File size: 44,285 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 | schema_version: "0.1"
created: "2026-05-16"
source_inventory: observables/observables.yaml
status: sparse_first_algorithm_rule_spec
purpose: >
Sparse support, counterevidence, and explanation rules for cases where only a
small number of non-core raw features are available. These rules can route
review, modestly adjust confidence, or explain candidate activity, but they do
not create a positive training label by themselves.
scope_policy:
trusted_local_inputs: "Only raw feature IDs from observables/observables.yaml are used as inputs."
relationship_to_aggregation_rules: >
This file emits weak support, counterevidence, and explanation screens for
the public aggregation layer. Support screens route, strengthen, or explain
candidates but do not create training identity by themselves.
modest_effect_policy: >
Single-feature support has score_effect or routing_effect capped to weak or
review-only. Pair rules can have larger but still non-primary effects when
time overlap and coverage are established.
support_categories:
- id: physical_activity_support
name: Power, energy, thermal, and cooling support
algorithm_treatment: "Use as physical consistency support or physical anomaly routing, not workload identity."
rule_ids:
- physical_power_load_support_screen
- cooling_thermal_response_support_screen
- physical_activity_pair_support_screen
- id: large_model_context_support
name: Large-model memory context
algorithm_treatment: "Use to strengthen large-model context only after core activity exists."
rule_ids:
- memory_residency_context_screen
- memory_capacity_residency_pair_screen
- id: data_lifecycle_support
name: Data staging, cache, and copy lifecycle
algorithm_treatment: "Use as weak dataset or artifact movement context; check explicit storage-operation explanations first."
rule_ids:
- storage_staging_support_screen
- object_copy_data_movement_support_screen
- cache_epoch_context_screen
- id: topology_context_support
name: Topology and distributed-work context
algorithm_treatment: "Use as distributed-work plausibility context, not fabric activity."
rule_ids:
- topology_distributed_work_support_screen
- id: serving_counterevidence
name: Serving-like counterevidence
algorithm_treatment: "Use to weaken non-serving or training interpretations when aligned with activity."
rule_ids:
- serving_like_network_counterevidence_screen
- id: storage_operation_explanation
name: Storage operation explanations
algorithm_treatment: "Use to explain storage, WAN, or internal-flow bursts before treating them as checkpoint or collective evidence."
rule_ids:
- storage_operation_explanation_screen
- id: benchmark_hpc_alternative
name: Benchmark and HPC alternatives
algorithm_treatment: "Use to route fabric-heavy candidates to benchmark/HPC review before promotion."
rule_ids:
- benchmark_like_alternative_screen
- hpc_mpi_alternative_screen
- id: lifecycle_operational_context
name: Lifecycle, preemption, and maintenance context
algorithm_treatment: "Use to explain starts, restarts, gaps, fragmented activity, and checkpoint bursts."
rule_ids:
- lifecycle_restart_or_preemption_context_screen
rule_category_membership:
membership_policy: >
Categories are evidentiary roles. Each sparse support rule has exactly one
primary category for default routing and may list additional categories for
scoring or suppressor checks.
rules:
- rule_id: physical_power_load_support_screen
primary_category: physical_activity_support
additional_categories: []
- rule_id: cooling_thermal_response_support_screen
primary_category: physical_activity_support
additional_categories: []
- rule_id: physical_activity_pair_support_screen
primary_category: physical_activity_support
additional_categories: []
- rule_id: memory_residency_context_screen
primary_category: large_model_context_support
additional_categories: []
- rule_id: memory_capacity_residency_pair_screen
primary_category: large_model_context_support
additional_categories: []
- rule_id: storage_staging_support_screen
primary_category: data_lifecycle_support
additional_categories: []
- rule_id: object_copy_data_movement_support_screen
primary_category: data_lifecycle_support
additional_categories: [storage_operation_explanation]
- rule_id: cache_epoch_context_screen
primary_category: data_lifecycle_support
additional_categories: []
- rule_id: topology_distributed_work_support_screen
primary_category: topology_context_support
additional_categories: []
- rule_id: serving_like_network_counterevidence_screen
primary_category: serving_counterevidence
additional_categories: []
- rule_id: storage_operation_explanation_screen
primary_category: storage_operation_explanation
additional_categories: [data_lifecycle_support]
- rule_id: benchmark_like_alternative_screen
primary_category: benchmark_hpc_alternative
additional_categories: []
- rule_id: hpc_mpi_alternative_screen
primary_category: benchmark_hpc_alternative
additional_categories: [topology_context_support]
- rule_id: lifecycle_restart_or_preemption_context_screen
primary_category: lifecycle_operational_context
additional_categories: [storage_operation_explanation]
training_support_rules:
- id: physical_power_load_support_screen
purpose: sparse_physical_activity_support
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [accelerator_device_power_energy]
- [server_system_input_power]
- [rack_pdu_it_power]
- [facility_or_room_it_power_rollup]
emitted_statement:
label: physical_activity_support_screen
category: physical_activity_support
max_confidence: screen
statement: "Power or energy telemetry shows material physical load in the monitored scope; workload identity is unresolved."
