File size: 65,294 Bytes
2955ecc | 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 | from __future__ import annotations
import json
import sys
import unittest
from copy import deepcopy
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
SRC = ROOT / "src"
if str(SRC) not in sys.path:
sys.path.insert(0, str(SRC))
from datacenter_verification.observable_algorithm import (
_capacity_claim_contradictions,
_concurrent_peak,
_max,
evaluate_site,
)
class ObservableAlgorithmTest(unittest.TestCase):
@classmethod
def setUpClass(cls) -> None:
payload = json.loads((ROOT / "synthetic" / "sites.json").read_text(encoding="utf-8"))
cls.sites = {site["scenario_key"]: site for site in payload["sites"]}
cls.results = {key: evaluate_site(site) for key, site in cls.sites.items()}
def result(self, key: str) -> dict:
return self.results[key]
def stage(self, key: str, stage: str) -> dict:
return self.result(key)["stage_outputs"][stage]
def suppress_identity_pathways(self, site: dict, keep_participant: bool = False) -> None:
signals = site.setdefault("normalized_signals", {})
signals["collective_cadence_score"] = 0.0
signals["checkpoint_periodicity_score"] = 0.0
signals["checkpoint_burst_count"] = 0.0
signals["activity_fabric_overlap_fraction"] = 0.0
signals["checkpoint_activity_adjacency_fraction"] = 0.0
if not keep_participant:
signals["participant_count"] = 0.0
for record in site.setdefault("raw_features", {}).get("fabric_port_device_sample_counters", []):
if record.get("counter_name") == "collective_cadence_score":
record["counter_value"] = 0.0
if not keep_participant and record.get("counter_name") == "participant_count":
record["counter_value"] = 0.0
def set_raw_rate_integral(self, site: dict, operations: float) -> None:
duration = 30 * 24 * 3600
for record in site.setdefault("raw_features", {}).get("generic_achieved_operation_rate", []):
record["operation_rate"] = operations / duration
def test_all_synthetic_expected_outputs_match(self) -> None:
for key, site in self.sites.items():
with self.subTest(key=key):
result = self.results[key]
expected = site["expected"]
self.assertEqual(expected["A_capacity_gate_label"], self.stage(key, "A_capacity_gate")["label"])
self.assertEqual(expected["final_route"], result["final_route"])
self.assertEqual(expected["capacity_short_circuit"], self.stage(key, "A_capacity_gate")["short_circuited"])
b_labels = self.stage(key, "B_training_candidate_detection")["labels"]
c_labels = self.stage(key, "C_discrepancy_and_explanation_review")["labels"]
for label in expected["B_training_candidate_detection_labels"]:
self.assertIn(label, b_labels)
for label in expected["C_discrepancy_and_explanation_review_labels"]:
self.assertIn(label, c_labels)
def test_capacity_ruleout_short_circuits_b_and_c(self) -> None:
result = self.result("C_capacity_ruled_out")
self.assertEqual("capacity_ruled_out_for_scope", result["final_route"])
self.assertTrue(self.stage("C_capacity_ruled_out", "A_capacity_gate")["short_circuited"])
self.assertEqual("skipped_due_to_capacity_ruleout", self.stage("C_capacity_ruled_out", "B_training_candidate_detection")["mode"])
self.assertEqual("skipped_due_to_capacity_ruleout", self.stage("C_capacity_ruled_out", "C_discrepancy_and_explanation_review")["mode"])
def test_clean_training_reaches_high_warning(self) -> None:
result = self.result("A_clean_threshold_training")
b = self.stage("A_clean_threshold_training", "B_training_candidate_detection")
c = self.stage("A_clean_threshold_training", "C_discrepancy_and_explanation_review")
self.assertEqual("high_training_like_warning", result["final_route"])
self.assertIn("distributed_training_like_candidate", b["labels"])
self.assertIn("checkpoint_training_like_candidate", b["labels"])
self.assertEqual("C2_candidate_conflict_adjudication", c["mode"])
self.assertFalse(c["discrepancies"])
self.assertFalse(c["missing_channels"])
def test_algorithm_version_is_v0_3(self) -> None:
self.assertEqual("observable_staged_v0.3", self.result("A_clean_threshold_training")["algorithm_version"])
def test_large_compute_alone_does_not_become_medium_or_high_training(self) -> None:
result = self.result("K_large_compute_alone")
b = self.stage("K_large_compute_alone", "B_training_candidate_detection")
self.assertEqual("weak_training_like_candidate", result["final_route"])
self.assertIn("large_compute_candidate", b["labels"])
self.assertNotIn("distributed_training_like_candidate", b["labels"])
self.assertNotIn("checkpoint_training_like_candidate", b["labels"])
self.assertIn("large_compute_training_identity_unresolved", result["caveats"])
def test_activity_alone_stays_clean_negative(self) -> None:
result = self.result("L_activity_alone")
b = self.stage("L_activity_alone", "B_training_candidate_detection")
c = self.stage("L_activity_alone", "C_discrepancy_and_explanation_review")
self.assertEqual("no_training_like_candidate_detected_in_covered_live_segment", result["final_route"])
self.assertEqual([], b["labels"])
self.assertEqual("C1_negative_screen_integrity", c["mode"])
self.assertIn("negative_screen_coverage_sufficient", c["labels"])
def test_fabric_alone_and_storage_alone_route_integrity(self) -> None:
cases = {
"M_fabric_alone": "fabric_without_job_or_topology_mapping_conflict",
"N_storage_writes_alone": "checkpoint_writes_without_activity_conflict",
}
for key, discrepancy in cases.items():
with self.subTest(case=key):
result = self.result(key)
b = self.stage(key, "B_training_candidate_detection")
c = self.stage(key, "C_discrepancy_and_explanation_review")
self.assertEqual("integrity_review_required", result["final_route"])
self.assertEqual([], b["labels"])
self.assertEqual("C1_negative_screen_integrity", c["mode"])
self.assertIn("negative_screen_incoherence_conflict", c["labels"])
self.assertIn(discrepancy, c["discrepancies"])
def test_storage_explanation_demotes_checkpoint_candidate(self) -> None:
result = self.result("E_storage_operation_explains_checkpoint")
b = self.stage("E_storage_operation_explains_checkpoint", "B_training_candidate_detection")
