schema_version: "0.1" created: "2026-05-16" source_inventory: observables/observables.yaml status: sparse_first_algorithm_rule_spec purpose: > Sparse discrepancy, integrity, and evasion screens for one-, two-, and three-feature settings. These rules express narrow contradictions, missingness warnings, and benign explanations that feed targeted aggregation review. scope_policy: trusted_local_inputs: "Only raw feature IDs from observables/observables.yaml are used as inputs." relationship_to_aggregation_rules: > This file emits sparse conflict and missingness statements for targeted C review in `aggregation_rules.yaml`. missingness_policy: > Missing telemetry is an output condition, not evidence of absence. Sparse discrepancy rules must report which channel was missing or only partially covered. severity_scale: [info, low, medium, high, critical] discrepancy_categories: - id: activity_attribution_conflict name: Activity attribution conflicts algorithm_treatment: "Escalate high accelerator activity with absent or sparse allocation, running, container, or billing attribution." rule_ids: - activity_without_allocation_or_usage_conflict - id: capacity_claim_conflict name: Capacity claim conflicts algorithm_treatment: "Compare achieved/claimed/running usage against sparse peak, quota, reservation, or availability limits." rule_ids: - achieved_ops_above_peak_capacity_conflict - usage_above_quota_or_reservation_conflict - id: physical_timeline_conflict name: Physical timeline conflicts algorithm_treatment: "Flag activity or usage before installation, energized service, or utility readiness." rule_ids: - activity_before_installation_or_service_conflict - id: power_activity_conflict name: Power and activity conflicts algorithm_treatment: "Compare physical load with accelerator activity and activity coverage." rule_ids: - high_power_low_or_missing_activity_conflict - id: fabric_mapping_conflict name: Fabric mapping conflicts algorithm_treatment: "Route high fabric activity with missing job/container/topology mapping to attribution review." rule_ids: - fabric_without_job_or_topology_mapping_conflict - id: storage_activity_conflict name: Storage activity conflicts and explanations algorithm_treatment: "Check checkpoint-like writes against accelerator activity and explicit storage operations." rule_ids: - checkpoint_writes_without_activity_conflict - fabric_or_storage_activity_explained_by_storage_operation - id: telemetry_missingness_conflict name: Candidate-window telemetry gaps algorithm_treatment: "Emit missingness warnings during candidate, high-activity, or high-power windows." rule_ids: - telemetry_gap_during_sparse_candidate_window - id: health_throttle_conflict name: Health, throttle, and limit conflicts algorithm_treatment: "Check high activity or claims against degraded, throttled, power-limited, or thermally limited states." rule_ids: - high_activity_during_degraded_or_throttled_state_conflict - id: topology_route_conflict name: Topology and route conflicts algorithm_treatment: "Treat route, NAT, gateway, and topology changes as interpretation-changing events during candidate windows." rule_ids: - route_or_topology_change_during_sparse_candidate rule_category_membership: membership_policy: > Categories are evidentiary roles. Each sparse discrepancy rule has exactly one primary category for severity routing and may list additional categories for explanation or missingness handling. rules: - rule_id: activity_without_allocation_or_usage_conflict primary_category: activity_attribution_conflict additional_categories: [telemetry_missingness_conflict] - rule_id: achieved_ops_above_peak_capacity_conflict primary_category: capacity_claim_conflict additional_categories: [] - rule_id: usage_above_quota_or_reservation_conflict primary_category: capacity_claim_conflict additional_categories: [] - rule_id: activity_before_installation_or_service_conflict primary_category: physical_timeline_conflict additional_categories: [capacity_claim_conflict] - rule_id: high_power_low_or_missing_activity_conflict primary_category: power_activity_conflict additional_categories: [telemetry_missingness_conflict] - rule_id: fabric_without_job_or_topology_mapping_conflict primary_category: fabric_mapping_conflict additional_categories: [activity_attribution_conflict, telemetry_missingness_conflict] - rule_id: checkpoint_writes_without_activity_conflict primary_category: storage_activity_conflict additional_categories: [telemetry_missingness_conflict] - rule_id: telemetry_gap_during_sparse_candidate_window primary_category: telemetry_missingness_conflict additional_categories: [] - rule_id: route_or_topology_change_during_sparse_candidate