark_report") or "")
if not benchmark_report:
raise ClaimReviewError(f"experiment report missing benchmark_report: {experiment_path}")
_require_existing_file(Path(benchmark_report), f"benchmark_report in {experiment_path}")
benchmark_rows = _load_benchmark_report(benchmark_report)
baselines = sorted({str(item.get("baseline") or "") for item in benchmark_rows if str(item.get("baseline") or "")})
benchmark_status_counts = _status_counts(benchmark_rows)
environment = dict(payload.get("environment") or {})
return ExperimentEvidence(
experiment_path=str(experiment_path),
experiment_id=str(payload.get("experiment_id") or experiment_path.stem),
eval_manifest=eval_manifest,
eval_manifest_sha256=eval_manifest_sha256,
recorded_eval_manifest_sha256=str(environment["eval_manifest_sha256"]) if environment.get("eval_manifest_sha256") else None,
eval_entries=eval_entries,
datasets=datasets,
slice_counts=slice_counts,
baselines=baselines,
benchmark_report=benchmark_report,
benchmark_rows=benchmark_rows,
benchmark_status_counts={str(key): int(value) for key, value in benchmark_status_counts.items()},
validation_status=str(payload["validation_status"]) if payload.get("validation_status") is not None else None,
validation_report=str(payload["validation_report"]) if payload.get("validation_report") is not None else None,
validation_config=str(environment["validation_config"]) if environment.get("validation_config") else None,
validation_config_sha256=str(environment["validation_config_sha256"]) if environment.get("validation_config_sha256") else None,
train_manifests=_environment_manifest_hashes(environment.get("train_manifests")),
candidate_model_config=str(environment["candidate_model_config"]) if environment.get("candidate_model_config") else None,
candidate_model_config_sha256=str(environment["candidate_model_config_sha256"]) if environment.get("candidate_model_config_sha256") else None,
fallback_model_config=str(environment["fallback_model_config"]) if environment.get("fallback_model_config") else None,
fallback_model_config_sha256=str(environment["fallback_model_config_sha256"]) if environment.get("fallback_model_config_sha256") else None,
)
def _shared_candidate_model_check(
experiments: list[ExperimentEvidence],
model_config: str | Path | None,
config: ClaimReviewConfig,
) -> ClaimReviewCheck:
candidate_refs = [item for item in experiments if any(baseline in {"candidate", "ours", "ocrtech"} for baseline in item.baselines)]
details: list[str] = []
shas = sorted({item.candidate_model_config_sha256 for item in candidate_refs if item.candidate_model_config_sha256})
missing = [item.experiment_id for item in candidate_refs if not item.candidate_model_config_sha256]
if missing:
return ClaimReviewCheck(
"candidate-model",
"fail",
"candidate experiments are missing candidate_model_config_sha256",
details=missing,
)
if config.require_shared_candidate_model and len(shas) > 1:
return ClaimReviewCheck("candidate-model", "fail", "candidate experiments do not share one model config", details=shas)
if model_config is not None:
model_path = Path(model_config)
if not model_path.exists():
return ClaimReviewCheck("candidate-model", "fail", f"explicit model config does not exist: {model_path}")
explicit_sha = sha256_file(model_path)
if shas and explicit_sha not in shas:
return ClaimReviewCheck(
"candidate-model",
"fail",
"explicit model config does not match experiment evidence",
details=[f"explicit_sha256={explicit_sha}", *shas],
)
if config.require_model_audit_passed:
audit = audit_model_card(model_path)
details.extend(audit.warnings)
if not audit.passed:
return ClaimReviewCheck("candidate-model", "fail", "explicit model card failed audit", details=[*audit.issues, *details])
return ClaimReviewCheck("candidate-model", "pass", f"explicit model card audited: {audit.model_id}", details=details)
return ClaimReviewCheck("candidate-model", "pass", "explicit model config matches experiment evidence")
if candidate_refs and config.require_model_audit_passed:
model_paths = sorted({item.candidate_model_config for item in candidate_refs if item.candidate_model_config})
if len(model_paths) != 1:
return ClaimReviewCheck(
"candidate-model",
"fail",
"claim review requires one explicit candidate model path across candidate experiments",
details=model_paths,
)
model_path = Path(model_paths[0])
if not model_path.exists():
return ClaimReviewCheck("candidate-model", "fail", f"candidate model config does not exist: {model_path}")
actual_sha = sha256_file(model_path)
if actual_sha not in shas:
return ClaimReviewCheck(
"candidate-model",
"fail",
"candidate model config sha256 does not match experiment evidence",
details=[f"path={model_path}", f"actual_sha256={actual_sha}", *shas],
)
audit = audit_model_card(model_path)
details.extend(audit.warnings)
if not audit.passed:
return ClaimReviewCheck("candidate-model", "fail", "candidate model card failed audit", details=[*audit.issues, *details])
return ClaimReviewCheck(
"candidate-model",
"pass",
f"candidate model card hash-checked and audited: {audit.model_id}",
details=[f"path={model_path}", f"sha256={actual_sha}", *details],
)
if candidate_refs:
return ClaimReviewCheck("candidate-model", "pass", "candidate experiments share one model config")
return ClaimReviewCheck("candidate-model", "warn", "no candidate baselines found in supplied experiments")
def _fallback_model_check(experiments: list[ExperimentEvidence], config: ClaimReviewConfig) -> ClaimReviewCheck:
candidate_refs = [item for item in experiments if any(baseline in {"candidate", "ours", "ocrtech"} for baseline in item.baselines)]
missing_required = [
item.experiment_id
for item in candidate_refs
if not item.fallback_model_config and not item.fallback_model_config_sha256
]
if config.require_fallback_model_present and missing_required:
return ClaimReviewCheck(
"fallback-model",
"fail",
"candidate experiments are missing required fallback_model_config provenance",
details=missing_required,
)
fallback_refs = [item for item in candidate_refs if item.fallback_model_config or item.fallback_model_config_sha256]
if not fallback_refs:
return ClaimReviewCheck("fallback-model", "pass", "no fallback model config recorded in candidate experiments")
missing_path = [item.experiment_id for item in fallback_refs if not item.fallback_model_config]
if missing_path:
