File size: 49,195 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 | schema_version: "0.1"
created: "2026-05-16"
source_inventory: observables/observables.yaml
status: sparse_first_algorithm_rule_spec
purpose: >
Sparse capacity and capability screens for settings where only one or two raw
observable feature families are present. These rules emit narrow capacity,
availability, topology, power-service, or policy-scale feasibility statements.
They do not emit workload identity or achieved-work labels.
scope_policy:
trusted_local_inputs: "Only raw feature IDs from observables/observables.yaml are used as inputs."
relationship_to_aggregation_rules: >
This file supplies sparse capacity facts and screens for the public
aggregation layer. Strong rule-out claims are made by derived capacity
bounds plus `aggregation_rules.yaml`, not by single sparse screens.
sparse_claim_policy: >
A sparse capacity rule may say that an observed feature is compatible with,
limits, or routes review of threshold-scale capacity. It must not say that
work occurred, that training occurred, or that no hidden/out-of-scope
capacity exists.
sparse_label_semantics:
capacity_screen:
meaning: "A sparse feature is relevant to capacity or physical feasibility."
max_confidence: screen
can_emit_training_label_by_itself: false
availability_screen:
meaning: "A sparse state changes whether capacity was available during the scoped window."
max_confidence: weak
can_emit_training_label_by_itself: false
capacity_limit_screen:
meaning: "A sparse feature can limit or reduce a capacity hypothesis, but cannot alone certify impossibility unless coverage and scope are established elsewhere."
max_confidence: weak
can_emit_training_label_by_itself: false
policy_scale_screen:
meaning: "A sparse capacity value is compatible or incompatible with a policy-scale compute hypothesis under optimistic assumptions."
max_confidence: weak
can_emit_training_label_by_itself: false
capacity_categories:
- id: capacity_count_or_shape
name: Accelerator count or cloud shape
algorithm_treatment: "Use as a first-pass capacity and scope screen; join with peak rate, duration, and activity before stronger conclusions."
rule_ids:
- accelerator_count_capacity_screen
- instance_shape_capacity_screen
- id: capacity_peak_rate
name: Advertised peak rate
algorithm_treatment: "Use only for upper-bound or threshold-time math; never as achieved activity."
rule_ids:
- peak_rate_capacity_screen
- id: capacity_memory
name: Memory capacity and bandwidth
algorithm_treatment: "Use as large-model context and memory-capability screening, not training identity."
rule_ids:
- memory_capacity_model_context_screen
- memory_bandwidth_capacity_screen
- id: capacity_topology
name: Local, scale-out, and cross-site topology
algorithm_treatment: "Use to establish whether distributed synchronized work is physically plausible."
rule_ids:
- local_interconnect_capacity_screen
- scaleout_domain_capacity_screen
- cross_site_connectivity_capacity_screen
- id: capacity_availability
name: Partition, reservation, running, health, and maintenance availability
algorithm_treatment: "Use to reduce, bound, or route capacity availability over time."
rule_ids:
- partition_capacity_screen
- health_availability_screen
- reservation_capacity_screen
- quota_capacity_limit_screen
- running_accelerator_hours_screen
- reservation_billing_capacity_screen
- maintenance_availability_screen
- id: capacity_power_service
name: Power, clocks, electrical service, and installation
algorithm_treatment: "Use to bound physical feasibility and cap-adjusted capacity."
rule_ids:
- power_clock_cap_screen
- electrical_service_capacity_screen
- installation_timeline_capacity_screen
- id: capacity_policy_scale
name: Policy-scale possibility screens
algorithm_treatment: "Use as sparse policy-scale triage; derived upper bounds and aggregation gates are required for strong rule-out claims."
rule_ids:
- accelerator_count_capacity_screen
- peak_rate_capacity_screen
- running_accelerator_hours_screen
rule_category_membership:
membership_policy: >
Categories are evidentiary roles. Each sparse capacity rule has exactly one
primary category for default routing and may list additional categories for
scoring or joins.