trigger:
type: one_feature_physical_load_logic
evaluation_window: physical_power_sample_window
conditions:
all:
- metric: baseline_subtracted_physical_load_score
op: ">="
value_ref: min_sparse_physical_load_score
- metric: physical_power_coverage_fraction
op: ">="
value_ref: min_sparse_physical_channel_coverage
calibration_defaults:
- name: min_sparse_physical_load_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S07, S08, S09, S40, S41]
rationale: "A moderate normalized, baseline-subtracted physical load is useful sparse support, but it is not workload-specific."
counterevidence_or_failure_modes:
- "CPU, storage, network, mixed-tenant load, burn-in, or cooling faults can raise power"
calibration_note: "Replace with rack/device/facility baselines and known workload power distributions."
- name: min_sparse_physical_channel_coverage
value: 0.60
unit: fraction
evidence_status: source_informed
source_refs: [S07, S08, S09, S40, S41]
rationale: "Sparse support can emit at lower coverage than strong discrepancy rules but must still report coverage."
counterevidence_or_failure_modes:
- "BMS/DCIM sampling can be sparse or scope-mismatched"
calibration_note: "Replace with source-specific coverage and sampling intervals."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.35
explanation_priority: low
routing_effect:
route_to: physical_consistency_join
missing_joins_to_report: [accelerator_activity, allocation_or_running, cooling_response]
false_positive_or_limitations:
- "Power alone cannot classify workload type."
- "Scope mapping determines whether power belongs to accelerator-bearing assets."
cannot_conclude: [training, accelerator_activity, achieved_compute]
source_refs: [S07, S08, S09, S40, S41]
- id: cooling_thermal_response_support_screen
purpose: sparse_thermal_or_cooling_support
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [accelerator_temperature_celsius]
- [inlet_outlet_air_or_coolant_temperature]
- [coolant_flow_rate]
- [coolant_pressure]
- [pump_speed_or_state]
- [valve_position_or_state]
- [cooling_alarm_state]
- [fan_speed_rpm_or_percent]
- [airflow_rate]
- [thermal_control_mode]
- [source_reported_airflow_state]
- [outdoor_air_temperature_celsius]
- [outdoor_humidity_percent]
- [hvac_control_mode]
- [economizer_state]
- [facility_temperature_setpoint_celsius]
emitted_statement:
label: thermal_or_cooling_context_screen
category: physical_activity_support
max_confidence: screen
statement: "Thermal or cooling telemetry provides physical context for load, cooling response, or alarms; it does not identify training."
trigger:
type: one_feature_cooling_thermal_logic
evaluation_window: thermal_or_cooling_sample_window
conditions:
any:
- metric: thermal_response_score
op: ">="
value_ref: min_thermal_response_score
- metric: cooling_alarm_active
op: "=="
value: true
- metric: cooling_control_change_score
op: ">="
value_ref: min_cooling_control_change_score
calibration_defaults:
- name: min_thermal_response_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S40, S41]
rationale: "A normalized temperature, flow, fan, or cooling response can support physical-load review."
counterevidence_or_failure_modes:
- "ambient conditions, economizer state, fixed fan policy, or unrelated heat load can drive response"
calibration_note: "Calibrate by cooling design, ambient conditions, and telemetry lag."