primary_category: topology_route_conflict additional_categories: [fabric_mapping_conflict, telemetry_missingness_conflict] - rule_id: fabric_or_storage_activity_explained_by_storage_operation primary_category: storage_activity_conflict additional_categories: [fabric_mapping_conflict] - rule_id: high_activity_during_degraded_or_throttled_state_conflict primary_category: health_throttle_conflict additional_categories: [capacity_claim_conflict] discrepancy_rules: - id: activity_without_allocation_or_usage_conflict purpose: sparse_activity_attribution_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [tensor_matrix_mxu_neuron_or_engine_active_fraction, allocated_accelerator_count_by_sku] - [tensor_matrix_mxu_neuron_or_engine_active_fraction, compute_running_intervals] - [tensor_matrix_mxu_neuron_or_engine_active_fraction, accelerator_compute_billing_usage_intervals] - [generic_achieved_operation_rate, allocated_accelerator_count_by_sku] - [generic_achieved_operation_rate, compute_running_intervals] - [generic_achieved_operation_rate, accelerator_compute_billing_usage_intervals] - [accelerator_busy_or_utilization_fraction, allocated_accelerator_count_by_sku] - [accelerator_busy_or_utilization_fraction, compute_running_intervals] - [accelerator_busy_or_utilization_fraction, accelerator_compute_billing_usage_intervals] - [container_resource_usage_quantity, allocated_accelerator_count_by_sku] emitted_statement: label: activity_attribution_conflict category: activity_attribution_conflict max_severity: high statement: "Sustained accelerator activity has no overlapping allocation, running, billing, or resource-usage attribution in the covered sparse channel." trigger: type: two_feature_activity_attribution_logic evaluation_window: activity_attribution_overlap_window conditions: all: - metric: sparse_activity_score op: ">=" value_ref: min_unattributed_activity_score - metric: activity_duration_seconds op: ">=" value_ref: min_unattributed_activity_duration_seconds - metric: attribution_overlap_fraction op: "<=" value_ref: max_attribution_overlap_fraction - metric: attribution_channel_coverage_fraction op: ">=" value_ref: min_attribution_channel_coverage_fraction none: - metric: maintenance_or_system_job_overlap_fraction op: ">=" value_ref: benign_attribution_explanation_overlap_fraction missing_data_behavior: "If attribution coverage is below the threshold, emit `attribution_coverage_gap` at severity medium instead of treating absence as proven." severity_mapping: low: "Activity is moderate and attribution coverage is partial." medium: "Activity is high but attribution coverage or scope mapping is incomplete." high: "Activity is high for the minimum duration with attribution overlap <= 0.05 and coverage >= 0.80." benign_explanation_or_suppressor_fields: - maintenance_operation_intervals - resource_container_lifecycle_events - scheduler_accounting_delay - provider_billing_delay calibration_defaults: - name: min_unattributed_activity_score value: 0.70 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S07, S08, S36, S37, S38, S39] rationale: "Attribution conflicts should require clear accelerator activity." counterevidence_or_failure_modes: - "activity-score normalization errors can inflate activity" - "serving, HPC, benchmarks, and burn-in can all trigger activity" calibration_note: "Calibrate by accelerator family and activity counter source." - name: min_unattributed_activity_duration_seconds value: 600 unit: seconds evidence_status: calibration_default source_refs: [S18, S19, S20, S21, S22, S23, S26, S35] rationale: "Filters short scheduler, lifecycle, or billing races while still catching material unattributed work." counterevidence_or_failure_modes: - "short unauthorized jobs can still matter" calibration_note: "Replace with scheduler/provider accounting delay distributions." - name: max_attribution_overlap_fraction value: 0.05 unit: fraction evidence_status: calibration_default source_refs: [S18, S19, S20, S23, S26, S35] rationale: "Allows small clock and aggregation mismatch while flagging effectively unmapped activity." counterevidence_or_failure_modes: - "scope mapping errors can lower apparent overlap" calibration_note: "Tune by clock drift, aggregation interval, and mapping quality." - name: min_attribution_channel_coverage_fraction value: 0.80 unit: fraction evidence_status: source_informed source_refs: [S18, S19, S20, S23, S26, S35] rationale: "No-attribution findings require adequate coverage of the attribution channel being checked." counterevidence_or_failure_modes: - "billing and scheduler exports can be delayed or incomplete" calibration_note: "Replace with source coverage and delivery-delay