rules:
- rule_id: accelerator_count_capacity_screen
primary_category: capacity_count_or_shape
additional_categories: [capacity_policy_scale]
- rule_id: instance_shape_capacity_screen
primary_category: capacity_count_or_shape
additional_categories: [capacity_policy_scale]
- rule_id: peak_rate_capacity_screen
primary_category: capacity_peak_rate
additional_categories: [capacity_policy_scale]
- rule_id: memory_capacity_model_context_screen
primary_category: capacity_memory
additional_categories: []
- rule_id: memory_bandwidth_capacity_screen
primary_category: capacity_memory
additional_categories: []
- rule_id: local_interconnect_capacity_screen
primary_category: capacity_topology
additional_categories: []
- rule_id: scaleout_domain_capacity_screen
primary_category: capacity_topology
additional_categories: [capacity_policy_scale]
- rule_id: cross_site_connectivity_capacity_screen
primary_category: capacity_topology
additional_categories: []
- rule_id: partition_capacity_screen
primary_category: capacity_availability
additional_categories: []
- rule_id: power_clock_cap_screen
primary_category: capacity_power_service
additional_categories: [capacity_availability]
- rule_id: health_availability_screen
primary_category: capacity_availability
additional_categories: [capacity_power_service]
- rule_id: reservation_capacity_screen
primary_category: capacity_availability
additional_categories: [capacity_policy_scale]
- rule_id: quota_capacity_limit_screen
primary_category: capacity_availability
additional_categories: [capacity_policy_scale]
- rule_id: running_accelerator_hours_screen
primary_category: capacity_availability
additional_categories: [capacity_policy_scale]
- rule_id: reservation_billing_capacity_screen
primary_category: capacity_availability
additional_categories: []
- rule_id: electrical_service_capacity_screen
primary_category: capacity_power_service
additional_categories: [capacity_policy_scale]
- rule_id: installation_timeline_capacity_screen
primary_category: capacity_power_service
additional_categories: [capacity_policy_scale]
- rule_id: maintenance_availability_screen
primary_category: capacity_availability
additional_categories: [capacity_power_service]
capability_rules:
- id: accelerator_count_capacity_screen
purpose: sparse_capacity_count_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [accelerator_count_by_family_sku]
emitted_statement:
label: accelerator_count_threshold_capacity_screen
category: capacity_count_or_shape
max_confidence: screen
statement: "Observed installed or provisioned accelerator count is a sparse capacity fact; it is compatible with threshold-scale compute only under SKU, duration, precision, availability, and topology assumptions."
trigger:
type: one_feature_capacity_count_logic
evaluation_window: inventory_validity_window
conditions:
all:
- metric: observed_accelerator_count
op: ">="
value_ref: min_count_to_emit_capacity_fact
any:
- metric: observed_accelerator_count
op: ">="
value_ref: min_accelerator_count_policy_scale_screen
- metric: count_only_threshold_time_unknown
op: "=="
value: true
calibration_defaults:
- name: min_count_to_emit_capacity_fact
value: 1
unit: accelerators
evidence_status: source_backed
source_refs: [S07, S09, S19, S20, S24]
rationale: "A directly observed nonzero accelerator count is already useful as a capacity fact."
counterevidence_or_failure_modes:
- "inventory can include failed, partitioned, powered-off, or out-of-scope devices"
calibration_note: "Replace only if the collection policy suppresses small counts."
- name: min_accelerator_count_policy_scale_screen
value: 1024
unit: accelerators
evidence_status: calibration_default
source_refs: [S01, S03, S04, S05, S06, S24]
rationale: "Executable optimistic screen for policy-scale review; true threshold depends on SKU peak rate, window length, precision, and available fraction."
counterevidence_or_failure_modes:
- "slow SKUs, short windows, or heavy partitioning can be insufficient above this count"
- "small counts can still reach the policy threshold over long enough windows"
calibration_note: "Replace with SKU/time-window-specific capacity_upper_bound_flop."
false_positive_or_limitations:
- "Count alone does not show activity."
- "Count alone cannot rule out hidden or out-of-scope capacity."
cannot_conclude: [activity, achieved_compute, training, no_threshold_run]
promotion_policy: "Promote to `capacity_upper_bound_flop` and the capacity gate when peak rate, availability, and time-window features are present."
source_refs: [S01, S03, S04, S05, S06, S07, S09, S19, S20, S24]
- id: instance_shape_capacity_screen
purpose: sparse_cloud_shape_capacity_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [instance_type_shape_machine_type]
emitted_statement:
label: cloud_shape_capacity_screen
category: capacity_count_or_shape
max_confidence: screen
statement: "Cloud instance or machine shape implies an accelerator count/type capacity fact when provider shape mapping is source-backed."
trigger:
type: one_feature_shape_capacity_logic
evaluation_window: shape_validity_window
conditions:
all:
- metric: shape_accelerator_count
op: ">="
value_ref: min_shape_accelerator_count_to_emit
- metric: provider_shape_mapping_confidence
op: ">="
value_ref: min_provider_shape_mapping_confidence
calibration_defaults:
- name: min_shape_accelerator_count_to_emit
value: 1
unit: accelerators
evidence_status: source_backed
source_refs: [S23, S24]
rationale: "A provider shape with at least one accelerator is a capacity input."
counterevidence_or_failure_modes:
- "shape records can be stale or not scoped to a running interval"
calibration_note: "Use provider SKU tables and local shape-normalization logic."
- name: min_provider_shape_mapping_confidence
value: 0.80
unit: score_0_to_1
evidence_status: source_informed
source_refs: [S23, S24, S35]
rationale: "Provider shape mapping should be source-backed before using it as a capacity screen."
counterevidence_or_failure_modes:
- "custom bare-metal or preview shapes can be absent from lookup tables"
calibration_note: "Replace with provider-specific shape lookup coverage."
false_positive_or_limitations:
- "Shape validity does not mean the instance was running or doing work."