c = self.stage("E_storage_operation_explains_checkpoint", "C_discrepancy_and_explanation_review")
self.assertIn("checkpoint_training_like_candidate", b["labels"])
self.assertIn("candidate_explained_by_storage_operation", c["labels"])
self.assertEqual("candidate_explained_or_demoted", result["final_route"])
def test_serving_counterevidence_demotes_large_compute_candidate(self) -> None:
result = self.result("F_serving_inference_counterevidence")
c = self.stage("F_serving_inference_counterevidence", "C_discrepancy_and_explanation_review")
self.assertIn("candidate_explained_by_serving", c["labels"])
self.assertEqual("candidate_explained_or_demoted", result["final_route"])
def test_benchmark_and_hpc_alternative_demotes_fabric_candidate(self) -> None:
result = self.result("G_hpc_mpi_benchmark_alternative")
c = self.stage("G_hpc_mpi_benchmark_alternative", "C_discrepancy_and_explanation_review")
self.assertIn("candidate_benchmark_like", c["labels"])
self.assertIn("candidate_hpc_mpi_alternative", c["labels"])
self.assertEqual("candidate_explained_or_demoted", result["final_route"])
def test_covered_negative_and_missing_negative_screen_routes(self) -> None:
covered = self.result("B_covered_negative")
missing = self.result("D_missingness_blocks_negative_screen")
self.assertEqual("no_training_like_candidate_detected_in_covered_live_segment", covered["final_route"])
self.assertIn(
"negative_screen_coverage_sufficient",
self.stage("B_covered_negative", "C_discrepancy_and_explanation_review")["labels"],
)
self.assertEqual("inconclusive_due_to_missingness", missing["final_route"])
self.assertIn(
"negative_screen_blocked_by_missingness",
self.stage("D_missingness_blocks_negative_screen", "C_discrepancy_and_explanation_review")["labels"],
)
def test_capacity_and_activity_attribution_conflicts_route_integrity(self) -> None:
capacity = self.result("H_capacity_claim_conflict")
attribution = self.result("I_activity_attribution_conflict")
self.assertEqual("integrity_review_required", capacity["final_route"])
self.assertIn(
"capacity_claim_conflict",
self.stage("H_capacity_claim_conflict", "C_discrepancy_and_explanation_review")["labels"],
)
self.assertEqual("integrity_review_required", attribution["final_route"])
self.assertIn(
"activity_attribution_conflict",
self.stage("I_activity_attribution_conflict", "C_discrepancy_and_explanation_review")["labels"],
)
# --- FIX 3 (F1, predicate A2'): C1 incoherence integrity ---
def _f1_probe(self, fabric: float | None = None, checkpoint: float | None = None) -> dict:
probe = deepcopy(self.sites["A_clean_threshold_training"])
probe["normalized_signals"]["activity_score"] = 0.30
for key in (
"accelerator_busy_or_utilization_fraction",
"tensor_matrix_mxu_neuron_or_engine_active_fraction",
):
if key in probe.get("raw_features", {}):
probe["raw_features"][key] = [{"value": 0.30}]
probe["normalized_signals"]["achieved_operations"] = 5e24
if fabric is not None:
probe["normalized_signals"]["collective_cadence_score"] = fabric
if checkpoint is not None:
probe["normalized_signals"]["checkpoint_periodicity_score"] = checkpoint
return probe
def test_f1_shaped_activity_does_not_certify_clean_negative(self) -> None:
result = evaluate_site(self._f1_probe())
self.assertEqual("integrity_review_required", result["final_route"])
self.assertEqual(
"C1_negative_screen_integrity",
result["stage_outputs"]["C_discrepancy_and_explanation_review"]["mode"],
)
def test_f1_fabric_only_variant_routes_integrity(self) -> None:
result = evaluate_site(self._f1_probe(checkpoint=0.0))
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("fabric_without_job_or_topology_mapping_conflict", result["discrepancy_findings"])
def test_f1_checkpoint_only_variant_routes_integrity(self) -> None:
result = evaluate_site(self._f1_probe(fabric=0.0))
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("checkpoint_writes_without_activity_conflict", result["discrepancy_findings"])
def test_f1_covered_negative_and_storage_explained_unaffected(self) -> None:
self.assertEqual(
"no_training_like_candidate_detected_in_covered_live_segment",
self.result("B_covered_negative")["final_route"],
)
self.assertEqual(
"candidate_explained_or_demoted",
self.result("E_storage_operation_explains_checkpoint")["final_route"],
)
# --- high-activity alignment shading cannot certify absence ---
def _alignment_shading_probe(
self,
activity: float = 0.91,
fabric_overlap: float = 0.49,
checkpoint_adjacency: float = 0.49,
achieved: float = 9.0e24,
fabric_score: float = 0.86,
checkpoint_score: float = 0.82,
) -> dict:
probe = deepcopy(self.sites["A_clean_threshold_training"])
sig = probe["normalized_signals"]
sig["activity_score"] = activity
sig["activity_fabric_overlap_fraction"] = fabric_overlap
sig["checkpoint_activity_adjacency_fraction"] = checkpoint_adjacency
sig["achieved_operations"] = achieved
sig["collective_cadence_score"] = fabric_score
sig["checkpoint_periodicity_score"] = checkpoint_score
for key in (
"accelerator_busy_or_utilization_fraction",
"tensor_matrix_mxu_neuron_or_engine_active_fraction",
):
if key in probe.get("raw_features", {}):
probe["raw_features"][key] = [{"value": activity}]
return probe
def test_fatal1_exact_alignment_shading_does_not_certify_absence(self) -> None:
result = evaluate_site(self._alignment_shading_probe())
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("fabric_activity_alignment_incoherence_conflict", result["discrepancy_findings"])
self.assertIn("checkpoint_activity_alignment_incoherence_conflict", result["discrepancy_findings"])
self.assertNotIn(
result["final_route"],
{"capacity_ruled_out_for_scope", "no_training_like_candidate_detected_in_covered_live_segment"},
)
def test_fatal1_alignment_shading_activity_band_variants(self) -> None:
for activity in (0.55, 0.70, 0.95):
with self.subTest(activity=activity):
result = evaluate_site(self._alignment_shading_probe(activity=activity))
self.assertEqual("integrity_review_required", result["final_route"])
def test_fatal1_single_alignment_shaded_variants_do_not_certify_absence(self) -> None:
cases = [
("fabric_only", 0.49, 0.78, "checkpoint_training_like_candidate"),
("checkpoint_only", 0.82, 0.49, "distributed_training_like_candidate"),
]
for name, fabric_overlap, checkpoint_adj, surviving_b_label in cases:
with self.subTest(name=name):
result = evaluate_site(
self._alignment_shading_probe(
fabric_overlap=fabric_overlap,
checkpoint_adjacency=checkpoint_adj,
)
)
self.assertNotIn(
result["final_route"],
{"capacity_ruled_out_for_scope", "no_training_like_candidate_detected_in_covered_live_segment"},
)
self.assertIn(surviving_b_label, result["stage_outputs"]["B_training_candidate_detection"]["labels"])
def test_fatal1_alignment_gate_boundary_forms_candidates_normally(self) -> None:
result = evaluate_site(
self._alignment_shading_probe(
fabric_overlap=0.50,
checkpoint_adjacency=0.50,
achieved=1.15e25,
)
)
b_labels = result["stage_outputs"]["B_training_candidate_detection"]["labels"]
self.assertIn("distributed_training_like_candidate", b_labels)
self.assertIn("checkpoint_training_like_candidate", b_labels)
self.assertEqual("high_training_like_warning", result["final_route"])
def test_fatal1_capacity_adjustment_deflation_with_raw_capacity_witness_is_blocked(self) -> None:
probe = deepcopy(self.sites["A_clean_threshold_training"])
probe["normalized_signals"]["capacity_adjustment_factor"] = 0.20
probe["normalized_signals"]["achieved_operations"] = 6.0e24
self.set_raw_rate_integral(probe, 6.0e24)
result = evaluate_site(probe)
self.assertEqual("integrity_review_required", result["final_route"])
contradictions = _capacity_claim_contradictions(result["derived_signals"], probe)
self.assertIn("capacity_adjustment_deflates_raw_threshold_capacity", contradictions)
def test_fatal1_benign_explanation_does_not_suppress_alignment_incoherence(self) -> None:
probe = self._alignment_shading_probe(
fabric_overlap=0.0,
checkpoint_adjacency=0.0,
achieved=0.0,
)
self.set_raw_rate_integral(probe, 0.0)
probe["normalized_signals"]["storage_operation_overlap_fraction"] = 0.80
probe["normalized_signals"]["bytes_explained_fraction"] = 0.70
result = evaluate_site(probe)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("negative_screen_incoherence_conflict", result["stage_outputs"]["C_discrepancy_and_explanation_review"]["labels"])
self.assertIn("fabric_activity_alignment_incoherence_conflict", result["discrepancy_findings"])
self.assertIn("checkpoint_activity_alignment_incoherence_conflict", result["discrepancy_findings"])
def test_fatal1_zero_count_does_not_capacity_ruleout_with_witnesses_standing(self) -> None:
probe = deepcopy(self.sites["A_clean_threshold_training"])
for record in probe["raw_features"].get("accelerator_count_by_family_sku", []):
record["count"] = 0.0
result = evaluate_site(probe)
a = result["stage_outputs"]["A_capacity_gate"]
self.assertEqual("capacity_unknown_due_to_missing_inputs", a["label"])
self.assertFalse(a["short_circuited"])
self.assertIn("accelerator_count_by_family_sku", a["missing_inputs"])
self.assertNotEqual("capacity_ruled_out_for_scope", result["final_route"])
def test_absent_serving_evidence_does_not_count_as_nonserving_identity(self) -> None:
probe = deepcopy(self.sites["A_clean_threshold_training"])
sig = probe["normalized_signals"]
sig["checkpoint_periodicity_score"] = 0.0
sig["checkpoint_burst_count"] = 0.0
sig["checkpoint_activity_adjacency_fraction"] = 0.0
sig.pop("serving_counterevidence_score", None)
sig.pop("non_serving_score", None)
result = evaluate_site(probe)
b = result["stage_outputs"]["B_training_candidate_detection"]
self.assertEqual("medium_training_like_warning", result["final_route"])
self.assertEqual(1, b["identity_category_count"])
self.assertIn("distributed_training_like_candidate", b["labels"])
self.assertNotIn("checkpoint_training_like_candidate", b["labels"])
def test_invalid_audit_window_does_not_capacity_ruleout(self) -> None:
probe = deepcopy(self.sites["C_capacity_ruled_out"])
probe["audit_window"] = {
"start": "2026-05-01T00:00:00Z",
"end": "2026-04-01T00:00:00Z",
}
result = evaluate_site(probe)
a = result["stage_outputs"]["A_capacity_gate"]
c = result["stage_outputs"]["C_discrepancy_and_explanation_review"]
self.assertEqual("capacity_unknown_due_to_missing_inputs", a["label"])
self.assertFalse(a["short_circuited"])
self.assertIn("audit_window", a["missing_inputs"])
self.assertEqual("inconclusive_due_to_missingness", result["final_route"])
self.assertIn("audit_window", c["missing_channels"])
def test_unknown_unit_raw_rate_does_not_create_large_compute_candidate(self) -> None:
probe = deepcopy(self.sites["A_clean_threshold_training"])
self.suppress_identity_pathways(probe)
probe["normalized_signals"].pop("achieved_operations", None)
duration = 30 * 24 * 3600
probe["raw_features"]["generic_achieved_operation_rate"] = [
{
"sample_time": "2026-04-16T00:00:00Z",
"operation_rate": 2.0e25 / duration,
"operation_unit": "unknown_vendor_units",
"counter_scope": "accelerator_pool",
}
]
result = evaluate_site(probe)
achieved = result["derived_signals"]["achieved_operation_integral"]
b = result["stage_outputs"]["B_training_candidate_detection"]
self.assertEqual(0.0, achieved["operation_count"])
self.assertFalse(achieved["unit_normalized"])
self.assertTrue(achieved["ignored_raw_rate_unit"])
self.assertNotIn("large_compute_candidate", b["labels"])
self.assertNotIn(result["final_route"], {"weak_training_like_candidate", "medium_training_like_warning", "high_training_like_warning"})
def test_negative_capacity_adjustment_factor_does_not_capacity_ruleout(self) -> None:
probe = deepcopy(self.sites["C_capacity_ruled_out"])
probe["normalized_signals"]["capacity_adjustment_factor"] = -1.0
result = evaluate_site(probe)
a = result["stage_outputs"]["A_capacity_gate"]
cap = result["derived_signals"]["capacity_upper_bound_flop"]
self.assertTrue(cap["invalid_capacity_adjustment_factor"])
self.assertEqual("capacity_unknown_due_to_missing_inputs", a["label"])
self.assertFalse(a["short_circuited"])
self.assertIn("capacity_adjustment_factor", a["missing_inputs"])