- name: min_cooling_control_change_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S40, S41]
rationale: "Control changes are weak context until joined to power or activity."
counterevidence_or_failure_modes:
- "control changes may be scheduled or ambient-driven"
calibration_note: "Replace with facility control-mode and setpoint change baselines."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.25
explanation_priority: low
routing_effect:
route_to: physical_context_review
missing_joins_to_report: [power, accelerator_activity, maintenance]
false_positive_or_limitations:
- "Cooling response can lag or precede IT load depending on control mode."
cannot_conclude: [training, accelerator_activity]
source_refs: [S40, S41]
- id: physical_activity_pair_support_screen
purpose: sparse_physical_activity_pair_support
raw_feature_budget: 2
required_feature_sets:
any_one_of:
- [accelerator_device_power_energy, tensor_matrix_mxu_neuron_or_engine_active_fraction]
- [accelerator_device_power_energy, accelerator_busy_or_utilization_fraction]
- [rack_pdu_it_power, tensor_matrix_mxu_neuron_or_engine_active_fraction]
- [rack_pdu_it_power, accelerator_busy_or_utilization_fraction]
- [server_system_input_power, tensor_matrix_mxu_neuron_or_engine_active_fraction]
- [facility_or_room_it_power_rollup, tensor_matrix_mxu_neuron_or_engine_active_fraction]
emitted_statement:
label: physical_activity_support_screen
category: physical_activity_support
max_confidence: weak
statement: "Power or energy load overlaps accelerator activity, supporting that substantial accelerator-bearing infrastructure work occurred."
trigger:
type: two_feature_physical_activity_overlap_logic
evaluation_window: lag_adjusted_power_activity_window
conditions:
all:
- metric: baseline_subtracted_physical_load_score
op: ">="
value_ref: min_pair_physical_load_score
- metric: sparse_activity_score
op: ">="
value_ref: min_pair_activity_score
- metric: lag_adjusted_power_activity_overlap_fraction
op: ">="
value_ref: min_pair_power_activity_overlap
calibration_defaults:
- name: min_pair_physical_load_score
value: 0.55
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S07, S08, S09, S40, S41]
rationale: "Pair support should require slightly stronger physical load than a one-feature screen."
counterevidence_or_failure_modes:
- "mixed-tenant and non-accelerator loads can still overlap activity"
calibration_note: "Calibrate by mapped rack/device power baselines."
- name: min_pair_activity_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S07, S08, S36, S37, S38, S39]
rationale: "Physical support should overlap material accelerator activity."
counterevidence_or_failure_modes:
- "serving, inference, HPC, benchmarks, and burn-in can pass"
calibration_note: "Replace with local activity-score distributions."
- name: min_pair_power_activity_overlap
value: 0.50
unit: fraction
evidence_status: calibration_default
source_refs: [S07, S08, S09, S40, S41]
rationale: "Thermal and power signals can lag activity; half-overlap is an executable sparse default."
counterevidence_or_failure_modes:
- "coarse timestamps can create false overlap"
calibration_note: "Tune lag and overlap by source sampling and cooling design."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.45
explanation_priority: low
routing_effect:
route_to: join_with_identity_shape
missing_joins_to_report: [fabric_identity_shape, checkpoint_identity_shape, serving_counterevidence]
false_positive_or_limitations:
- "Aligned physical load and activity still do not distinguish training from other accelerator work."
cannot_conclude: [training, model_development]
source_refs: [S07, S08, S09, S36, S37, S38, S39, S40, S41]
- id: memory_residency_context_screen
purpose: sparse_large_model_memory_context
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [hbm_or_device_memory_used]
emitted_statement:
label: large_model_memory_context_screen
category: large_model_context_support
max_confidence: screen
statement: "High device-memory use is large-model context, but can also be serving, cache, allocation, or idle residency."
trigger:
type: one_feature_memory_residency_logic
evaluation_window: memory_used_sample_window
conditions:
all:
- metric: hbm_or_device_memory_used_bytes
op: ">="
value_ref: min_sparse_memory_used_bytes
- metric: memory_value_type
op: in
value_ref: supportive_memory_value_types
calibration_defaults:
- name: min_sparse_memory_used_bytes
value: 40000000000
unit: bytes
evidence_status: calibration_default
source_refs: [S04, S05, S06, S07, S08, S09]
rationale: "A 40 GB sparse screen catches materially large accelerator memory residency while remaining below current high-end capacity."
counterevidence_or_failure_modes:
- "large inference, embeddings, caching, and idle loaded models can pass"
- "adapter or sharded training can be below this threshold"
calibration_note: "Replace with local memory-use distributions by workload and accelerator."