- "Cloud shape capacity can be constrained by reservation, quota, placement, or billing scope."
cannot_conclude: [activity, achieved_compute, training]
promotion_policy: "Join with compute_running_intervals or billing intervals for use-over-time facts, and with peak rate for upper bounds."
source_refs: [S23, S24, S35]
- id: peak_rate_capacity_screen
purpose: sparse_peak_rate_upper_bound_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [advertised_peak_rate_by_precision]
emitted_statement:
label: advertised_peak_rate_capacity_screen
category: capacity_peak_rate
max_confidence: screen
statement: "Source-backed advertised peak rate can support optimistic upper-bound math, but it is not achieved activity."
trigger:
type: one_feature_peak_rate_logic
evaluation_window: peak_rate_validity_window
conditions:
all:
- metric: advertised_peak_rate_ops_per_second
op: ">="
value_ref: min_nonzero_peak_rate_ops_per_second
- metric: precision_or_mode_normalized
op: "=="
value: true
compute:
- metric: optimistic_seconds_to_policy_threshold
formula: "policy_threshold_operations / advertised_peak_rate_ops_per_second"
calibration_defaults:
- name: min_nonzero_peak_rate_ops_per_second
value: 1
unit: operations_per_second
evidence_status: source_backed
source_refs: [S04, S05, S06, S24]
rationale: "Any source-backed positive peak rate is usable for upper-bound arithmetic."
counterevidence_or_failure_modes:
- "incorrect precision/mode normalization can inflate capacity"
calibration_note: "Use source-backed SKU and precision/mode mappings."
- name: policy_threshold_operations
value: 1.0e+25
unit: operations
evidence_status: source_backed
source_refs: [S01, S03]
rationale: "Policy review scale for large compute; not a workload identity threshold."
counterevidence_or_failure_modes:
- "the policy threshold does not distinguish training from other large compute"
calibration_note: "Replace only if the governing policy threshold changes."
false_positive_or_limitations:
- "Advertised peak is an upper bound, not measured use."
- "Sustained training efficiency can be far below peak."
cannot_conclude: [activity, achieved_compute, training]
promotion_policy: "Promote to `capacity_upper_bound_flop` and the capacity gate when count and time-window features are present."
source_refs: [S01, S03, S04, S05, S06, S24]
- id: memory_capacity_model_context_screen
purpose: sparse_memory_capacity_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [memory_capacity_bytes]
emitted_statement:
label: large_model_memory_capacity_context_screen
category: capacity_memory
max_confidence: screen
statement: "Device memory capacity is compatible with some large-model residency classes, but memory capacity alone does not indicate training."
trigger:
type: one_feature_memory_capacity_logic
evaluation_window: memory_capacity_sample_window
conditions:
all:
- metric: device_memory_capacity_bytes
op: ">="
value_ref: min_large_model_context_memory_bytes
calibration_defaults:
- name: min_large_model_context_memory_bytes
value: 80000000000
unit: bytes
evidence_status: source_informed
source_refs: [S04, S05, S06]
rationale: "Modern high-end training accelerators expose tens to hundreds of GB of device memory; 80 GB is an executable large-memory context screen."
counterevidence_or_failure_modes:
- "large inference and vector databases can also require high memory"
- "small or adapter fine-tuning can run below this memory class"
calibration_note: "Replace with local model-size, precision, sharding, and accelerator-memory distributions."
false_positive_or_limitations:
- "Memory capacity is hardware context, not workload evidence."
cannot_conclude: [activity, training, model_size]
promotion_policy: "Join with hbm_or_device_memory_used and checkpoint-size evidence for large-model context support."
source_refs: [S04, S05, S06]
- id: memory_bandwidth_capacity_screen
purpose: sparse_memory_bandwidth_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [memory_bandwidth_bytes_per_sec]
emitted_statement:
label: memory_bandwidth_capacity_screen
category: capacity_memory
max_confidence: screen
statement: "Source-reported memory bandwidth is a hardware capability fact that may support memory-bound model-workload plausibility."
trigger:
type: one_feature_memory_bandwidth_logic
evaluation_window: memory_bandwidth_sample_window
conditions:
all:
- metric: memory_bandwidth_bytes_per_sec
op: ">="
value_ref: min_high_bandwidth_context_bytes_per_sec
calibration_defaults:
- name: min_high_bandwidth_context_bytes_per_sec
value: 1000000000000
unit: bytes_per_second
evidence_status: source_informed
source_refs: [S04, S05, S06]
rationale: "A 1 TB/s screen separates high-bandwidth accelerator memory classes from ordinary host memory, while remaining below current top-end published bandwidth."
counterevidence_or_failure_modes:
- "memory bandwidth capability does not imply memory activity"
- "bandwidth needs vary sharply by model architecture and precision"
calibration_note: "Replace with accelerator-family capability lookup and workload memory-intensity baselines."