return ClaimReviewCheck("fallback-model", "fail", "fallback model sha256 is recorded but fallback_model_config path is missing", details=missing_path)
missing_sha = [item.experiment_id for item in fallback_refs if not item.fallback_model_config_sha256]
if missing_sha:
return ClaimReviewCheck("fallback-model", "fail", "fallback model config path is recorded but sha256 is missing", details=missing_sha)
shas = sorted({item.fallback_model_config_sha256 for item in fallback_refs if item.fallback_model_config_sha256})
if config.require_shared_fallback_model and len(shas) > 1:
return ClaimReviewCheck("fallback-model", "fail", "candidate experiments do not share one fallback model config", details=shas)
details: list[str] = []
model_paths = sorted({item.fallback_model_config for item in fallback_refs if item.fallback_model_config})
for model_path_text in model_paths:
model_path = Path(model_path_text)
matching_shas = sorted({item.fallback_model_config_sha256 for item in fallback_refs if item.fallback_model_config == model_path_text})
if not model_path.exists():
return ClaimReviewCheck("fallback-model", "fail", f"fallback model config does not exist: {model_path}")
actual_sha = sha256_file(model_path)
if actual_sha not in matching_shas:
return ClaimReviewCheck(
"fallback-model",
"fail",
"fallback model config sha256 does not match experiment evidence",
details=[f"path={model_path}", f"actual_sha256={actual_sha}", *matching_shas],
)
if config.require_fallback_model_audit_passed:
audit = audit_model_card(model_path)
details.extend(f"{model_path}: {warning}" for warning in audit.warnings)
if not audit.passed:
return ClaimReviewCheck(
"fallback-model",
"fail",
"fallback model card failed audit",
details=[f"{model_path}: {issue}" for issue in audit.issues] + details,
)
details.append(f"{model_path}: audited model_id={audit.model_id}")
return ClaimReviewCheck("fallback-model", "pass", "fallback model configs are present, hash-checked, and audited", details=details)
def _candidate_fallback_provenance_check(
experiments: list[ExperimentEvidence],
model_config: str | Path | None,
config: ClaimReviewConfig,
) -> ClaimReviewCheck:
if not config.require_candidate_fallback_provenance:
return ClaimReviewCheck("candidate-fallback-provenance", "pass", "candidate fallback provenance is not required by config")
model_path = _single_candidate_model_path(experiments, model_config)
if model_path is None:
return ClaimReviewCheck("candidate-fallback-provenance", "fail", "candidate fallback provenance requires one candidate model path")
if not model_path.exists():
return ClaimReviewCheck("candidate-fallback-provenance", "fail", f"candidate model config does not exist: {model_path}")
card = ModelCard.from_path(model_path)
payload = card.provenance.get("fallback_model_card")
if not isinstance(payload, dict):
return ClaimReviewCheck("candidate-fallback-provenance", "fail", "candidate model card missing provenance.fallback_model_card")
path_value = payload.get("path")
if not isinstance(path_value, str) or not path_value:
return ClaimReviewCheck("candidate-fallback-provenance", "fail", "candidate fallback_model_card.path is missing")
fallback_path = Path(path_value)
if not fallback_path.is_absolute():
fallback_path = model_path.parent / fallback_path
if not fallback_path.exists():
return ClaimReviewCheck("candidate-fallback-provenance", "fail", f"candidate fallback model card does not exist: {fallback_path}")
declared_sha = payload.get("sha256")
actual_sha = sha256_file(fallback_path)
if isinstance(declared_sha, str) and declared_sha and declared_sha != actual_sha:
return ClaimReviewCheck(
"candidate-fallback-provenance",
"fail",
"candidate fallback model card sha256 mismatch",
details=[f"expected={declared_sha}", f"actual={actual_sha}", f"path={fallback_path}"],
)
fallback_refs = [item for item in experiments if item.fallback_model_config_sha256]
mismatches = [
f"{item.experiment_id}: experiment_fallback_sha256={item.fallback_model_config_sha256} candidate_fallback_sha256={actual_sha}"
for item in fallback_refs
if item.fallback_model_config_sha256 != actual_sha
]
if mismatches:
return ClaimReviewCheck(
"candidate-fallback-provenance",
"fail",
"candidate fallback model card does not match experiment fallback_model_config",
details=mismatches,
)
audit = audit_model_card(fallback_path)
if not audit.passed:
return ClaimReviewCheck(
"candidate-fallback-provenance",
"fail",
"candidate fallback model card failed audit",
details=[*audit.issues, *audit.warnings],
)
return ClaimReviewCheck(
"candidate-fallback-provenance",
"pass",
"candidate fallback model provenance matches reviewed experiment fallback configs",
details=[f"fallback_model_card={fallback_path}", f"sha256={actual_sha}", *audit.warnings],
)
def _trained_recognizer_check(
experiments: list[ExperimentEvidence],
model_config: str | Path | None,
config: ClaimReviewConfig,
) -> ClaimReviewCheck:
if not config.require_trained_recognizer:
return ClaimReviewCheck("trained-recognizer", "pass", "trained recognizer provenance is not required by config")
model_path = _single_candidate_model_path(experiments, model_config)
if model_path is None:
return ClaimReviewCheck("trained-recognizer", "fail", "trained recognizer review requires one candidate model path")
if not model_path.exists():
return ClaimReviewCheck("trained-recognizer", "fail", f"candidate model config does not exist: {model_path}")
card = ModelCard.from_path(model_path)
status = trained_recognizer_provenance_status(card, model_path)
if not status.passed:
return ClaimReviewCheck(
"trained-recognizer",
"fail",
"candidate model lacks trained local recognizer provenance",
details=status.details,
)
return ClaimReviewCheck("trained-recognizer", "pass", "candidate includes trained local recognizer provenance", details=status.details)
def _model_admission_review_check(
experiments: list[ExperimentEvidence],
model_config: str | Path | None,
config: ClaimReviewConfig,
) -> ClaimReviewCheck:
if not config.require_model_admission:
return ClaimReviewCheck("model-admission", "pass", "model admission validation is not required by config")
model_path = _single_candidate_model_path(experiments, model_config)
if model_path is None:
return ClaimReviewCheck("model-admission", "fail", "model admission review requires one candidate model path")