- name: supportive_memory_value_types
value: [used, allocated, high_water, resident]
unit: categorical_set
evidence_status: source_informed
source_refs: [S07, S08, S09]
rationale: "Used, allocated, resident, or high-water values can support memory-residency context."
counterevidence_or_failure_modes:
- "free-memory values require inversion and capacity context"
calibration_note: "Normalize telemetry source memory value semantics."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.30
explanation_priority: low
routing_effect:
route_to: large_model_context_join
missing_joins_to_report: [memory_capacity_bytes, accelerator_activity, serving_counterevidence]
false_positive_or_limitations:
- "Memory residency alone is not training evidence."
cannot_conclude: [training, activity, model_size]
source_refs: [S04, S05, S06, S07, S08, S09]
- id: memory_capacity_residency_pair_screen
purpose: sparse_large_model_memory_pair_support
raw_feature_budget: 2
required_feature_sets:
any_one_of:
- [hbm_or_device_memory_used, memory_capacity_bytes]
emitted_statement:
label: large_model_memory_context_screen
category: large_model_context_support
max_confidence: weak
statement: "Device memory residency occupies a material fraction of memory capacity, supporting large-model context if joined with activity evidence."
trigger:
type: two_feature_memory_fraction_logic
evaluation_window: memory_used_capacity_overlap_window
conditions:
all:
- metric: memory_residency_fraction
op: ">="
value_ref: min_sparse_memory_residency_fraction
- metric: memory_capacity_overlap_fraction
op: ">="
value_ref: min_memory_capacity_overlap_fraction
calibration_defaults:
- name: min_sparse_memory_residency_fraction
value: 0.60
unit: fraction_of_device_memory
evidence_status: calibration_default
source_refs: [S04, S05, S06, S07, S08, S09]
rationale: "High fractional memory residency is stronger context than raw bytes, but remains workload-ambiguous."
counterevidence_or_failure_modes:
- "serving and cache-heavy workloads can occupy most memory"
- "ZeRO, offload, and adapter training can use less memory"
calibration_note: "Calibrate by model family, sharding, serving, and cache behavior."
- name: min_memory_capacity_overlap_fraction
value: 0.80
unit: fraction
evidence_status: source_informed
source_refs: [S07, S08, S09]
rationale: "The capacity and use samples should cover the same memory window."
counterevidence_or_failure_modes:
- "sample timing mismatch can distort fractions"
calibration_note: "Tune by scrape period and device mapping."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.45
explanation_priority: low
routing_effect:
route_to: join_with_activity_and_checkpoint
missing_joins_to_report: [accelerator_activity, checkpoint_identity_shape, serving_counterevidence]
false_positive_or_limitations:
- "Fractional residency is not a workload identity signal by itself."
cannot_conclude: [training, activity]
source_refs: [S04, S05, S06, S07, S08, S09]
- id: storage_staging_support_screen
purpose: sparse_data_staging_support
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [storage_read_operation_bytes]
- [object_storage_operation_counts]
- [north_south_external_egress]
- [data_center_interconnect_wan_usage]
emitted_statement:
label: data_staging_support_screen
category: data_lifecycle_support
max_confidence: screen
statement: "Large or repeated read, object, egress, or DCI movement can be dataset or artifact staging context, but many non-training workflows share it."
trigger:
type: one_feature_storage_staging_logic
evaluation_window: storage_or_egress_interval
conditions:
all:
- metric: normalized_staging_or_transfer_score
op: ">="
value_ref: min_sparse_staging_score
- metric: storage_or_network_coverage_fraction
op: ">="
value_ref: min_sparse_staging_coverage_fraction
calibration_defaults:
- name: min_sparse_staging_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S28, S29, S31, S32, S33, S34, S35]
rationale: "Use normalized volume relative to local baseline or participant memory rather than raw bytes."
counterevidence_or_failure_modes:
- "backup, restore, ETL, replication, logs, and artifact export can pass"
calibration_note: "Replace with storage and network baselines by service, tenant, and workload class."