false_positive_or_limitations:
- "Capability alone does not prove memory-bound activity."
cannot_conclude: [activity, achieved_compute, training]
promotion_policy: "Join with memory_bandwidth_or_dram_active for support, or with SKU capability for upper bounds."
source_refs: [S04, S05, S06]
- id: local_interconnect_capacity_screen
purpose: sparse_local_topology_capacity_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [local_accelerator_interconnect_domain]
emitted_statement:
label: local_dense_interconnect_capacity_screen
category: capacity_topology
max_confidence: screen
statement: "A local accelerator interconnect domain exists and may make intra-node synchronized work plausible."
trigger:
type: one_feature_local_interconnect_logic
evaluation_window: local_interconnect_validity_window
conditions:
all:
- metric: local_fabric_device_count
op: ">="
value_ref: min_local_fabric_device_count
- metric: local_fabric_bandwidth_gbps
op: ">="
value_ref: min_local_fabric_bandwidth_gbps
calibration_defaults:
- name: min_local_fabric_device_count
value: 2
unit: devices
evidence_status: mechanism_inferred
source_refs: [S04, S05, S06, S11]
rationale: "At least two locally connected accelerators are needed for local accelerator collectives or peer movement."
counterevidence_or_failure_modes:
- "single-accelerator training and CPU-hosted communication remain possible outside this screen"
calibration_note: "Replace with hardware topology and local workload requirements."
- name: min_local_fabric_bandwidth_gbps
value: 100
unit: gigabits_per_second
evidence_status: calibration_default
source_refs: [S04, S05, S06]
rationale: "Executable low bar for dense accelerator interconnect screening; exact usefulness depends on topology and workload."
counterevidence_or_failure_modes:
- "PCIe-only or degraded links can still run distributed work slowly"
calibration_note: "Calibrate by fabric type and measured peer bandwidth."
false_positive_or_limitations:
- "Local interconnect supports physical possibility, not activity."
cannot_conclude: [distributed_training, fabric_activity, training]
promotion_policy: "Join with local_interconnect_tx_rx_bytes and activity for fabric evidence."
source_refs: [S04, S05, S06, S11]
- id: scaleout_domain_capacity_screen
purpose: sparse_scaleout_topology_capacity_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [scaleout_fabric_domain_graph]
emitted_statement:
label: scaleout_distributed_work_capacity_screen
category: capacity_topology
max_confidence: screen
statement: "A connected scale-out fabric domain is large enough to make distributed synchronized work physically plausible."
trigger:
type: one_feature_scaleout_domain_logic
evaluation_window: scaleout_fabric_validity_window
conditions:
all:
- metric: scaleout_node_count
op: ">="
value_ref: min_scaleout_node_count
- metric: scaleout_link_count
op: ">="
value_ref: min_scaleout_link_count
- metric: scaleout_link_bandwidth_gbps
op: ">="
value_ref: min_scaleout_link_bandwidth_gbps
calibration_defaults:
- name: min_scaleout_node_count
value: 8
unit: nodes
evidence_status: calibration_default
source_refs: [S11, S15, S16, S17, S25, S42, S43]
rationale: "Eight nodes is a sparse distributed-work screen, not a policy threshold; it catches topology domains large enough for multi-node collectives."
counterevidence_or_failure_modes:
- "smaller domains can train smaller models"
- "large domains can serve HPC, inference, or benchmarks"
calibration_note: "Replace with local topology sizes for training, HPC, inference, and benchmark jobs."
- name: min_scaleout_link_count
value: 8
unit: links
evidence_status: calibration_default
source_refs: [S25, S42, S43]
rationale: "Requires a nontrivial connected fabric surface before emitting scale-out capacity support."
counterevidence_or_failure_modes:
- "link count can be reported differently across fabrics"
calibration_note: "Use graph-derived connected components and fabric-specific link semantics."
- name: min_scaleout_link_bandwidth_gbps
value: 100
unit: gigabits_per_second
evidence_status: source_informed
source_refs: [S25, S42, S43]
rationale: "Low-latency high-throughput fabrics for ML/HPC expose high bandwidth links; 100 Gbps is a conservative executable screen."
counterevidence_or_failure_modes:
- "bandwidth alone does not imply low latency or correct routing"
calibration_note: "Replace with fabric-generation-specific thresholds and bisection bandwidth."
false_positive_or_limitations:
- "HPC, storage, benchmark, and inference clusters can have the same topology."
cannot_conclude: [fabric_activity, collective_cadence, training]
promotion_policy: "Join with topology health and participant mapping in derived capacity and aggregation review."
source_refs: [S11, S15, S16, S17, S25, S42, S43]
- id: cross_site_connectivity_capacity_screen
purpose: sparse_dci_capacity_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [cross_site_or_dci_connectivity]
emitted_statement:
label: cross_site_or_dci_capacity_context_screen
category: capacity_topology
max_confidence: screen
statement: "Cross-site or DCI connectivity can support data movement or multi-site context, but low-latency synchronous training plausibility remains unresolved."