if not model_path.exists():
return ClaimReviewCheck("model-admission", "fail", f"candidate model config does not exist: {model_path}")
card = ModelCard.from_path(model_path)
status = model_admission_status(card, model_path)
if not status.passed:
return ClaimReviewCheck("model-admission", "fail", "candidate model has not passed admission validation", details=status.details)
return ClaimReviewCheck("model-admission", "pass", "candidate model passed admission validation", details=status.details)
def _single_candidate_model_path(experiments: list[ExperimentEvidence], model_config: str | Path | None) -> Path | None:
if model_config is not None:
return Path(model_config)
candidate_refs = [item for item in experiments if any(baseline in {"candidate", "ours", "ocrtech"} for baseline in item.baselines)]
distinct_paths = sorted({item.candidate_model_config for item in candidate_refs if item.candidate_model_config})
if len(distinct_paths) != 1:
return None
return Path(distinct_paths[0])
def _required_experiment_checks(experiments: list[ExperimentEvidence], config: ClaimReviewConfig) -> list[ClaimReviewCheck]:
checks: list[ClaimReviewCheck] = []
for required in config.required_experiments:
matches = [item for item in experiments if _matches_required_experiment(item, required)]
if not matches:
checks.append(
ClaimReviewCheck(
f"required:{required.name}",
"fail",
"no experiment matches the required dataset/pattern constraints",
details=[*required.datasets, *required.eval_manifest_patterns],
)
)
continue
if len(matches) > 1:
checks.append(
ClaimReviewCheck(
f"required:{required.name}",
"fail",
"multiple experiments match one required slot; tighten the claim config",
details=[item.experiment_path for item in matches],
)
)
continue
evidence = matches[0]
checks.append(_evaluate_required_experiment(evidence, required))
return checks
def _candidate_calibration_review_check(
experiments: list[ExperimentEvidence],
model_config: str | Path | None,
) -> ClaimReviewCheck:
candidate_refs = [item for item in experiments if any(baseline in {"candidate", "ours", "ocrtech"} for baseline in item.baselines)]
if not candidate_refs:
return ClaimReviewCheck("candidate-calibration", "warn", "no candidate experiments found; calibration review skipped")
model_path = Path(model_config) if model_config is not None else None
if model_path is None:
distinct_paths = sorted({item.candidate_model_config for item in candidate_refs if item.candidate_model_config})
if len(distinct_paths) != 1:
return ClaimReviewCheck("candidate-calibration", "warn", "candidate calibration review skipped because no single model path is available")
model_path = Path(distinct_paths[0])
if not model_path.exists():
return ClaimReviewCheck("candidate-calibration", "fail", f"candidate model config does not exist: {model_path}")
card = ModelCard.from_path(model_path)
if card.backend not in {"quality_select_composite", "quality_ranked_ensemble", "script_select_composite"}:
return ClaimReviewCheck("candidate-calibration", "pass", "candidate backend does not require routed-threshold calibration review")
calibration = card.provenance.get("calibration")
if not isinstance(calibration, dict):
return ClaimReviewCheck("candidate-calibration", "fail", f"{card.backend} candidate is missing provenance.calibration")
eval_payload = calibration.get("eval_manifest")
if not isinstance(eval_payload, dict):
return ClaimReviewCheck("candidate-calibration", "fail", f"{card.backend} calibration is missing eval_manifest provenance")
calibration_sha = eval_payload.get("sha256")
if not isinstance(calibration_sha, str) or not calibration_sha:
return ClaimReviewCheck("candidate-calibration", "fail", f"{card.backend} calibration eval manifest is missing sha256")
overlaps = [item.experiment_id for item in candidate_refs if item.eval_manifest_sha256 and item.eval_manifest_sha256 == calibration_sha]
if overlaps:
return ClaimReviewCheck(
"candidate-calibration",
"fail",
f"{card.backend} calibration eval manifest matches a reviewed experiment eval manifest",
details=overlaps,
)
threshold = calibration.get("selected_threshold")
details = []
if isinstance(threshold, int | float):
details.append(f"selected_threshold={float(threshold):.6f}")
collapse_reason, collapse_details = inspect_routed_calibration_collapse(card, model_path)
if collapse_reason is not None:
return ClaimReviewCheck("candidate-calibration", "fail", collapse_reason, details=[*details, *collapse_details])
return ClaimReviewCheck("candidate-calibration", "pass", f"{card.backend} calibration provenance is separate from reviewed eval manifests", details=details)
def _candidate_baseline_equivalence_check(
experiments: list[ExperimentEvidence],
model_config: str | Path | None,
) -> ClaimReviewCheck:
candidate_refs = [item for item in experiments if any(baseline in {"candidate", "ours", "ocrtech"} for baseline in item.baselines)]
if not candidate_refs:
return ClaimReviewCheck("candidate-distinctness", "warn", "no candidate experiments found; baseline-equivalence review skipped")
model_path = Path(model_config) if model_config is not None else None
if model_path is None:
distinct_paths = sorted({item.candidate_model_config for item in candidate_refs if item.candidate_model_config})
if len(distinct_paths) != 1:
return ClaimReviewCheck("candidate-distinctness", "warn", "candidate distinctness review skipped because no single model path is available")
model_path = Path(distinct_paths[0])
if not model_path.exists():
return ClaimReviewCheck("candidate-distinctness", "fail", f"candidate model config does not exist: {model_path}")
card = ModelCard.from_path(model_path)
component_baselines = _component_baselines_for_card(card)
if not component_baselines:
return ClaimReviewCheck("candidate-distinctness", "pass", "candidate backend does not require component-baseline distinctness review")
checked: list[str] = []
for evidence in candidate_refs:
candidate_baseline = _candidate_baseline_name(evidence.baselines)
if candidate_baseline is None:
continue
for component_baseline in component_baselines:
if component_baseline not in evidence.baselines:
continue
paired = _paired_metric_outcomes(
evidence.benchmark_rows,
candidate_baseline,
component_baseline,
metric="cer",
slice_name="all",
)
if paired is None:
continue
checked.append(f"{evidence.experiment_id}:{component_baseline}")