- name: min_sparse_staging_coverage_fraction
value: 0.60
unit: fraction
evidence_status: source_informed
source_refs: [S28, S29, S31, S34]
rationale: "Sparse staging support should be retained with moderate coverage but carry explicit caveats."
counterevidence_or_failure_modes:
- "flow and object logs can be sampled, delayed, best-effort, or skipped"
calibration_note: "Replace with per-source logging completeness."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.35
explanation_priority: medium
routing_effect:
route_to: data_lifecycle_join
missing_joins_to_report: [compute_window, storage_operation_intervals, accelerator_activity]
false_positive_or_limitations:
- "Staging support only strengthens training after core activity exists."
cannot_conclude: [training, checkpointing, model_development]
source_refs: [S28, S29, S31, S32, S33, S34, S35]
- id: object_copy_data_movement_support_screen
purpose: sparse_copy_movement_context
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [object_or_file_copy_movement_records]
emitted_statement:
label: data_copy_movement_context_screen
category: data_lifecycle_support
max_confidence: screen
statement: "Object or file copy/movement records provide data lifecycle context and may also explain storage or network bursts."
trigger:
type: one_feature_copy_movement_logic
evaluation_window: copy_movement_interval
conditions:
all:
- metric: bytes_moved
op: ">="
value_ref: min_copy_bytes_to_emit
- metric: route_class_known
op: "=="
value: true
calibration_defaults:
- name: min_copy_bytes_to_emit
value: 0
unit: bytes
evidence_status: source_informed
source_refs: [S31, S32, S33, S34, S35]
rationale: "Even zero or unknown bytes with a source-emitted copy interval can route explanation review; byte materiality is scored separately."
counterevidence_or_failure_modes:
- "metadata-only or failed copy events may not explain traffic"
calibration_note: "Normalize success status, bytes moved, route class, and source coverage."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.25
explanation_priority: medium
routing_effect:
route_to: storage_or_network_explanation_join
missing_joins_to_report: [storage_operation_intervals, flow_or_fabric_bytes, compute_window]
false_positive_or_limitations:
- "Copy movement can be dataset staging, backup, migration, artifact export, or ETL."
cannot_conclude: [training, checkpointing]
source_refs: [S31, S32, S33, S34, S35]
- id: cache_epoch_context_screen
purpose: sparse_cache_epoch_context
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [cache_hit_miss_counters]
emitted_statement:
label: cache_epoch_context_screen
category: data_lifecycle_support
max_confidence: screen
statement: "Cache hit/miss curves provide weak data-reuse or input-bottleneck context; they are highly workload and tenant dependent."
trigger:
type: one_feature_cache_context_logic
evaluation_window: cache_counter_interval
conditions:
any:
- metric: cache_miss_burst_score
op: ">="
value_ref: min_cache_miss_burst_score
- metric: cache_hit_rate_change_score
op: ">="
value_ref: min_cache_hit_rate_change_score
calibration_defaults:
- name: min_cache_miss_burst_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S33, S34, S35]
rationale: "Cache miss bursts can route data-loading review but are not portable across storage and cache systems."
counterevidence_or_failure_modes:
- "serving cold starts, ETL, package installs, and multi-tenant cache churn can pass"
calibration_note: "Replace with cache baselines and known dataset-pass traces."
- name: min_cache_hit_rate_change_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S33, S34, S35]
rationale: "Hit-rate changes can support epoch-like reuse only after joining with compute windows."
counterevidence_or_failure_modes:
- "cache warming and serving can produce similar curves"
calibration_note: "Calibrate by cache policy and workload class."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.20
explanation_priority: low
routing_effect:
route_to: data_reuse_review
missing_joins_to_report: [storage_reads, accelerator_activity, compute_window]
false_positive_or_limitations:
- "Cache signals are weak without compute and storage context."
cannot_conclude: [training, epoch_count]
source_refs: [S33, S34, S35]
- id: topology_distributed_work_support_screen
purpose: sparse_topology_context_support
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [scaleout_fabric_domain_graph]
- [local_accelerator_interconnect_domain]
- [cross_site_or_dci_connectivity]
emitted_statement:
label: topology_distributed_work_support_screen
category: topology_context_support
max_confidence: screen
statement: "Topology capability supports distributed-work plausibility, but observed activity and identity-shape evidence are still required."
trigger:
type: one_feature_topology_context_logic
evaluation_window: topology_validity_or_sample_window
conditions:
all:
- metric: topology_context_score
op: ">="
value_ref: min_sparse_topology_context_score
- metric: topology_coverage_fraction
op: ">="
value_ref: min_sparse_topology_coverage_fraction
calibration_defaults:
- name: min_sparse_topology_context_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S11, S15, S16, S17, S25, S42, S43]
rationale: "A moderate topology score routes distributed-work review without claiming activity."
counterevidence_or_failure_modes:
- "HPC, benchmark, inference, and storage clusters can have similar topology"
calibration_note: "Replace with local topology scores by workload and fabric generation."