trigger:
type: one_feature_cross_site_connectivity_logic
evaluation_window: cross_site_connectivity_sample_window
conditions:
all:
- metric: link_capacity_gbps
op: ">="
value_ref: min_cross_site_capacity_gbps
- metric: service_state
op: in
value_ref: active_connectivity_states
calibration_defaults:
- name: min_cross_site_capacity_gbps
value: 100
unit: gigabits_per_second
evidence_status: calibration_default
source_refs: [S25, S28, S29, S30]
rationale: "Executable high-capacity DCI screen for large data movement; exact threshold is provider and route specific."
counterevidence_or_failure_modes:
- "cross-site capacity can support replication or backups rather than training"
- "high bandwidth does not imply low enough latency for synchronous training"
calibration_note: "Replace with measured DCI throughput, latency, and route class."
- name: active_connectivity_states
value: [active, available, in_service, up]
unit: categorical_set
evidence_status: source_informed
source_refs: [S25, S28, S29, S30]
rationale: "Only active/up service states should support a connectivity capacity screen."
counterevidence_or_failure_modes:
- "provider state names vary and can lag actual path availability"
calibration_note: "Normalize local service-state labels."
false_positive_or_limitations:
- "Cross-site connectivity more often supports staging, replication, or federation than tight synchronous training."
cannot_conclude: [training, synchronous_training, fabric_activity]
promotion_policy: "Join with WAN usage or storage operation records for data movement support or explanations."
source_refs: [S25, S28, S29, S30]
- id: partition_capacity_screen
purpose: sparse_partition_capacity_limit_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [accelerator_partitioning_intervals]
emitted_statement:
label: accelerator_partition_capacity_screen
category: capacity_availability
max_confidence: weak
statement: "Partitioning or sharing state can reduce single-run capacity and changes how accelerator counts map to usable compute and memory."
trigger:
type: one_feature_partition_logic
evaluation_window: partition_interval
conditions:
any:
- metric: inferred_partition_compute_fraction
op: "<="
value_ref: max_partition_fraction_for_capacity_limit_screen
- metric: sharing_mode
op: in
value_ref: capacity_limiting_sharing_modes
calibration_defaults:
- name: max_partition_fraction_for_capacity_limit_screen
value: 0.50
unit: fraction
evidence_status: calibration_default
source_refs: [S10, S19, S20]
rationale: "Unknown or small partitions should route review because logical accelerator count can overstate full-device training capacity."
counterevidence_or_failure_modes:
- "some logical partitions can still be sufficient for fine-tuning"
- "partition profile names must be mapped per accelerator family"
calibration_note: "Replace with MIG, vGPU, device-plugin, or scheduler partition profile lookup."
- name: capacity_limiting_sharing_modes
value: [shared, time_sliced, mig, vgpu, fractional]
unit: categorical_set
evidence_status: source_informed
source_refs: [S10, S19, S20]
rationale: "Shared or fractional modes are capacity context, especially for single large-run hypotheses."
counterevidence_or_failure_modes:
- "labels vary by source and can mix policy with hardware partitioning"
calibration_note: "Normalize local partition and sharing labels."
false_positive_or_limitations:
- "Partitioning can support many small jobs rather than one large job."
cannot_conclude: [no_activity, no_training, no_threshold_run]
promotion_policy: "Join with count or allocation features in derived capacity and aggregation review."
source_refs: [S10, S19, S20]
- id: power_clock_cap_screen
purpose: sparse_power_clock_capacity_limit_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [accelerator_power_cap_watts]
- [accelerator_core_clock_hz]
- [accelerator_memory_clock_hz]
- [accelerator_p_state]
- [accelerator_throttle_state]
- [accelerator_performance_limit_state]
emitted_statement:
label: power_clock_limit_capacity_screen
category: capacity_power_service
max_confidence: weak
statement: "Power caps, clocks, P-states, throttle states, or performance-limit states can reduce cap-adjusted capacity during overlapping windows."
trigger:
type: one_feature_power_clock_limit_logic
evaluation_window: power_clock_or_limit_sample_window
conditions:
any:
- metric: normalized_clock_or_power_factor
op: "<="
value_ref: max_capacity_factor_for_limit_screen
- metric: active_limit_state
op: in
value_ref: capacity_limiting_states
calibration_defaults:
- name: max_capacity_factor_for_limit_screen
value: 0.80
unit: fraction
evidence_status: calibration_default
source_refs: [S07, S08, S09]
rationale: "A 20 percent or larger cap/clock reduction is material enough for sparse capacity review."
counterevidence_or_failure_modes:
- "some active limit states are transient or non-binding"
- "reference clocks and power limits are SKU-specific"
calibration_note: "Replace with per-SKU cap, clock, and achieved-rate curves."