if paired["pairs"] > 0 and paired["win_count"] == 0 and paired["loss_count"] == 0:
return ClaimReviewCheck(
"candidate-distinctness",
"fail",
f"candidate is indistinguishable from component baseline {component_baseline} on reviewed holdout CER",
details=[
f"experiment={evidence.experiment_id}",
f"pairs={paired['pairs']}",
f"candidate_mean={paired['candidate_mean']:.6f}",
f"baseline_mean={paired['baseline_mean']:.6f}",
"metric=cer",
"slice=all",
],
)
if checked:
return ClaimReviewCheck(
"candidate-distinctness",
"pass",
"candidate is measurably distinct from reviewed component baselines on holdout CER",
details=checked,
)
return ClaimReviewCheck(
"candidate-distinctness",
"warn",
"no reviewed experiment includes component baselines needed for distinctness checks",
details=component_baselines,
)
def _component_baselines_for_card(card: ModelCard) -> list[str]:
return component_baseline_names(card)
def _candidate_baseline_name(baselines: list[str]) -> str | None:
for name in ("candidate", "ours", "ocrtech"):
if name in baselines:
return name
return None
def _paired_metric_outcomes(
rows: list[dict[str, Any]],
candidate_baseline: str,
baseline: str,
*,
metric: str,
slice_name: str,
tolerance: float = 1e-12,
) -> dict[str, float | int] | None:
candidate_values = _rows_by_sample(rows, candidate_baseline, metric, slice_name)
baseline_values = _rows_by_sample(rows, baseline, metric, slice_name)
shared = sorted(set(candidate_values) & set(baseline_values))
if not shared:
return None
candidate_series = [candidate_values[key] for key in shared]
baseline_series = [baseline_values[key] for key in shared]
wins = 0
losses = 0
ties = 0
for candidate_value, baseline_value in zip(candidate_series, baseline_series, strict=True):
delta = baseline_value - candidate_value
if delta > tolerance:
wins += 1
elif delta < -tolerance:
losses += 1
else:
ties += 1
return {
"pairs": len(shared),
"candidate_mean": sum(candidate_series) / len(candidate_series),
"baseline_mean": sum(baseline_series) / len(baseline_series),
"win_count": wins,
"loss_count": losses,
"tie_count": ties,
}
def _rows_by_sample(rows: list[dict[str, Any]], baseline: str, metric: str, slice_name: str) -> dict[str, float]:
values: dict[str, float] = {}
for row in rows:
if row.get("baseline") != baseline or row.get("status") != "ok":
continue
if slice_name != "all" and slice_name not in set(row.get("slices") or []):
continue
metrics = row.get("metrics")
if not isinstance(metrics, dict):
continue
value = metrics.get(metric)
if isinstance(value, int | float) and math.isfinite(float(value)):
values[sample_key(row)] = float(value)
return values
def _matches_required_experiment(evidence: ExperimentEvidence, required: RequiredExperiment) -> bool:
if required.datasets and not (set(required.datasets) & set(evidence.datasets)):
return False
if required.eval_manifest_patterns and not any(fnmatch(evidence.eval_manifest, pattern) for pattern in required.eval_manifest_patterns):
return False
return True
def _evaluate_required_experiment(evidence: ExperimentEvidence, required: RequiredExperiment) -> ClaimReviewCheck:
details: list[str] = [f"experiment={evidence.experiment_id}", f"eval_manifest={evidence.eval_manifest}"]
failures: list[str] = []
if not evidence.recorded_eval_manifest_sha256:
failures.append("experiment missing environment.eval_manifest_sha256")
elif evidence.recorded_eval_manifest_sha256 != evidence.eval_manifest_sha256:
failures.append(
"experiment eval_manifest sha256 does not match environment.eval_manifest_sha256"
)
failures.extend(_benchmark_eval_manifest_binding_failures(evidence))
if required.require_validation_status is not None and evidence.validation_status != required.require_validation_status:
failures.append(
f"validation_status {evidence.validation_status or 'not_run'} != required {required.require_validation_status}"
)
if required.required_validation_config is not None:
failures.extend(_required_validation_config_failures(evidence, required.required_validation_config))
if required.require_validation_status is not None:
if not evidence.validation_report:
failures.append("validation_report is missing")
elif not Path(evidence.validation_report).exists():
failures.append(f"validation_report does not exist: {evidence.validation_report}")
else:
failures.extend(_validation_report_failures(evidence.validation_report, required.require_validation_status, evidence))
failures.extend(_validation_gate_failures(evidence.validation_report, required.required_validation_gates))
missing_baselines = [baseline for baseline in required.require_baselines if baseline not in evidence.baselines]
if missing_baselines:
failures.append(f"missing baselines: {', '.join(missing_baselines)}")
missing_baseline_samples = _missing_required_baseline_sample_coverage(evidence, required.require_baselines)
if missing_baseline_samples:
failures.append(f"missing baseline sample coverage: {', '.join(missing_baseline_samples)}")
if required.require_no_errors and evidence.benchmark_status_counts.get("error", 0) > 0:
failures.append(f"benchmark has {evidence.benchmark_status_counts['error']} error rows")
missing_slices = [slice_name for slice_name in required.required_slices if evidence.slice_counts.get(slice_name, 0) < 1]
if missing_slices:
failures.append(f"missing slices: {', '.join(missing_slices)}")
metric_baselines = required.require_baselines or (["candidate"] if "candidate" in evidence.baselines else evidence.baselines)
missing_metrics = _missing_required_metrics(evidence.benchmark_rows, metric_baselines, required.required_metrics)
if missing_metrics:
failures.append(f"missing metrics: {', '.join(missing_metrics)}")
missing_metric_slices = _missing_required_metric_slices(evidence.benchmark_rows, metric_baselines, required.required_metric_slices)
if missing_metric_slices:
failures.append(f"missing metric slice coverage: {', '.join(missing_metric_slices)}")
missing_paired_metrics = _missing_required_paired_metric_coverage(
evidence.benchmark_rows,
metric_baselines,
required.required_metrics,
required.required_metric_slices,
)
if missing_paired_metrics:
failures.append(f"missing paired metric coverage: {', '.join(missing_paired_metrics)}")
missing_artifacts = _missing_required_output_artifacts(evidence, required)
if missing_artifacts:
failures.append(f"missing output artifacts: {', '.join(missing_artifacts)}")
failures.extend(_candidate_fallback_policy_failures(evidence.benchmark_rows, evidence.baselines, required))
details.append(f"datasets={', '.join(evidence.datasets)}")
details.append(f"baselines={', '.join(evidence.baselines)}")
if required.required_metrics:
details.append(f"required_metrics={', '.join(required.required_metrics)}")
if required.required_validation_config is not None:
details.append(f"required_validation_config={required.required_validation_config}")
if required.required_metric_slices:
details.append(
"required_metric_slices="
+ ", ".join(f"{metric}@{'+'.join(slices)}" for metric, slices in sorted(required.required_metric_slices.items()))
)
if required.required_output_artifacts:
details.append(f"required_output_artifacts={', '.join(required.required_output_artifacts)}")
if required.required_output_artifact_slices:
details.append(
"required_output_artifact_slices="
+ ", ".join(f"{artifact}@{'+'.join(slices)}" for artifact, slices in sorted(required.required_output_artifact_slices.items()))
)
if required.required_validation_gates:
details.append(
"required_validation_gates="
+ ", ".join(
f"{gate.baseline}:{gate.metric}@{'+'.join(gate.slices) if gate.slices else 'any'}"
for gate in required.required_validation_gates
)
)
if required.require_candidate_fallback_metrics:
details.append("require_candidate_fallback_metrics=true")
if required.max_candidate_fallback_trigger_rate is not None:
details.append(f"max_candidate_fallback_trigger_rate={required.max_candidate_fallback_trigger_rate:.6f}")
if required.max_candidate_fallback_failed_rate is not None:
details.append(f"max_candidate_fallback_failed_rate={required.max_candidate_fallback_failed_rate:.6f}")
if failures:
return ClaimReviewCheck(f"required:{required.name}", "fail", "; ".join(failures), details=details)
summary = "required experiment evidence is present and passes configured checks"
return ClaimReviewCheck(f"required:{required.name}", "pass", summary, details=details)
def _load_benchmark_report(path: str | Path) -> list[dict[str, Any]]:
_require_existing_file(Path(path), "benchmark report")
try:
payload = json.loads(Path(path).read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
raise ClaimReviewError(f"invalid benchmark report JSON {path}: {exc}") from exc
if not isinstance(payload, list):
raise ClaimReviewError(f"benchmark report must be a JSON list: {path}")
rows: list[dict[str, Any]] = []
for index, item in enumerate(payload):
if not isinstance(item, dict):
raise ClaimReviewError(f"benchmark row {index} must be an object: {path}")
rows.append(dict(item))
return rows
def _benchmark_eval_manifest_binding_failures(evidence: ExperimentEvidence) -> list[str]:
entries_by_sample = {entry.sample_id: entry for entry in evidence.eval_entries}
failures: list[str] = []
for row in evidence.benchmark_rows:
if row.get("status") != "ok":
continue
sample_id_value = row.get("sample_id")
if not isinstance(sample_id_value, str) or not sample_id_value:
continue
entry = entries_by_sample.get(sample_id_value)
if entry is None:
_append_once(failures, f"benchmark row sample_id not in eval_manifest: {sample_id_value}")
continue
input_path_value = row.get("input_path")
if not isinstance(input_path_value, str) or not input_path_value:
_append_once(failures, f"benchmark row missing input_path for sample_id: {sample_id_value}")
continue
if not _same_path(input_path_value, entry.image_path):
_append_once(failures, f"benchmark row input_path does not match eval_manifest for sample_id: {sample_id_value}")
row_slices = row.get("slices")
if not isinstance(row_slices, list):
_append_once(failures, f"benchmark row slices missing or invalid for sample_id: {sample_id_value}")
continue
if {str(item) for item in row_slices if str(item)} != set(_entry_slices(entry)):
_append_once(failures, f"benchmark row slices do not match eval_manifest for sample_id: {sample_id_value}")
return failures
def _append_once(items: list[str], value: str) -> None:
if value not in items:
items.append(value)
def _missing_required_baseline_sample_coverage(evidence: ExperimentEvidence, required_baselines: list[str]) -> list[str]:
if not required_baselines:
return []
required_samples = {entry.sample_id for entry in evidence.eval_entries}
if not required_samples:
return []
missing: list[str] = []
for baseline in required_baselines:
ok_samples = {
str(row.get("sample_id"))
for row in evidence.benchmark_rows
if row.get("baseline") == baseline
and row.get("status") == "ok"
and isinstance(row.get("sample_id"), str)
and row.get("sample_id")
}
baseline_missing = sorted(required_samples - ok_samples)
if not baseline_missing:
continue
preview = ", ".join(baseline_missing[:5])
suffix = f" (+{len(baseline_missing) - 5} more)" if len(baseline_missing) > 5 else ""
missing.append(f"{baseline}:{preview}{suffix}")
return missing
def _same_path(left: str, right: str) -> bool:
left_path = Path(left)
right_path = Path(right)
try:
if left_path.exists() and right_path.exists():
return left_path.resolve() == right_path.resolve()
except OSError:
pass
return str(left_path) == str(right_path)
def _required_validation_config_failures(evidence: ExperimentEvidence, required_validation_config: str) -> list[str]:
config_path = Path(required_validation_config)
if not config_path.exists():
return [f"required validation config does not exist: {required_validation_config}"]
if not config_path.is_file():
return [f"required validation config is not a file: {required_validation_config}"]
expected_sha = sha256_file(config_path)
failures: list[str] = []
if not evidence.validation_config:
failures.append(f"experiment missing required validation_config: {required_validation_config}")
if not evidence.validation_config_sha256:
failures.append(f"experiment missing validation_config_sha256 for required config: {required_validation_config}")
elif evidence.validation_config_sha256 != expected_sha:
failures.append(
f"experiment validation_config sha256 does not match required config {required_validation_config}: "
f"expected={expected_sha} actual={evidence.validation_config_sha256}"
)
return failures
def _validation_report_failures(path: str | Path, required_status: str, evidence: ExperimentEvidence) -> list[str]:
report_path = Path(path)
try:
payload = json.loads(report_path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