- name: min_sparse_topology_coverage_fraction
value: 0.70
unit: fraction
evidence_status: source_informed
source_refs: [S25, S42, S43]
rationale: "Topology capability must cover most of the sparse context window."
counterevidence_or_failure_modes:
- "route changes or link failures can invalidate topology capability"
calibration_note: "Calibrate by topology inventory update cadence."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.35
explanation_priority: medium
routing_effect:
route_to: fabric_activity_join
missing_joins_to_report: [scaleout_port_tx_rx_bytes_packets, accelerator_activity, allocation_or_running]
false_positive_or_limitations:
- "Topology capability alone is not fabric telemetry."
cannot_conclude: [training, collective_activity]
source_refs: [S11, S15, S16, S17, S25, S42, S43]
- id: serving_like_network_counterevidence_screen
purpose: sparse_serving_counterevidence
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [load_balancer_gateway_flow_activity]
- [north_south_external_egress]
- [vpc_or_nsg_flow_records]
emitted_statement:
label: serving_like_counterevidence_screen
category: serving_counterevidence
max_confidence: screen
statement: "Serving-like gateway, external egress, or flow shape is counterevidence for non-serving training interpretations when aligned with accelerator activity."
trigger:
type: one_feature_serving_counterevidence_logic
evaluation_window: serving_network_interval
conditions:
all:
- metric: serving_counterevidence_score
op: ">="
value_ref: min_sparse_serving_counterevidence_score
- metric: serving_channel_coverage_fraction
op: ">="
value_ref: min_sparse_serving_coverage_fraction
calibration_defaults:
- name: min_sparse_serving_counterevidence_score
value: 0.65
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S27, S28, S29, S30, S34]
rationale: "Serving-like network shape should be retained as counterevidence, but one channel should not fully suppress a candidate alone."
counterevidence_or_failure_modes:
- "dataset fetches, artifact export, batch APIs, service mesh, and logging can resemble serving"
calibration_note: "Replace with local serving, batch inference, export, and training network traces."
- name: min_sparse_serving_coverage_fraction
value: 0.60
unit: fraction
evidence_status: source_informed
source_refs: [S27, S28, S29, S30, S34]
rationale: "Presence of serving-like records can be useful at moderate coverage; absence still cannot imply non-serving."
counterevidence_or_failure_modes:
- "flow logs can be sampled, aggregated, skipped, delayed, or disabled"
calibration_note: "Replace with source-specific flow/gateway coverage."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.40
explanation_priority: high
suppresses_confidence_at_or_above: medium_only_when_joined_to_activity
routing_effect:
route_to: serving_false_positive_review
missing_joins_to_report: [accelerator_activity, storage_operation_intervals, compute_window]
false_positive_or_limitations:
- "Serving counterevidence should be time-aligned with activity before weakening a training candidate."
cannot_conclude: [not_training, serving_workload]
source_refs: [S27, S28, S29, S30, S34]
- id: storage_operation_explanation_screen
purpose: sparse_storage_operation_explanation
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [storage_operation_intervals]
emitted_statement:
label: storage_operation_explanation_screen
category: storage_operation_explanation
max_confidence: screen
statement: "Explicit storage backup, restore, rebuild, replication, migration, copy, or ETL intervals can explain storage, WAN, or fabric-like traffic."
trigger:
type: one_feature_storage_operation_explanation_logic
evaluation_window: storage_operation_interval
conditions:
all:
- metric: operation_type
op: in
value_ref: explanatory_storage_operation_types
- metric: bytes_moved
op: ">="
value_ref: min_storage_operation_bytes_to_emit
calibration_defaults:
- name: explanatory_storage_operation_types
value: [backup, restore, rebuild, replication, migration, disaster_recovery, snapshot, object_copy, etl_export]
unit: categorical_set
evidence_status: source_informed
source_refs: [S31, S32, S33, S34, S35]
rationale: "Source-emitted storage operation labels are direct explanation candidates for storage and network movement."
counterevidence_or_failure_modes:
- "operation labels can be coarse and overlap real training"
calibration_note: "Normalize local storage operation taxonomy."