self.assertNotEqual("capacity_ruled_out_for_scope", result["final_route"])
def test_fatal1_raw_threshold_rate_blocks_covered_negative(self) -> None:
probe = deepcopy(self.sites["A_clean_threshold_training"])
probe["normalized_signals"]["achieved_operations"] = 0.0
probe["normalized_signals"]["activity_duration_seconds"] = 1799.0
probe["normalized_signals"]["checkpoint_periodicity_score"] = 0.54
result = evaluate_site(probe)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn(
"raw_rate_threshold_compute_without_candidate_conflict",
result["discrepancy_findings"],
)
# --- FIX 1 (M1): achieved_ops in the negative screen ---
def test_negative_screen_blocks_on_missing_achieved_ops(self) -> None:
site = deepcopy(self.sites["B_covered_negative"])
site["coverage"]["achieved_ops"] = 0.0
result = evaluate_site(site)
self.assertEqual("inconclusive_due_to_missingness", result["final_route"])
self.assertIn("achieved_ops", result["missing_channels"])
# --- FIX 2 (M3): polarity-correct certification coverage ---
def test_omitted_certification_key_does_not_increase_confidence(self) -> None:
site = deepcopy(self.sites["B_covered_negative"])
del site["coverage"]["achieved_ops"]
self.assertEqual("inconclusive_due_to_missingness", evaluate_site(site)["final_route"])
def test_omitting_attribution_does_not_disable_integrity_guard(self) -> None:
site = deepcopy(self.sites["I_activity_attribution_conflict"])
del site["coverage"]["attribution"]
self.assertEqual("integrity_review_required", evaluate_site(site)["final_route"])
def test_omitting_capacity_blocks_capacity_ruleout(self) -> None:
site = deepcopy(self.sites["C_capacity_ruled_out"])
del site["coverage"]["capacity"]
result = evaluate_site(site)
self.assertNotEqual("capacity_ruled_out_for_scope", result["final_route"])
self.assertFalse(result["stage_outputs"]["A_capacity_gate"]["short_circuited"])
def test_omitting_achieved_ops_on_integrity_fixture_routes_inconclusive(self) -> None:
site = deepcopy(self.sites["I_activity_attribution_conflict"])
del site["coverage"]["achieved_ops"]
result = evaluate_site(site)
self.assertEqual("inconclusive_due_to_missingness", result["final_route"])
self.assertIn("achieved_ops", result["missing_channels"])
# --- FIX 4 (M-2): per-category serving/storage carve-outs ---
def test_serving_shape_does_not_demote_independent_checkpoint_pathway(self) -> None:
instance = deepcopy(self.sites["A_clean_threshold_training"])
sig = instance["normalized_signals"]
sig["activity_fabric_overlap_fraction"] = 0.49
sig["serving_counterevidence_score"] = 0.75
sig["serving_activity_overlap_fraction"] = 0.80
result = evaluate_site(instance)
self.assertTrue(result["final_route"].endswith("_warning"))
self.assertNotEqual("candidate_explained_or_demoted", result["final_route"])
c = result["stage_outputs"]["C_discrepancy_and_explanation_review"]
self.assertIn("candidate_explained_by_serving", c["labels"])
self.assertTrue(c["surviving_identity_pathway"])
self.assertIn("checkpoint_training_like_candidate", result["stage_outputs"]["B_training_candidate_detection"]["labels"])
def test_storage_relabel_does_not_demote_live_fabric_pathway(self) -> None:
instance = deepcopy(self.sites["A_clean_threshold_training"])
sig = instance["normalized_signals"]
sig["checkpoint_activity_adjacency_fraction"] = 0.49
sig["storage_operation_overlap_fraction"] = 0.85
sig["bytes_explained_fraction"] = 0.80
result = evaluate_site(instance)
self.assertTrue(result["final_route"].endswith("_warning"))
self.assertNotEqual("candidate_explained_or_demoted", result["final_route"])
c = result["stage_outputs"]["C_discrepancy_and_explanation_review"]
self.assertIn("candidate_explained_by_storage_operation", c["labels"])
self.assertTrue(c["surviving_identity_pathway"])
self.assertIn("distributed_training_like_candidate", result["stage_outputs"]["B_training_candidate_detection"]["labels"])
def test_model_parallel_inference_serving_still_suppressed(self) -> None:
result = self.result("F_serving_inference_counterevidence")
c = self.stage("F_serving_inference_counterevidence", "C_discrepancy_and_explanation_review")
self.assertEqual("candidate_explained_or_demoted", result["final_route"])
self.assertIn("candidate_explained_by_serving", c["labels"])
self.assertFalse(c["surviving_identity_pathway"])
# --- FIX 5 (F2): forged capacity rule-out via under-reported accelerator count ---
def _f2_count_shaded_probe(self, count: float = 100.0) -> dict:
probe = deepcopy(self.sites["A_clean_threshold_training"])
raw = probe.get("raw_features", {})
for record in raw.get("accelerator_count_by_family_sku", []):
record["count"] = count
for record in raw.get("allocated_accelerator_count_by_sku", []):
if "count" in record:
record["count"] = count
return probe
def test_f2_count_shaded_ruleout_routes_integrity_without_disturbing_legit_ruleouts(self) -> None:
# (1) F2 PROBE: a genuine positive whose accelerator count is shaded down so the
# capacity upper bound drops below T_sys must NOT certify a capacity rule-out while the
# untouched, well-covered achieved-operation integral still exceeds that shaded bound.
probe = evaluate_site(self._f2_count_shaded_probe())
achieved = probe["derived_signals"]["achieved_operation_integral"]
self.assertGreater(achieved["operation_count_to_capacity_upper_bound_ratio"], 1.0)
self.assertGreaterEqual(achieved["coverage_fraction"], 0.75)
self.assertEqual("integrity_review_required", probe["final_route"])
self.assertNotEqual("capacity_ruled_out_for_scope", probe["final_route"])
self.assertFalse(probe["stage_outputs"]["A_capacity_gate"]["short_circuited"])
self.assertIn("capacity_claim_conflict", probe["discrepancy_findings"])
# (2) LEGITIMATE RULE-OUT: the synthetic clean rule-out fixture must STILL rule out.
# Additional low-ratio and sequential-reuse rule-out locks below replace the old
# external sweep dependency.
fixture = self.result("C_capacity_ruled_out")
self.assertEqual("capacity_ruled_out_for_scope", fixture["final_route"])
self.assertTrue(fixture["stage_outputs"]["A_capacity_gate"]["short_circuited"])