- name: capacity_limiting_states
value: [active, throttled, limited, degraded, power_limited, thermal_limited]
unit: categorical_set
evidence_status: source_informed
source_refs: [S07, S08, S09]
rationale: "Vendor telemetry exposes power, thermal, throttle, performance-state, and limit-state surfaces that can constrain capacity."
counterevidence_or_failure_modes:
- "state names and semantics vary by vendor, driver, and source"
calibration_note: "Normalize limit-state labels and required duration per source."
false_positive_or_limitations:
- "Limit states explain reduced capacity but do not prove or disprove activity."
cannot_conclude: [training, no_activity, achieved_compute]
promotion_policy: "Join with reference clocks, power, and overlap windows in derived capacity and aggregation review."
source_refs: [S07, S08, S09]
- id: health_availability_screen
purpose: sparse_health_capacity_limit_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [accelerator_health_error_state]
emitted_statement:
label: health_degraded_capacity_availability_screen
category: capacity_availability
max_confidence: weak
statement: "Device health, reset, error, link-error, throttle, or degraded-state counters can reduce available accelerator capacity or explain capacity gaps."
trigger:
type: one_feature_health_availability_logic
evaluation_window: health_sample_window
conditions:
any:
- metric: degraded_state
op: "=="
value: true
- metric: health_error_delta_count
op: ">="
value_ref: min_health_error_delta_count_for_review
- metric: reset_or_throttle_event_delta_count
op: ">="
value_ref: min_reset_or_throttle_delta_count_for_review
calibration_defaults:
- name: min_health_error_delta_count_for_review
value: 1
unit: events
evidence_status: source_informed
source_refs: [S07, S08, S09]
rationale: "Any new source-reported health/error event is useful sparse availability context."
counterevidence_or_failure_modes:
- "transient corrected errors may not affect capacity"
calibration_note: "Replace with source-specific severity and drain-policy thresholds."
- name: min_reset_or_throttle_delta_count_for_review
value: 1
unit: events
evidence_status: source_informed
source_refs: [S07, S08, S09]
rationale: "Reset or throttle events can affect capacity and time alignment even when brief."
counterevidence_or_failure_modes:
- "short events can be operationally benign"
calibration_note: "Calibrate by event type, duration, and affected device fraction."
false_positive_or_limitations:
- "Health errors require source-specific severity interpretation."
cannot_conclude: [training, no_training, achieved_compute]
promotion_policy: "Join with affected asset counts and overlap windows in derived capacity and aggregation review."
source_refs: [S07, S08, S09]
- id: reservation_capacity_screen
purpose: sparse_reservation_capacity_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [reservation_state_intervals]
- [capacity_reservation_intervals]
emitted_statement:
label: reservation_capacity_availability_screen
category: capacity_availability
max_confidence: screen
statement: "Reservation records indicate capacity availability or holding, not observed accelerator activity."
trigger:
type: one_feature_reservation_logic
evaluation_window: reservation_interval
conditions:
all:
- metric: reserved_or_recorded_accelerator_count
op: ">="
value_ref: min_reserved_accelerator_count_to_emit
- metric: reservation_state
op: in
value_ref: active_reservation_states
calibration_defaults:
- name: min_reserved_accelerator_count_to_emit
value: 1
unit: accelerators
evidence_status: source_backed
source_refs: [S23, S24, S26, S35]
rationale: "A nonzero reservation count is a capacity-availability fact."
counterevidence_or_failure_modes:
- "reserved capacity may not be launched or used"
calibration_note: "Map provider and scheduler reservation fields to accelerator counts."
- name: active_reservation_states
value: [active, reserved, fulfilled, committed, available]
unit: categorical_set
evidence_status: source_informed
source_refs: [S23, S24, S26, S35]
rationale: "Only active or fulfilled reservation states support availability."
counterevidence_or_failure_modes:
- "provider state names and effective start times vary"
calibration_note: "Normalize local scheduler/provider reservation states."
false_positive_or_limitations:
- "Reservation is not a running or billing record."
cannot_conclude: [activity, achieved_compute, training]
promotion_policy: "Join with running and billing intervals for accelerator-hour facts."
source_refs: [S23, S24, S26, S35]
- id: quota_capacity_limit_screen
purpose: sparse_quota_capacity_limit_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [quota_limit_intervals]
emitted_statement:
label: quota_capacity_limit_screen
category: capacity_availability
max_confidence: weak
statement: "Quota intervals can allow, cap, or forbid accelerator capacity only when the quota scope and resource type match the monitored claim."
trigger:
type: one_feature_quota_logic
evaluation_window: quota_interval
conditions:
all:
- metric: quota_limit_value
op: ">="
value_ref: min_quota_limit_value_to_emit
- metric: quota_resource_type_is_accelerator
op: "=="
value: true
calibration_defaults:
- name: min_quota_limit_value_to_emit
value: 0
unit: quota_units
evidence_status: source_backed
source_refs: [S23, S24, S35]
rationale: "Even a zero quota is useful as a capacity-limit fact if the scope and resource type match."
counterevidence_or_failure_modes:
- "quotas can be changed outside the observed interval or split across accounts/projects"
calibration_note: "Normalize quota unit, account/project/region scope, and effective-time semantics."
false_positive_or_limitations:
- "Quota does not prove actual capacity or activity."