- name: min_storage_operation_bytes_to_emit
value: 0
unit: bytes
evidence_status: source_informed
source_refs: [S31, S32, S33, S34, S35]
rationale: "Even operation intervals without byte fields can route explanation checks; byte materiality is handled when traffic features are present."
counterevidence_or_failure_modes:
- "zero-byte or failed operations may not explain traffic"
calibration_note: "Use source status, bytes moved, and object count where available."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.35
explanation_priority: high
suppresses_confidence_at_or_above: medium_only_when_overlapping_activity
routing_effect:
route_to: storage_explanation_join
missing_joins_to_report: [storage_write_operation_bytes, scaleout_port_tx_rx_bytes_packets, vpc_or_nsg_flow_records]
false_positive_or_limitations:
- "Explicit storage operations explain only overlapping and attributable bytes."
cannot_conclude: [not_training, no_checkpointing]
source_refs: [S31, S32, S33, S34, S35]
- id: benchmark_like_alternative_screen
purpose: sparse_benchmark_like_counterevidence
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [scaleout_port_tx_rx_bytes_packets]
- [local_interconnect_tx_rx_bytes]
- [fabric_port_device_sample_counters]
emitted_statement:
label: benchmark_like_alternative_screen
category: benchmark_hpc_alternative
max_confidence: screen
statement: "Highly regular fabric bursts can be benchmark-like or burn-in-like and should suppress overpromotion until lifecycle and checkpoint evidence are checked."
trigger:
type: one_feature_benchmark_like_fabric_logic
evaluation_window: fabric_burst_interval
conditions:
all:
- metric: fabric_burst_regularity_score
op: ">="
value_ref: min_sparse_benchmark_regularity_score
- metric: fabric_burst_duration_seconds
op: "<="
value_ref: max_sparse_benchmark_duration_seconds
calibration_defaults:
- name: min_sparse_benchmark_regularity_score
value: 0.85
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S11, S15, S16, S17, S25, S42, S43]
rationale: "Synthetic communication tests are often very regular; exact threshold is not portable."
counterevidence_or_failure_modes:
- "some training steps are regular and some benchmarks are long or irregular"
calibration_note: "Calibrate with local NCCL/MPI benchmark, burn-in, and training traces."
- name: max_sparse_benchmark_duration_seconds
value: 7200
unit: seconds
evidence_status: calibration_default
source_refs: [S11, S15, S16, S17, S25]
rationale: "Short high-regularity fabric windows are useful benchmark-like screens."
counterevidence_or_failure_modes:
- "short fine-tunes and long benchmarks exist"
calibration_note: "Replace with local benchmark and job-duration distributions."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.40
explanation_priority: high
suppresses_confidence_at_or_above: medium_only_when_joined_to_candidate
routing_effect:
route_to: benchmark_or_burnin_review
missing_joins_to_report: [accelerator_activity, checkpoint_identity_shape, allocation_or_running]
false_positive_or_limitations:
- "Fabric regularity alone cannot identify benchmark or training."
cannot_conclude: [not_training, benchmark]
source_refs: [S11, S15, S16, S17, S25, S42, S43]
- id: hpc_mpi_alternative_screen
purpose: sparse_hpc_mpi_alternative
raw_feature_budget: 2
required_feature_sets:
any_one_of:
- [scaleout_port_tx_rx_bytes_packets, tensor_matrix_mxu_neuron_or_engine_active_fraction]
- [scaleout_port_tx_rx_bytes_packets, generic_achieved_operation_rate]
- [scaleout_port_tx_rx_bytes_packets, accelerator_busy_or_utilization_fraction]
- [local_interconnect_tx_rx_bytes, tensor_matrix_mxu_neuron_or_engine_active_fraction]
- [fabric_port_device_sample_counters, generic_achieved_operation_rate]
emitted_statement:
label: hpc_mpi_alternative_screen
category: benchmark_hpc_alternative
max_confidence: weak
statement: "Fabric cadence and accelerator activity can also match HPC or MPI-style workloads; checkpoint or model-lifecycle support is needed before training promotion."
trigger:
type: two_feature_fabric_activity_hpc_alternative_logic
evaluation_window: fabric_activity_overlap_window
conditions:
all:
- metric: sparse_fabric_cadence_score
op: ">="
value_ref: min_sparse_hpc_fabric_cadence_score
- metric: sparse_activity_score
op: ">="
value_ref: min_sparse_hpc_activity_score
- metric: fabric_activity_overlap_fraction
op: ">="
value_ref: min_sparse_hpc_overlap_fraction
calibration_defaults:
- name: min_sparse_hpc_fabric_cadence_score
value: 0.60
unit: score_0_to_1
evidence_status: mechanism_inferred
source_refs: [S11, S15, S16, S17, S25]
rationale: "HPC/MPI and distributed training can both produce collective-like fabric cadence."
counterevidence_or_failure_modes:
- "without domain/runtime context the alternative remains weak"
calibration_note: "Calibrate with local HPC, MPI, benchmark, and training traces."