# (3) I2 FIXTURE: the existing capacity_claim_conflict fixture still routes to integrity.
i2 = self.result("H_capacity_claim_conflict")
self.assertEqual("integrity_review_required", i2["final_route"])
self.assertIn(
"capacity_claim_conflict",
self.stage("H_capacity_claim_conflict", "C_discrepancy_and_explanation_review")["labels"],
)
# --- service-power and low-coverage achieved witnesses ---
def _ruleout_forge_base(
self,
count: float = 1000.0,
achieved: float = 1e23,
suppress_identity: bool = False,
) -> dict:
probe = deepcopy(self.sites["A_clean_threshold_training"])
raw = probe.setdefault("raw_features", {})
sig = probe.setdefault("normalized_signals", {})
for record in raw.get("accelerator_count_by_family_sku", []):
record["count"] = count
sig["achieved_operations"] = achieved
sig["participant_count"] = 0
for record in raw.get("fabric_port_device_sample_counters", []):
if record.get("counter_name") == "participant_count":
record["counter_value"] = 0
for key in (
"allocated_accelerator_count_by_sku",
"accelerator_compute_billing_usage_intervals",
"compute_running_intervals",
"scaleout_fabric_domain_graph",
"capacity_reservation_intervals",
"reservation_state_intervals",
"instance_type_shape_machine_type",
"local_accelerator_interconnect_domain",
):
raw.pop(key, None)
for record in raw.get("generic_achieved_operation_rate", []):
record["operation_rate"] = 0.0
if suppress_identity:
self.suppress_identity_pathways(probe)
probe["coverage"]["achieved_ops"] = 0.74
return probe
def test_major2_large_achieved_override_blocks_ruleout_under_shaded_coverage(self) -> None:
probe = self._ruleout_forge_base(achieved=1.15e25)
probe["raw_features"].pop("generic_achieved_operation_rate", None)
result = evaluate_site(probe)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
self.assertIn(
"achieved_operations_exceed_capacity_bound",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
def test_major2_raw_rate_blocks_ruleout_under_shaded_coverage(self) -> None:
probe = self._ruleout_forge_base(achieved=1e23)
duration = 30 * 24 * 3600
probe["raw_features"]["generic_achieved_operation_rate"] = [
{
"sample_time": "2026-04-16T00:00:00Z",
"operation_rate": 1.15e25 / duration,
"operation_unit": "synthetic_normalized_operations",
"counter_scope": "accelerator_pool",
}
]
result = evaluate_site(probe)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn(
"raw_achieved_rate_integral_exceeds_capacity_bound",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
def test_major2_below_bound_low_coverage_ruleout_preserved(self) -> None:
probe = self._ruleout_forge_base(achieved=1e23, suppress_identity=True)
probe["raw_features"].pop("electrical_service_status_intervals", None)
result = evaluate_site(probe)
self.assertEqual("capacity_ruled_out_for_scope", result["final_route"])
self.assertEqual([], result["discrepancy_findings"])
def test_major2_unit_safety_for_achieved_and_raw_rate(self) -> None:
override = self._ruleout_forge_base(achieved=1.15e25, suppress_identity=True)
override["normalized_signals"]["achieved_operations_unit_normalized"] = False
override["raw_features"].pop("generic_achieved_operation_rate", None)
override["raw_features"].pop("electrical_service_status_intervals", None)
self.assertEqual("capacity_ruled_out_for_scope", evaluate_site(override)["final_route"])
raw = self._ruleout_forge_base(achieved=1e23, suppress_identity=True)
duration = 30 * 24 * 3600
raw["raw_features"]["generic_achieved_operation_rate"] = [
{
"sample_time": "2026-04-16T00:00:00Z",
"operation_rate": 1.15e25 / duration,
"operation_unit": "unknown_vendor_units",
"counter_scope": "accelerator_pool",
}
]
raw["raw_features"].pop("electrical_service_status_intervals", None)
self.assertEqual("capacity_ruled_out_for_scope", evaluate_site(raw)["final_route"])
def test_major1_scoped_service_power_floor_blocks_ruleout(self) -> None:
probe = self._ruleout_forge_base(achieved=1e23)
result = evaluate_site(probe)
self.assertEqual("integrity_review_required", result["final_route"])
notes = " ".join(result["stage_outputs"]["A_capacity_gate"]["notes"])
self.assertIn("Independent capacity-scale witnesses", notes)
self.assertNotIn("achieved-operation integral exceeds", notes)
self.assertIn(
"electrical_service_power_floor_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
def test_major1_service_power_floor_requires_scope_and_supporting_status(self) -> None:
for field, value in (("service_class", "unscoped_facility_service"), ("service_status", "pending")):
with self.subTest(field=field):
probe = self._ruleout_forge_base(achieved=1e23, suppress_identity=True)
for record in probe["raw_features"].get("electrical_service_status_intervals", []):
record[field] = value
result = evaluate_site(probe)
self.assertEqual("capacity_ruled_out_for_scope", result["final_route"])
self.assertNotIn(
"electrical_service_power_floor_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
def test_major1_split_concurrent_scoped_service_floor_blocks_ruleout(self) -> None:
probe = self._ruleout_forge_base(count=100.0, achieved=1e23, suppress_identity=True)
proto = probe["raw_features"]["electrical_service_status_intervals"][0]
probe["raw_features"]["electrical_service_status_intervals"] = [
dict(
proto,
service_status="energized",
service_class="synthetic_datacenter_service",
service_capacity_mw=0.072,
power_mw=0.0,
mean_power_mw=0.0,
max_power_mw=0.0,
)
for _ in range(40)
]
result = evaluate_site(probe)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn(
"electrical_service_power_floor_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
def test_major1_sequential_scoped_service_floor_does_not_overcount(self) -> None:
probe = self._ruleout_forge_base(count=100.0, achieved=1e23, suppress_identity=True)
proto = probe["raw_features"]["electrical_service_status_intervals"][0]
probe["raw_features"]["electrical_service_status_intervals"] = [
dict(
proto,
start_time="2026-04-01T00:00:00Z",
end_time="2026-04-16T00:00:00Z",
service_capacity_mw=0.072,
),
dict(
proto,
start_time="2026-04-16T00:00:00Z",
end_time="2026-05-01T00:00:00Z",
service_capacity_mw=0.072,
),
]
result = evaluate_site(probe)
self.assertEqual("capacity_ruled_out_for_scope", result["final_route"])
self.assertNotIn(
"electrical_service_power_floor_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
# --- FIX 6 (F3): generalized capacity-claim contradiction (coordinated forge) ---
def _f3_probe(
self,
count: float = 100.0,
achieved: float = 1e23,
suppress_participant: bool = False,
suppress_allocated: bool = False,
suppress_billing: bool = False,
suppress_running: bool = False,
suppress_fabric_graph: bool = False,
suppress_service: bool = False,
suppress_identity: bool = False,
keep_participant_identity: bool = False,
) -> dict:
# Coordinated forge: shade the accelerator count down AND shade achieved-ops down in
# lockstep so the achieved/capacity ratio stays below 1.0 (defeating the F2 guard).
probe = deepcopy(self.sites["A_clean_threshold_training"])
raw = probe.setdefault("raw_features", {})
signals = probe.setdefault("normalized_signals", {})
for record in raw.get("accelerator_count_by_family_sku", []):
record["count"] = count
signals["achieved_operations"] = achieved
# Shade BOTH the override and the raw achieved-rate witness in lockstep, so the
# F4 raw-rate sub-check is also defeated (the full coordinated shading move).
for record in raw.get("generic_achieved_operation_rate", []):
record["operation_rate"] = 0.0
if suppress_participant:
signals["participant_count"] = 0
for record in raw.get("fabric_port_device_sample_counters", []):
if record.get("counter_name") == "participant_count":
record["counter_value"] = 0
if suppress_allocated:
raw.pop("allocated_accelerator_count_by_sku", None)
if suppress_billing:
raw.pop("accelerator_compute_billing_usage_intervals", None)