- "Unenforced or mismatched quota scopes cannot support strong limits."
cannot_conclude: [activity, training, physical_capacity]
promotion_policy: "Join with usage, reservation, and quota records in capacity and discrepancy aggregation."
source_refs: [S23, S24, S35]
- id: running_accelerator_hours_screen
purpose: sparse_running_usage_capacity_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [compute_running_intervals]
- [accelerator_compute_billing_usage_intervals]
emitted_statement:
label: running_or_billed_accelerator_hours_screen
category: capacity_availability
max_confidence: screen
statement: "Running or billed accelerator-bearing compute intervals imply accelerator-hours existed, but not the workload type or achieved operations."
trigger:
type: one_feature_accelerator_hour_logic
evaluation_window: running_or_billing_interval
conditions:
all:
- metric: accelerator_hours
op: ">="
value_ref: min_accelerator_hours_to_emit
- metric: interval_record_coverage_fraction
op: ">="
value_ref: min_running_or_billing_coverage_fraction
calibration_defaults:
- name: min_accelerator_hours_to_emit
value: 1
unit: accelerator_hours
evidence_status: source_backed
source_refs: [S23, S24, S26, S35]
rationale: "One accelerator-hour is enough to emit a sparse use-of-capacity fact."
counterevidence_or_failure_modes:
- "billing records can aggregate or lag actual running intervals"
calibration_note: "Normalize usage quantity units and provider/scheduler time boundaries."
- name: min_running_or_billing_coverage_fraction
value: 0.50
unit: fraction
evidence_status: source_informed
source_refs: [S23, S26, S35]
rationale: "Sparse capacity facts can be useful with partial coverage, but should carry coverage caveats."
counterevidence_or_failure_modes:
- "partial billing export can hide start/end times"
calibration_note: "Replace with provider export completeness and aggregation interval."
false_positive_or_limitations:
- "Running/billing can cover idle, serving, inference, HPC, or setup time."
cannot_conclude: [training, achieved_compute]
promotion_policy: "Join with activity counters for run-existence candidates and with peak rate for capacity upper bounds."
source_refs: [S23, S24, S26, S35]
- id: reservation_billing_capacity_screen
purpose: sparse_reserved_capacity_billing_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [accelerator_reservation_billing_usage_intervals]
emitted_statement:
label: reserved_capacity_billing_screen
category: capacity_availability
max_confidence: screen
statement: "Accelerator reservation billing shows paid or held capacity over time, not observed compute activity."
trigger:
type: one_feature_reservation_billing_logic
evaluation_window: reservation_billing_interval
conditions:
all:
- metric: reserved_capacity_usage_quantity
op: ">="
value_ref: min_reserved_capacity_usage_quantity_to_emit
calibration_defaults:
- name: min_reserved_capacity_usage_quantity_to_emit
value: 1
unit: provider_usage_units
evidence_status: source_backed
source_refs: [S23, S26, S35]
rationale: "Nonzero reservation usage quantity is a capacity-holding fact."
counterevidence_or_failure_modes:
- "billing units may be dollars, hours, instance-hours, or reservation-hours depending on export settings"
calibration_note: "Normalize billing usage units and reservation SKU mapping."
false_positive_or_limitations:
- "Reservation billing can occur without active launched instances."
cannot_conclude: [activity, training, achieved_compute]
promotion_policy: "Join with reservation, running, or quota records for provider capacity consistency."
source_refs: [S23, S26, S35]
- id: electrical_service_capacity_screen
purpose: sparse_physical_power_service_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [utility_interconnection_status_intervals]
- [electrical_service_status_intervals]
- [electrical_permit_status_intervals]
emitted_statement:
label: electrical_service_capacity_screen
category: capacity_power_service
max_confidence: weak
statement: "Utility, electrical-service, or permit records can support or limit physical capacity timing and IT-load plausibility."
trigger:
type: one_feature_electrical_service_logic
evaluation_window: electrical_or_utility_interval
conditions:
any:
- all:
- metric: service_or_contracted_capacity_mw
op: ">="
value_ref: min_service_capacity_mw_to_emit
- metric: service_or_permit_state
op: in
value_ref: capacity_supporting_service_states
- all:
- metric: service_or_permit_state
op: in
value_ref: capacity_limiting_service_states
calibration_defaults:
- name: min_service_capacity_mw_to_emit
value: 1.0
unit: megawatts
evidence_status: calibration_default
source_refs: [S40, S41, S44]
rationale: "One megawatt is a useful sparse facility-scale capacity fact, while exact accelerator capacity depends on power chain and PUE/IT-load mapping."
counterevidence_or_failure_modes:
- "service capacity can be shared with non-accelerator loads"
- "temporary generation or alternate feeds can be out of records"
calibration_note: "Replace with site electrical one-line, IT-load allocation, and commissioning records."