return [f"validation_report invalid JSON: {report_path}: {exc}"]
if not isinstance(payload, dict):
return [f"validation_report must be a JSON object: {report_path}"]
claim_status = str(payload.get("claim_status") or "")
if claim_status != required_status:
return [f"validation_report claim_status {claim_status or 'not_run'} != required {required_status}"]
if required_status == "validated" and payload.get("passed") is not True:
return ["validation_report passed is not true"]
failures: list[str] = []
provenance = payload.get("provenance")
if not isinstance(provenance, dict):
return ["validation_report missing provenance"]
benchmark_payload = provenance.get("benchmark_report")
if not isinstance(benchmark_payload, dict):
failures.append("validation_report missing provenance.benchmark_report")
else:
benchmark_sha = benchmark_payload.get("sha256")
expected_benchmark_sha = sha256_file(Path(evidence.benchmark_report))
if benchmark_sha != expected_benchmark_sha:
failures.append("validation_report benchmark_report sha256 does not match experiment benchmark_report")
eval_payload = provenance.get("eval_manifest")
if not isinstance(eval_payload, dict):
failures.append("validation_report missing provenance.eval_manifest")
else:
eval_sha = eval_payload.get("sha256")
if eval_sha != evidence.eval_manifest_sha256:
failures.append("validation_report eval_manifest sha256 does not match experiment eval_manifest")
config_payload = provenance.get("validation_config")
if evidence.validation_config_sha256 is not None:
if not isinstance(config_payload, dict):
failures.append("validation_report missing provenance.validation_config")
elif config_payload.get("sha256") != evidence.validation_config_sha256:
failures.append("validation_report validation_config sha256 does not match experiment validation_config")
elif config_payload not in (None, {}):
failures.append("validation_report validation_config provenance is present but experiment has no validation_config")
validation_train_manifests = _environment_manifest_hashes(provenance.get("train_manifests"))
if _manifest_hash_signature(validation_train_manifests) != _manifest_hash_signature(evidence.train_manifests):
failures.append("validation_report train_manifests provenance does not match experiment train_manifests")
_append_optional_provenance_hash_failure(
failures,
provenance,
"candidate_model_config",
evidence.candidate_model_config_sha256,
"candidate_model_config",
)
_append_optional_provenance_hash_failure(
failures,
provenance,
"fallback_model_config",
evidence.fallback_model_config_sha256,
"fallback_model_config",
)
return failures
def _append_optional_provenance_hash_failure(
failures: list[str],
provenance: dict[str, Any],
key: str,
expected_sha256: str | None,
label: str,
) -> None:
payload = provenance.get(key)
if expected_sha256 is not None:
if not isinstance(payload, dict):
failures.append(f"validation_report missing provenance.{key}")
elif payload.get("sha256") != expected_sha256:
failures.append(f"validation_report {label} sha256 does not match experiment {label}")
elif payload not in (None, {}):
failures.append(f"validation_report {label} provenance is present but experiment has no {label}")
def _validation_gate_failures(path: str | Path, required_gates: list[RequiredValidationGate]) -> list[str]:
if not required_gates:
return []
report_path = Path(path)
try:
payload = json.loads(report_path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
return [f"validation_report invalid JSON: {report_path}: {exc}"]
if not isinstance(payload, dict):
return [f"validation_report must be a JSON object: {report_path}"]
gate_results = payload.get("gates")
if not isinstance(gate_results, list):
return ["validation_report missing gates"]
failures: list[str] = []
for required in required_gates:
if not any(_validation_gate_matches(item, required) for item in gate_results):
slice_suffix = "@" + "+".join(required.slices) if required.slices else ""
failures.append(f"missing passing validation gate: {required.baseline}:{required.metric}{slice_suffix}")
return failures
def _validation_gate_matches(item: Any, required: RequiredValidationGate) -> bool:
if not isinstance(item, dict) or item.get("passed") is not True:
return False
gate = item.get("gate")
if not isinstance(gate, dict):
return False
if gate.get("baseline") != required.baseline or gate.get("metric") != required.metric:
return False
if not required.slices:
return gate_has_quantitative_evidence(item)
gate_slices = gate.get("slices")
if not isinstance(gate_slices, list):
gate_slice = gate.get("slice")
gate_slices = [gate_slice] if gate_slice else []
return set(required.slices).issubset({str(slice_name) for slice_name in gate_slices}) and gate_has_quantitative_evidence(item)
def _require_existing_file(path: Path, label: str) -> None:
if not path.exists():
raise ClaimReviewError(f"{label} does not exist: {path}")
if not path.is_file():
raise ClaimReviewError(f"{label} is not a file: {path}")
def _missing_required_metrics(rows: list[dict[str, Any]], baselines: list[str], required_metrics: list[str]) -> list[str]:
missing: list[str] = []
if not required_metrics:
return missing
for baseline in baselines:
ok_rows = [row for row in rows if row.get("baseline") == baseline and row.get("status") == "ok"]
for metric in required_metrics:
if not any(_row_has_finite_metric(row, metric) for row in ok_rows):
missing.append(f"{baseline}:{metric}")
return missing
def _missing_required_metric_slices(rows: list[dict[str, Any]], baselines: list[str], requirements: dict[str, list[str]]) -> list[str]:
missing: list[str] = []
if not requirements:
return missing
for baseline in baselines:
ok_rows = [row for row in rows if row.get("baseline") == baseline and row.get("status") == "ok"]
for metric, slices in requirements.items():
for slice_name in slices:
if not any(slice_name in set(row.get("slices") or []) and _row_has_finite_metric(row, metric) for row in ok_rows):
missing.append(f"{baseline}:{metric}@{slice_name}")
return missing
def _missing_required_paired_metric_coverage(
rows: list[dict[str, Any]],
baselines: list[str],
required_metrics: list[str],
required_metric_slices: dict[str, list[str]],
) -> list[str]:
missing: list[str] = []
if not required_metrics:
return missing
candidate = _candidate_baseline_name(baselines)