# F5: the remaining device-population witnesses the four count-comparison checks
# never read. The running-accelerator interval count and the scale-out fabric
# node/switch counts must also be suppressed for the rule-out to forge.
if suppress_running:
raw.pop("compute_running_intervals", None)
if suppress_fabric_graph:
raw.pop("scaleout_fabric_domain_graph", None)
if suppress_service:
raw.pop("electrical_service_status_intervals", None)
if suppress_identity:
self.suppress_identity_pathways(probe, keep_participant=keep_participant_identity)
return probe
def _f3_full_suppression(self, count: float = 100.0, achieved: float = 1e23) -> dict:
return self._f3_probe(
count=count,
achieved=achieved,
suppress_participant=True,
suppress_allocated=True,
suppress_billing=True,
suppress_running=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
)
def test_f3_coordinated_forge_does_not_rule_out_while_scale_signal_survives(self) -> None:
# The pure F3 maneuver: count shaded AND achieved-ops shaded so the F2 ratio guard is
# below 1.0, yet participant_count / allocated / billed device-seconds still exceed the
# claimed count. Must NOT certify a capacity rule-out; routes to integrity.
base = evaluate_site(self._f3_probe())
ratio = base["derived_signals"]["achieved_operation_integral"][
"operation_count_to_capacity_upper_bound_ratio"
]
self.assertLessEqual(ratio, 1.0) # the F2 ratio guard does NOT fire here
self.assertNotEqual("capacity_ruled_out_for_scope", base["final_route"])
self.assertEqual("integrity_review_required", base["final_route"])
self.assertIn("capacity_claim_conflict", base["discrepancy_findings"])
# Each independent scale signal alone is sufficient to block the forge. Suppress the
# others and confirm the surviving one still routes to integrity.
participant_only = evaluate_site(
self._f3_probe(
suppress_allocated=True,
suppress_billing=True,
suppress_running=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
keep_participant_identity=True,
)
)
self.assertEqual("integrity_review_required", participant_only["final_route"])
allocated_only = evaluate_site(
self._f3_probe(
suppress_participant=True,
suppress_billing=True,
suppress_running=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
)
)
self.assertEqual("integrity_review_required", allocated_only["final_route"])
billing_only = evaluate_site(
self._f3_probe(
suppress_participant=True,
suppress_allocated=True,
suppress_running=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
)
)
self.assertEqual("integrity_review_required", billing_only["final_route"])
def test_f3_irreducible_residual_requires_suppressing_every_scale_signal(self) -> None:
# The forge only succeeds when EVERY enumerated device-population witness is coherently
# suppressed at once: achieved-ops shaded (override + raw rate), participant_count zeroed,
# allocated removed, billing removed, the running-accelerator interval removed, and the
# scale-out fabric graph removed, and strong fabric/checkpoint identity suppressed. With
# no contradicting evidence present, the rule-out is vacuously consistent. This documents
# the disclosed irreducible residual (the true limit of telemetry-grounded verification);
# if a new witness were added the guard would need to cover it and this test would flip.
residual = evaluate_site(self._f3_full_suppression())
self.assertEqual("capacity_ruled_out_for_scope", residual["final_route"])
self.assertEqual([], residual["discrepancy_findings"])
def test_f5_running_accelerator_count_alone_blocks_forge(self) -> None:
# F5 regression: the F3 minimal recipe that suppresses participant/allocated/billing/
# achieved but leaves compute_running_intervals.accelerator_count honest must NOT forge a
# rule-out. The running-accelerator count exceeds the shaded claimed count and routes to
# integrity. (The fabric graph is also suppressed here so the running count is the ONLY
# surviving witness, proving it alone is load-bearing.)
probe = self._f3_probe(
suppress_participant=True,
suppress_allocated=True,
suppress_billing=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
)
result = evaluate_site(probe)
derived = result["derived_signals"]
running = _max(probe["raw_features"].get("compute_running_intervals", []), "accelerator_count")
count = derived["capacity_upper_bound_flop"]["count"]
self.assertGreater(running, count)
self.assertIn(
"running_accelerator_count_exceeds_capacity_count",
_capacity_claim_contradictions(derived, probe),
)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
self.assertFalse(result["stage_outputs"]["A_capacity_gate"]["short_circuited"])
def test_f5_fabric_node_count_alone_blocks_forge(self) -> None:
# F5 regression: with every other witness suppressed (including compute_running), the
# scale-out fabric node/switch counts still exceed the shaded claimed count and block the
# forge. Each fabric node hosts at least one accelerator, so node_count is a physical
# floor on the accelerator population.
probe = self._f3_probe(
suppress_participant=True,
suppress_allocated=True,
suppress_billing=True,
suppress_running=True,
suppress_service=True,
suppress_identity=True,
)
result = evaluate_site(probe)
self.assertIn(
"fabric_node_count_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
self.assertFalse(result["stage_outputs"]["A_capacity_gate"]["short_circuited"])
# --- FIX 7 (F6): additive witnesses aggregated by concurrent semantics, not _max ---
def _f6_split_pool_probe(
self, claimed: float = 100.0, per_record: int = 80, true_pop: int = 3072
) -> dict:
# The F6 forge: shade the accelerator count to a small claim, mask both achieved
# channels, falsify every _sum / counter / fabric witness to a value consistent
# with the small claim, and report the genuinely large running pool HONESTLY as
# many CONCURRENT sub-pool records (identical window), each accelerator_count
# below the claimed count. Under _max the guard read only the largest sub-record
# (per_record) and certified a rule-out; the overlap-aware concurrent peak now
# sums the concurrent sub-pools back to the true population.
probe = deepcopy(self.sites["A_clean_threshold_training"])
raw = probe.setdefault("raw_features", {})
signals = probe.setdefault("normalized_signals", {})
for record in raw.get("accelerator_count_by_family_sku", []):
record["count"] = claimed
signals["achieved_operations"] = 0.0
raw.pop("generic_achieved_operation_rate", None)
half = int(claimed) // 2
for record in raw.get("allocated_accelerator_count_by_sku", []):
record["count"] = half
signals["participant_count"] = half
for record in raw.get("fabric_port_device_sample_counters", []):
if record.get("counter_name") == "participant_count":
record["counter_value"] = half
for record in raw.get("scaleout_fabric_domain_graph", []):
record["node_count"] = half
record["switch_count"] = half
self.suppress_identity_pathways(probe)
proto = raw["compute_running_intervals"][0]
records = []
remaining = true_pop
while remaining > 0:
chunk = min(per_record, remaining)
records.append(dict(proto, accelerator_count=chunk))
remaining -= chunk
raw["compute_running_intervals"] = records
return probe
def test_f6_concurrent_subpool_split_routes_integrity(self) -> None:
# Locking test: the split-pool / concurrent-sub-pool forge must route to integrity.
# Every sub-record is individually below the claimed count (so _max would have read
# 80 < 100 and stayed silent), but the records are concurrent so the true population
# is their sum (3072), far above the claim.
probe = self._f6_split_pool_probe()
derived = evaluate_site(probe)["derived_signals"]
records = probe["raw_features"]["compute_running_intervals"]
count = derived["capacity_upper_bound_flop"]["count"]
# each sub-record is below the claimed count
self.assertTrue(all(r["accelerator_count"] < count for r in records))
# _max would have been silent, the concurrent peak is the true population
self.assertLessEqual(_max(records, "accelerator_count"), count)
self.assertGreater(_concurrent_peak(records, "accelerator_count"), count)
result = evaluate_site(probe)
self.assertIn(
"running_accelerator_count_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
self.assertFalse(result["stage_outputs"]["A_capacity_gate"]["short_circuited"])
def test_f6_multidomain_fabric_split_routes_integrity(self) -> None:
# The multi-domain fabric variant: a genuinely large cluster reported HONESTLY as
# many concurrent fabric domains, each node_count below the claimed count. Domains
# are spatial/concurrent, so node_count sums across them (the old _max read one
# domain and stayed silent).
probe = deepcopy(self.sites["A_clean_threshold_training"])
raw = probe["raw_features"]
signals = probe.setdefault("normalized_signals", {})
for record in raw.get("accelerator_count_by_family_sku", []):
record["count"] = 100.0
signals["achieved_operations"] = 0.0
raw.pop("generic_achieved_operation_rate", None)
signals["participant_count"] = 50
for record in raw.get("fabric_port_device_sample_counters", []):
if record.get("counter_name") == "participant_count":
record["counter_value"] = 50
for record in raw.get("allocated_accelerator_count_by_sku", []):
record["count"] = 50
self.suppress_identity_pathways(probe)
raw.pop("compute_running_intervals", None)
proto = raw["scaleout_fabric_domain_graph"][0]
raw["scaleout_fabric_domain_graph"] = [
dict(proto, node_count=80, switch_count=10, link_count=80) for _ in range(40)
]
result = evaluate_site(probe)
self.assertIn(
"fabric_node_count_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
def test_f6_sequential_reuse_of_small_pool_still_rules_out(self) -> None:
# Locking the overlap-aware semantics from the OTHER side: a genuinely small pool
# reused SEQUENTIALLY (back-to-back, non-overlapping windows) must NOT be summed
# into a false contradiction. The concurrent peak stays at the small per-window
# value, so an honest small-scope rule-out is preserved. A plain _sum would have
# over-counted the sequential reuse and falsely blocked this legitimate rule-out.
probe = deepcopy(self.sites["C_capacity_ruled_out"])
raw = probe["raw_features"]
count = raw["accelerator_count_by_family_sku"][0]["count"]
small = int(count) // 2 or 1
windows = [
("2026-04-01T00:00:00Z", "2026-04-08T00:00:00Z"),
("2026-04-08T00:00:00Z", "2026-04-15T00:00:00Z"),
("2026-04-15T00:00:00Z", "2026-04-22T00:00:00Z"),
("2026-04-22T00:00:00Z", "2026-04-29T00:00:00Z"),
]
raw["compute_running_intervals"] = [
{
"start_time": start,
"end_time": end,
"compute_resource_state": "running",
"accelerator_count": small,
"accelerator_shape_or_sku": "SYN-ACCEL",
}
for start, end in windows
]
records = raw["compute_running_intervals"]
# the naive sum (4 * small) would exceed the count, but the genuine concurrent peak
# is just `small` because the windows do not overlap
self.assertGreater(sum(r["accelerator_count"] for r in records), count)
self.assertLessEqual(_concurrent_peak(records, "accelerator_count"), count)
result = evaluate_site(probe)
self.assertNotIn(
"running_accelerator_count_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], probe),
)
self.assertEqual("capacity_ruled_out_for_scope", result["final_route"])
self.assertEqual([], result["discrepancy_findings"])
def test_f4_peak_or_adjust_deflation_with_raw_rate_witness_routes_integrity(self) -> None:
# F4 variant: deflate the bound through peak_rate / capacity_adjustment_factor (count
# left honest, so participant/allocated/billing do NOT exceed it) and mask the
# achieved_operations override down so the F2 ratio guard is below 1.0. The raw
# generic_achieved_operation_rate witness still integrates above the claimed bound and
# must block the forge.
for channel in ("peak", "adjust"):
with self.subTest(channel=channel):
probe = deepcopy(self.sites["A_clean_threshold_training"])
if channel == "peak":
for record in probe["raw_features"]["advertised_peak_rate_by_precision"]:
record["peak_rate"] *= 0.01
else:
probe["normalized_signals"]["capacity_adjustment_factor"] = 0.01
# mask the override only; leave the raw operation-rate witness intact
probe["normalized_signals"]["achieved_operations"] = 1e23
result = evaluate_site(probe)
ratio = result["derived_signals"]["achieved_operation_integral"][
"operation_count_to_capacity_upper_bound_ratio"
]
self.assertLessEqual(ratio, 1.0) # the F2 override-ratio guard does NOT fire
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
def test_participant_count_exceeding_inventory_routes_integrity(self) -> None:
# A distributed-training participant set larger than the claimed accelerator inventory is
# physically incoherent: it must block the rule-out even with achieved-ops fully shaded.
# Every other device-population witness is suppressed so participant_count is the ONLY
# surviving witness, proving it alone is load-bearing.
probe = self._f3_probe(
achieved=0.0,
suppress_allocated=True,
suppress_billing=True,
suppress_running=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
keep_participant_identity=True,
)
result = evaluate_site(probe)
participant = result["derived_signals"]["collective_cadence_score"]["participant_count"]
count = result["derived_signals"]["capacity_upper_bound_flop"]["count"]
self.assertGreater(participant, count)
self.assertEqual(
["participant_count_exceeds_capacity_count"],
_capacity_claim_contradictions(result["derived_signals"], probe),
)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
def test_raw_participant_count_floor_is_not_masked_by_normalized_zero(self) -> None:
probe = self._f3_probe(
count=100.0,
achieved=0.0,
suppress_allocated=True,
suppress_billing=True,
suppress_running=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
)
probe["normalized_signals"]["participant_count"] = 0.0
for record in probe["raw_features"].get("fabric_port_device_sample_counters", []):
if record.get("counter_name") == "participant_count":
record["counter_value"] = 6144.0
result = evaluate_site(probe)
self.assertEqual(
["participant_count_exceeds_capacity_count"],
_capacity_claim_contradictions(result["derived_signals"], probe),
)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
def test_negative_capacity_count_row_cannot_cancel_positive_inventory(self) -> None:
probe = self._f3_full_suppression(count=100.0, achieved=0.0)
proto = probe["raw_features"]["accelerator_count_by_family_sku"][0]
probe["raw_features"]["accelerator_count_by_family_sku"] = [
dict(proto, count=8192.0),
dict(proto, count=-8092.0),
]
result = evaluate_site(probe)
a = result["stage_outputs"]["A_capacity_gate"]
self.assertEqual("capacity_unknown_due_to_missing_inputs", a["label"])
self.assertIn("accelerator_count_by_family_sku", a["missing_inputs"])
self.assertFalse(a["short_circuited"])
self.assertNotEqual("capacity_ruled_out_for_scope", result["final_route"])
def test_allocated_count_exceeding_inventory_routes_integrity(self) -> None:
# Every other device-population witness suppressed so allocated_count is the only one left.
probe = self._f3_probe(
achieved=0.0,
suppress_participant=True,
suppress_billing=True,
suppress_running=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
)
result = evaluate_site(probe)
self.assertEqual(
["allocated_count_exceeds_capacity_count"],
_capacity_claim_contradictions(result["derived_signals"], probe),
)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
def test_negative_rows_do_not_cancel_positive_population_floors(self) -> None:
duration = 30 * 24 * 3600
def base() -> dict:
probe = self._f3_full_suppression(count=100.0, achieved=0.0)
raw = probe["raw_features"]
for key in ("instance_type_shape_machine_type", "local_accelerator_interconnect_domain"):
raw.pop(key, None)
return probe
cases = []
allocated = base()
allocated["raw_features"]["allocated_accelerator_count_by_sku"] = [
{"accelerator_sku": "SYN", "count": 8192.0},
{"accelerator_sku": "SYN", "count": -8092.0},
]
cases.append(("allocated", allocated, "allocated_count_exceeds_capacity_count"))
instance_shape = base()
instance_shape["raw_features"]["instance_type_shape_machine_type"] = [
{"machine_type": "synthetic-large", "accelerator_count": 8192.0},
{"machine_type": "synthetic-large", "accelerator_count": -8092.0},
]
cases.append(("instance_shape", instance_shape, "instance_shape_accelerator_count_exceeds_capacity_count"))
local_fabric = base()
local_fabric["raw_features"]["local_accelerator_interconnect_domain"] = [
{"fabric_domain_id": "domain-a", "local_fabric_device_count": 8192.0},
{"fabric_domain_id": "domain-b", "local_fabric_device_count": -8092.0},
]
cases.append(("local_fabric", local_fabric, "local_fabric_device_count_exceeds_capacity_count"))
fabric_graph = base()
fabric_graph["raw_features"]["scaleout_fabric_domain_graph"] = [
{"fabric_domain_id": "domain-a", "node_count": 8192.0, "switch_count": 8192.0, "link_count": 0.0},
{"fabric_domain_id": "domain-b", "node_count": -8092.0, "switch_count": -8092.0, "link_count": 0.0},
]
cases.append(("fabric_graph", fabric_graph, "fabric_node_count_exceeds_capacity_count"))
cases.append(("fabric_graph", fabric_graph, "fabric_switch_count_exceeds_capacity_count"))
billing = base()
billing["raw_features"]["accelerator_compute_billing_usage_intervals"] = [
{"usage_unit": "accelerator_seconds", "usage_quantity": 8192.0 * duration},
{"usage_unit": "accelerator_seconds", "usage_quantity": -8092.0 * duration},
]
cases.append(("billing", billing, "billing_device_hours_exceed_capacity_count"))
for label, probe, expected in cases:
with self.subTest(label=label, expected=expected):
result = evaluate_site(probe)
contradictions = _capacity_claim_contradictions(result["derived_signals"], probe)
self.assertIn(expected, contradictions)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn("capacity_claim_conflict", result["discrepancy_findings"])
def test_offset_timestamps_are_parsed_for_concurrent_interval_floors(self) -> None:
start = "2026-04-02T00:30:00+14:00"
end = "2026-04-01T23:30:00-12:00"
running = self._f3_probe(
count=100.0,
achieved=0.0,
suppress_participant=True,
suppress_allocated=True,
suppress_billing=True,
suppress_fabric_graph=True,
suppress_service=True,
suppress_identity=True,
)
running["raw_features"]["compute_running_intervals"] = [
{
"compute_resource_state": "running",
"accelerator_shape_or_sku": "SYN-ACCEL",
"accelerator_count": 80.0,
"start_time": start,
"end_time": end,
}
for _ in range(700)
]
result = evaluate_site(running)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn(
"running_accelerator_count_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], running),
)
service = self._ruleout_forge_base(count=100.0, achieved=1e23, suppress_identity=True)
proto = service["raw_features"]["electrical_service_status_intervals"][0]
service["raw_features"]["electrical_service_status_intervals"] = [
dict(
proto,
service_status="energized",
service_class="synthetic_datacenter_service",
service_capacity_mw=0.072,
power_mw=0.0,
mean_power_mw=0.0,
max_power_mw=0.0,
start_time=start,
end_time=end,
)
for _ in range(700)
]
result = evaluate_site(service)
self.assertEqual("integrity_review_required", result["final_route"])
self.assertIn(
"electrical_service_power_floor_exceeds_capacity_count",
_capacity_claim_contradictions(result["derived_signals"], service),
)
def test_provider_account_style_honest_ruleout_with_no_scale_signals_still_rules_out(self) -> None:
# An honest small-scope rule-out (the provider-account analogue) reports NONE of the scale
# signals the guard checks, so no contradiction is present and the rule-out is preserved.
site = {
"site_id": "provider_account_style_honest_slice",
"scenario_key": "provider_account_style_honest_slice",
"scenario_name": "Honest small-scope rule-out with no participant/allocated/billing signals",
"scope": "account_slice/accelerator_pool",
"audit_window": {"start": "2026-06-04T00:00:00Z", "end": "2026-06-10T00:00:00Z"},
"raw_features": {
"accelerator_count_by_family_sku": [{"accelerator_sku": "SYN", "count": 370}],
"advertised_peak_rate_by_precision": [{"peak_rate": 2.0e15}],
"accelerator_busy_or_utilization_fraction": [{"value": 0.45}],
},
"coverage": {
"capacity": 0.96,
"activity": 0.9,
"achieved_ops": 0.0,
"fabric": 0.0,
"storage": 0.0,
"scope_mapping": 0.99,
"clock_alignment": 0.93,
},
"normalized_signals": {
"hidden_or_unmonitored_capacity_possible": False,
"capacity_unit_normalized": True,
"activity_score": 0.45,
"achieved_operations": 3.0e23,
"achieved_operations_unit_normalized": True,
},
}
result = evaluate_site(site)
cap = result["derived_signals"]["capacity_upper_bound_flop"]["capacity_upper_bound_operations"]
self.assertLess(cap, 1e25)
self.assertEqual("capacity_ruled_out_for_scope", result["final_route"])
self.assertEqual([], result["discrepancy_findings"])
if __name__ == "__main__":
unittest.main()
|