- name: capacity_supporting_service_states
value: [active, energized, approved, in_service, connected, commissioned]
unit: categorical_set
evidence_status: source_informed
source_refs: [S40, S41, S44]
rationale: "Supportive service states indicate plausible power availability."
counterevidence_or_failure_modes:
- "public or permit records can lag actual energization"
calibration_note: "Normalize utility, permit, and facility state names."
- name: capacity_limiting_service_states
value: [requested, pending, denied, suspended, not_energized, disconnected, expired]
unit: categorical_set
evidence_status: source_informed
source_refs: [S40, S41, S44]
rationale: "Limiting states route physical feasibility review."
counterevidence_or_failure_modes:
- "records can be incomplete or apply to only part of a site"
calibration_note: "Normalize state labels and effective dates per source."
false_positive_or_limitations:
- "Electrical service records are capacity context, not activity."
cannot_conclude: [activity, training, no_threshold_run]
promotion_policy: "Join with asset and power telemetry in capacity-segment and discrepancy aggregation."
source_refs: [S40, S41, S44]
- id: installation_timeline_capacity_screen
purpose: sparse_asset_installation_timeline_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [asset_receiving_installation_events]
emitted_statement:
label: asset_installation_capacity_timeline_screen
category: capacity_power_service
max_confidence: weak
statement: "Asset receipt or installation events bound the earliest plausible local capacity timeline, subject to commissioning and service records."
trigger:
type: one_feature_asset_timeline_logic
evaluation_window: asset_event_time_with_validity_extension
conditions:
all:
- metric: asset_quantity
op: ">="
value_ref: min_asset_quantity_to_emit
- metric: event_type
op: in
value_ref: capacity_supporting_asset_event_types
calibration_defaults:
- name: min_asset_quantity_to_emit
value: 1
unit: assets
evidence_status: source_backed
source_refs: [S40, S41, S44]
rationale: "One source-emitted receipt or installation event is useful timeline context."
counterevidence_or_failure_modes:
- "receipt does not equal installed, tested, powered, or commissioned capacity"
calibration_note: "Map asset categories to accelerator, server, rack, and power-chain capacity."
- name: capacity_supporting_asset_event_types
value: [received, installed, commissioned, racked, accepted, in_service]
unit: categorical_set
evidence_status: source_informed
source_refs: [S40, S41, S44]
rationale: "These event types can establish a plausible capacity timeline."
counterevidence_or_failure_modes:
- "event naming varies; commissioning can lag receipt or installation"
calibration_note: "Normalize asset event taxonomy and commissioning delay."
false_positive_or_limitations:
- "Asset event records can be administrative and not operational."
cannot_conclude: [activity, energized_service, training]
promotion_policy: "Join with electrical service and inventory for physical timeline checks."
source_refs: [S40, S41, S44]
- id: maintenance_availability_screen
purpose: sparse_maintenance_capacity_screen
raw_feature_budget: 1
required_feature_sets:
any_one_of:
- [maintenance_operation_intervals]
emitted_statement:
label: maintenance_capacity_availability_screen
category: capacity_availability
max_confidence: weak
statement: "Maintenance intervals can reduce available capacity, split topology validity, or explain telemetry gaps during the affected window."
trigger:
type: one_feature_maintenance_availability_logic
evaluation_window: maintenance_interval
conditions:
any:
- metric: affected_asset_count
op: ">="
value_ref: min_maintenance_affected_asset_count
- metric: maintenance_status
op: in
value_ref: active_or_capacity_affecting_maintenance_states
calibration_defaults:
- name: min_maintenance_affected_asset_count
value: 1
unit: assets
evidence_status: source_informed
source_refs: [S18, S40, S41, S42, S43]
rationale: "Any affected asset can matter for sparse capacity routing when the scope is an accelerator cluster or fabric domain."
counterevidence_or_failure_modes:
- "maintenance records may be broad or unrelated to accelerator capacity"
calibration_note: "Map maintenance asset categories to capacity, topology, telemetry, and storage surfaces."
- name: active_or_capacity_affecting_maintenance_states
value: [active, in_progress, draining, offline, degraded, completed_with_impact]
unit: categorical_set
evidence_status: source_informed
source_refs: [S18, S40, S41, S42, S43]
rationale: "Active or impact states can reduce capacity or explain sparse-rule contradictions."
counterevidence_or_failure_modes:
- "completed or scheduled maintenance may not overlap actual impact"
calibration_note: "Normalize maintenance status and affected asset semantics."
false_positive_or_limitations:
- "Maintenance can over-explain if intervals are coarse."
cannot_conclude: [training, no_training, achieved_compute]
promotion_policy: "Join with health, topology, power, or telemetry coverage for derived capacity and discrepancy rules."
source_refs: [S18, S40, S41, S42, S43]
|