if candidate is None:
return missing
comparison_baselines = [baseline for baseline in baselines if baseline != candidate]
for baseline in comparison_baselines:
for metric in required_metrics:
slice_names = required_metric_slices.get(metric) or ["all"]
for slice_name in slice_names:
if _paired_metric_outcomes(rows, candidate, baseline, metric=metric, slice_name=slice_name) is None:
missing.append(f"{candidate}:{baseline}:{metric}@{slice_name}")
return missing
def _missing_required_output_artifacts(evidence: ExperimentEvidence, required: RequiredExperiment) -> list[str]:
if not required.required_output_artifacts and not required.required_output_artifact_slices:
return []
candidate = _candidate_baseline_name(evidence.baselines)
if candidate is None:
return ["candidate output artifacts require a candidate/ours baseline"]
ok_rows = [row for row in evidence.benchmark_rows if row.get("baseline") == candidate and row.get("status") == "ok"]
if not ok_rows:
return [f"candidate output artifacts have no ok rows for {candidate}"]
benchmark_dir = Path(evidence.benchmark_report).parent
missing: list[str] = []
for row in ok_rows:
output_dir = benchmark_dir / candidate / Path(str(row.get("input_path") or "")).stem
sample = sample_key(row)
for artifact in required.required_output_artifacts:
artifact_issue = _output_artifact_issue(output_dir, artifact)
if artifact_issue is not None:
missing.append(f"{candidate}:{sample}:{artifact} ({artifact_issue})")
row_slices = {str(item) for item in row.get("slices", [])}
for artifact, slices in required.required_output_artifact_slices.items():
for slice_name in slices:
if slice_name not in row_slices:
continue
artifact_issue = _output_artifact_issue(output_dir, artifact)
if artifact_issue is not None:
missing.append(f"{candidate}:{sample}:{artifact}@{slice_name} ({artifact_issue})")
return missing
def _output_artifact_issue(output_dir: Path, artifact: str) -> str | None:
if artifact == "document_json":
return _json_file_issue(output_dir / "document.json", expected_type=dict, required_keys=("pages", "tables", "figures"))
if artifact == "document_md":
return _nonempty_file_issue(output_dir / "document.md")
if artifact == "quality_json":
return _json_file_issue(output_dir / "quality.json", expected_type=dict)
if artifact == "table_csv_html":
tables_dir = output_dir / "tables"
if not tables_dir.is_dir():
return "tables directory missing"
csv_issue = _any_nonempty_file_issue(tables_dir, "*.csv")
html_issue = _any_nonempty_file_issue(tables_dir, "*.html")
if csv_issue is not None:
return f"csv {csv_issue}"
if html_issue is not None:
return f"html {html_issue}"
return None
if artifact == "figure_metadata":
return _json_file_issue(output_dir / "figures" / "metadata.json", expected_type=list, require_nonempty=True)
if artifact == "figure_images":
figures_dir = output_dir / "figures"
if not figures_dir.is_dir():
return "figures directory missing"
image_files = [
path
for path in figures_dir.iterdir()
if path.is_file() and path.suffix.lower() in {".png", ".jpg", ".jpeg", ".webp"}
]
if not image_files:
return "no image files"
if not any(path.stat().st_size > 0 for path in image_files):
return "image files are empty"
return None
return "unknown artifact type"
def _json_file_issue(
path: Path,
*,
expected_type: type[dict[str, Any]] | type[list[Any]],
required_keys: tuple[str, ...] = (),
require_nonempty: bool = False,
) -> str | None:
if not path.is_file():
return "missing"
if path.stat().st_size == 0:
return "empty"
try:
payload = json.loads(path.read_text(encoding="utf-8"))
except json.JSONDecodeError as exc:
return f"invalid json: {exc}"
if not isinstance(payload, expected_type):
return f"expected {expected_type.__name__}"
if require_nonempty and not payload:
return "empty payload"
if isinstance(payload, dict):
missing_keys = [key for key in required_keys if key not in payload]
if missing_keys:
return f"missing keys: {', '.join(missing_keys)}"
return None
def _nonempty_file_issue(path: Path) -> str | None:
if not path.is_file():
return "missing"
if path.stat().st_size == 0:
return "empty"
return None
def _any_nonempty_file_issue(directory: Path, pattern: str) -> str | None:
files = [path for path in directory.glob(pattern) if path.is_file()]
if not files:
return f"missing {pattern}"
if not any(path.stat().st_size > 0 for path in files):
return f"empty {pattern}"
return None
def _candidate_fallback_policy_failures(rows: list[dict[str, Any]], baselines: list[str], required: RequiredExperiment) -> list[str]:
if (
not required.require_candidate_fallback_metrics
and required.max_candidate_fallback_trigger_rate is None
and required.max_candidate_fallback_failed_rate is None
):
return []
candidate = _candidate_baseline_name(baselines)
if candidate is None:
return ["candidate fallback policy requires a candidate/ours baseline"]
ok_rows = [row for row in rows if row.get("baseline") == candidate and row.get("status") == "ok"]
if not ok_rows:
return [f"candidate fallback policy has no ok rows for {candidate}"]
failures: list[str] = []
if required.require_candidate_fallback_metrics:
missing_triggered = [sample_key(row) for row in ok_rows if not _row_has_finite_metric(row, "fallback_triggered")]
missing_failed = [sample_key(row) for row in ok_rows if not _row_has_finite_metric(row, "fallback_failed")]
if missing_triggered:
failures.append(f"missing candidate fallback_triggered metrics on {len(missing_triggered)} rows")
if missing_failed:
failures.append(f"missing candidate fallback_failed metrics on {len(missing_failed)} rows")
if required.max_candidate_fallback_trigger_rate is not None:
trigger_rate = _mean_metric(ok_rows, "fallback_triggered")
if trigger_rate is None:
failures.append("candidate fallback_triggered metrics are missing for trigger-rate limit")
elif trigger_rate > required.max_candidate_fallback_trigger_rate:
failures.append(
f"candidate fallback trigger rate {trigger_rate:.6f} exceeds max {required.max_candidate_fallback_trigger_rate:.6f}"
)
if required.max_candidate_fallback_failed_rate is not None:
failed_rate = _mean_metric(ok_rows, "fallback_failed")
if failed_rate is None:
failures.append("candidate fallback_failed metrics are missing for failure-rate limit")
elif failed_rate > required.max_candidate_fallback_failed_rate:
failures.append(