- name: min_sparse_hpc_activity_score
value: 0.50
unit: score_0_to_1
evidence_status: calibration_default
source_refs: [S07, S08, S36, S37, S38, S39]
rationale: "HPC alternative should overlap material accelerator activity."
counterevidence_or_failure_modes:
- "communication-only tests can have lower accelerator activity"
calibration_note: "Tune by workload class and accelerator family."
- name: min_sparse_hpc_overlap_fraction
value: 0.50
unit: fraction
evidence_status: calibration_default
source_refs: [S11, S15, S16, S17]
rationale: "Fabric and compute should overlap materially before routing to HPC/MPI alternative review."
counterevidence_or_failure_modes:
- "coarse telemetry can create false overlap"
calibration_note: "Replace with source-specific clock alignment and aggregation behavior."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.45
explanation_priority: medium
suppresses_confidence_at_or_above: medium_only_when_checkpoint_support_missing
routing_effect:
route_to: hpc_or_mpi_review
missing_joins_to_report: [checkpoint_identity_shape, serving_counterevidence, domain_runtime_context]
false_positive_or_limitations:
- "HPC/MPI alternative is a routing screen, not a definitive label."
cannot_conclude: [not_training, hpc]
source_refs: [S07, S08, S11, S15, S16, S17, S25, S36, S37, S38, S39]
- id: lifecycle_restart_or_preemption_context_screen
purpose: sparse_lifecycle_operational_context
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [resource_container_lifecycle_events]
- [allocation_preemption_events]
- [maintenance_operation_intervals]
emitted_statement:
label: lifecycle_restart_or_preemption_context_screen
category: lifecycle_operational_context
max_confidence: screen
statement: "Lifecycle, preemption, or maintenance events can explain starts, restarts, fragmented activity, checkpoint bursts, or telemetry gaps."
trigger:
type: one_feature_lifecycle_context_logic
evaluation_window: lifecycle_or_preemption_event_window
conditions:
any:
- metric: lifecycle_event_count
op: ">="
value_ref: min_lifecycle_event_count_to_emit
- metric: preempted_resource_count
op: ">="
value_ref: min_preempted_resource_count_to_emit
- metric: affected_asset_count
op: ">="
value_ref: min_lifecycle_maintenance_affected_assets
calibration_defaults:
- name: min_lifecycle_event_count_to_emit
value: 1
unit: events
evidence_status: source_backed
source_refs: [S21, S22, S38]
rationale: "Any source-emitted lifecycle event can bound candidate windows or explain fragmented activity."
counterevidence_or_failure_modes:
- "generic lifecycle events may be unrelated to accelerator work"
calibration_note: "Normalize event reasons and container-to-resource mapping."
- name: min_preempted_resource_count_to_emit
value: 1
unit: resources
evidence_status: source_informed
source_refs: [S18, S21, S22]
rationale: "Preemption of any accelerator-relevant resource can explain restarts or checkpoint bursts."
counterevidence_or_failure_modes:
- "preemption records can refer to CPU or non-accelerator resources"
calibration_note: "Map resource_type to accelerator and workload scope."
- name: min_lifecycle_maintenance_affected_assets
value: 1
unit: assets
evidence_status: source_informed
source_refs: [S18, S40, S41]
rationale: "Maintenance affecting one mapped asset can explain gaps or restarts in sparse telemetry."
counterevidence_or_failure_modes:
- "broad maintenance intervals can over-explain unrelated events"
calibration_note: "Join maintenance asset categories to cluster, storage, power, and fabric scopes."
score_effect:
training_core_confidence_delta: 0
support_score_cap: 0.30
explanation_priority: medium
routing_effect:
route_to: lifecycle_explanation_join
missing_joins_to_report: [compute_window, checkpoint_write_window, telemetry_coverage_fraction]
false_positive_or_limitations:
- "Lifecycle context explains timing but does not identify workload type."
cannot_conclude: [training, no_training]
source_refs: [S18, S21, S22, S38, S40, S41]
|