f"candidate fallback failure rate {failed_rate:.6f} exceeds max {required.max_candidate_fallback_failed_rate:.6f}"
)
return failures
def _mean_metric(rows: list[dict[str, Any]], metric: str) -> float | None:
values: list[float] = []
for row in rows:
metrics = row.get("metrics")
if not isinstance(metrics, dict):
continue
value = metrics.get(metric)
if isinstance(value, int | float) and math.isfinite(float(value)):
values.append(float(value))
if not values:
return None
return sum(values) / len(values)
def _row_has_finite_metric(row: dict[str, Any], metric: str) -> bool:
metrics = row.get("metrics")
if not isinstance(metrics, dict):
return False
value = metrics.get(metric)
return isinstance(value, int | float) and math.isfinite(float(value))
def _optional_rate(value: Any, field_name: str) -> float | None:
if value is None:
return None
if not isinstance(value, int | float):
raise ClaimReviewError(f"{field_name} must be a number between 0 and 1")
rate = float(value)
if not math.isfinite(rate) or rate < 0.0 or rate > 1.0:
raise ClaimReviewError(f"{field_name} must be a number between 0 and 1")
return rate
def _status_counts(rows: list[dict[str, Any]]) -> dict[str, int]:
counts: dict[str, int] = {}
for item in rows:
status = str(item.get("status") or "")
if not status:
continue
counts[status] = counts.get(status, 0) + 1
return dict(sorted(counts.items()))
def _slice_counts(entries: list[ManifestEntry]) -> dict[str, int]:
counts: dict[str, int] = {}
seen: set[tuple[str, str]] = set()
for entry in entries:
for slice_name in _entry_slices(entry):
pair = (entry.sample_id, slice_name)
if pair in seen:
continue
seen.add(pair)
counts[slice_name] = counts.get(slice_name, 0) + 1
return dict(sorted(counts.items()))
def _entry_slices(entry: ManifestEntry) -> list[str]:
values: set[str] = set()
metadata = entry.metadata or {}
for key in ("slice", "script", "document_type", "degradation", "language"):
value = metadata.get(key)
if isinstance(value, str) and value:
values.add(value)
for key in ("slices", "scripts", "tags"):
value = metadata.get(key)
if isinstance(value, list):
values.update(str(item) for item in value if str(item))
elif isinstance(value, str) and value:
values.add(value)
return sorted(values)
def _string_list(value: Any) -> list[str]:
if value is None:
return []
if not isinstance(value, list):
raise ClaimReviewError("expected a list of strings")
result = [str(item) for item in value if str(item)]
if len(result) != len(value):
raise ClaimReviewError("list items must be non-empty strings")
return result
def _environment_manifest_hashes(value: Any) -> list[dict[str, str | None]]:
if not isinstance(value, list):
return []
manifests: list[dict[str, str | None]] = []
for item in value:
if not isinstance(item, dict):
continue
path = item.get("path")
sha = item.get("sha256")
manifests.append(
{
"path": str(path) if path is not None else None,
"sha256": str(sha) if sha is not None else None,
}
)
return manifests
def _manifest_hash_signature(manifests: list[dict[str, str | None]]) -> list[tuple[str | None, str | None]]:
return sorted(
((item.get("path"), item.get("sha256")) for item in manifests),
key=lambda item: (item[0] or "", item[1] or ""),
)
def _metric_slice_requirements(value: Any) -> dict[str, list[str]]:
if value is None:
return {}
if not isinstance(value, dict):
raise ClaimReviewError("required_metric_slices must be an object mapping metric names to slice lists")
result: dict[str, list[str]] = {}
for metric, slices in value.items():
metric_name = str(metric)
if not metric_name:
raise ClaimReviewError("required_metric_slices metric names must be non-empty strings")
parsed_slices = _string_list(slices)
if not parsed_slices:
raise ClaimReviewError(f"required_metric_slices for {metric_name!r} must define at least one slice")
result[metric_name] = parsed_slices
return result
def _required_validation_gates(value: Any) -> list[RequiredValidationGate]:
if value is None:
return []
if not isinstance(value, list):
raise ClaimReviewError("required_validation_gates must be a list")
gates: list[RequiredValidationGate] = []
for item in value:
if not isinstance(item, dict):
raise ClaimReviewError("required_validation_gates items must be objects")
gates.append(RequiredValidationGate.from_dict(item))
return gates
def _validate_required_validation_gates(
experiment_name: str,
require_validation_status: str | None,
require_baselines: list[str],
required_metrics: list[str],
required_validation_gates: list[RequiredValidationGate],
) -> None:
if not required_validation_gates:
return
if require_validation_status is None:
raise ClaimReviewError(f"required experiment {experiment_name!r} defines required_validation_gates without require_validation_status")
baseline_set = set(require_baselines)
missing_baselines = sorted({gate.baseline for gate in required_validation_gates if gate.baseline not in baseline_set})
if missing_baselines:
raise ClaimReviewError(
f"required experiment {experiment_name!r} has validation gate baselines not listed in require_baselines: {', '.join(missing_baselines)}"
)
metric_set = set(required_metrics)
missing_metrics = sorted({gate.metric for gate in required_validation_gates if gate.metric not in metric_set})
if missing_metrics:
raise ClaimReviewError(
f"required experiment {experiment_name!r} has validation gate metrics not listed in required_metrics: {', '.join(missing_metrics)}"
)
def _write_claim_review(report: ClaimReviewReport, output_dir: Path) -> None:
(output_dir / "claim-review.json").write_text(json.dumps(report.to_dict(), ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
lines = [
"# OCR Claim Review",
"",
f"Passed: `{'yes' if report.passed else 'no'}`",
f"Model config: `{report.model_config or ''}`",
"",
"| check | status | summary |",
"| --- | --- | --- |",
]
for check in report.checks:
summary = check.summary.replace("|", "\\|")
lines.append(f"| {check.name} | {check.status} | {summary} |")
for detail in check.details:
escaped_detail = detail.replace("|", "\\|")
lines.append(f"| | | {escaped_detail} |")
lines.extend(["", "## Experiments", ""])
for experiment in report.experiments:
lines.append(f"- `{experiment.experiment_id}` datasets=`{', '.join(experiment.datasets)}` baselines=`{', '.join(experiment.baselines)}`")
(output_dir / "claim-review.md").write_text("\n".join(lines) + "\n", encoding="utf-8")
|