measurements." - name: benign_attribution_explanation_overlap_fraction value: 0.80 unit: fraction evidence_status: calibration_default source_refs: [S18, S21, S22, S40, S41] rationale: "A benign explanation should cover most of the suspicious activity before reducing severity." counterevidence_or_failure_modes: - "broad maintenance windows can over-explain unrelated activity" calibration_note: "Tune by maintenance and system-job record precision." required_output_fields: - activity_window - checked_attribution_channel - attribution_overlap_fraction - coverage_fraction - missing_attribution_channels - benign_explanations_checked false_positive_or_limitations: - "A single attribution channel may not be the authoritative source for all jobs." source_refs: [S07, S08, S18, S19, S20, S21, S22, S23, S26, S35, S36, S37, S38, S39] - id: achieved_ops_above_peak_capacity_conflict purpose: sparse_capacity_claim_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [generic_achieved_operation_rate, advertised_peak_rate_by_precision] - [generic_achieved_operation_rate, instance_type_shape_machine_type] - [generic_achieved_operation_rate, accelerator_count_by_family_sku] derived_signals: - achieved_operation_integral - capacity_upper_bound_flop emitted_statement: label: capacity_claim_conflict category: capacity_claim_conflict max_severity: critical statement: "Achieved-operation integral or claim exceeds sparse peak-derived capacity for the same scope/window by more than the uncertainty margin." trigger: type: two_feature_achieved_ops_capacity_logic evaluation_window: achieved_ops_peak_overlap_window conditions: all: - metric: achieved_operation_integral.unit_normalized op: "==" value: true - metric: achieved_ops_to_sparse_capacity_ratio op: ">" value_ref: max_sparse_achieved_to_capacity_ratio - metric: sparse_capacity_input_coverage_fraction op: ">=" value_ref: min_sparse_capacity_conflict_coverage_fraction none: - metric: unit_mismatch_or_hidden_capacity_explanation_score op: ">=" value_ref: unit_or_hidden_capacity_suppression_score missing_data_behavior: "If operation units or peak mode are not normalized, emit `unit_normalization_required` and cap severity at medium." severity_mapping: medium: "Exceedance with incomplete count/shape scope or possible unit mismatch." high: "Ratio > 1.10 with normalized units and coverage >= 0.75." critical: "Ratio > 1.25 with normalized units, coverage >= 0.90, and no hidden-capacity explanation." benign_explanation_or_suppressor_fields: - operation_unit_or_precision_mismatch - precision_or_mode_mismatch - hidden_or_unmonitored_capacity - counter_reset_or_rollover calibration_defaults: - name: max_sparse_achieved_to_capacity_ratio value: 1.10 unit: ratio evidence_status: calibration_default source_refs: [S01, S03, S04, S05, S06, S07, S08, S09, S24] rationale: "Allows sparse measurement uncertainty while flagging achieved work above optimistic capacity." counterevidence_or_failure_modes: - "mixed precision, unit mismatch, counter reset, or hidden capacity can create apparent exceedance" calibration_note: "Replace with source-specific unit normalization and capacity uncertainty model." - name: min_sparse_capacity_conflict_coverage_fraction value: 0.75 unit: fraction evidence_status: calibration_default source_refs: [S04, S05, S06, S07, S08, S09, S24] rationale: "Sparse conflict can emit before full capacity coverage but should not become high confidence without coverage." counterevidence_or_failure_modes: - "missing count, shape, or peak records can understate capacity" calibration_note: "Raise to 0.90 for critical severity in site policy." - name: unit_or_hidden_capacity_suppression_score value: 0.70 unit: score_0_to_1 evidence_status: project_assumption source_refs: [S01, S03, S04, S05, S06, S24] rationale: "Strong evidence of unit mismatch or out-of-scope capacity should suppress high severity until reviewed." counterevidence_or_failure_modes: - "claimed hidden capacity can be opportunistic or unverifiable" calibration_note: "Tie suppression to concrete source records and monitored-boundary definitions." required_output_fields: - achieved_operation_count - sparse_capacity_upper_bound_operations - exceedance_ratio - unit_normalization_status - capacity_scope - suppression_reason false_positive_or_limitations: - "Sparse peak capacity can be incomplete without count, shape, and duration." source_refs: [S01, S03, S04, S05, S06, S07, S08, S09, S24] - id: usage_above_quota_or_reservation_conflict purpose: sparse_cloud_capacity_consistency_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [quota_limit_intervals, compute_running_intervals] - [quota_limit_intervals, accelerator_compute_billing_usage_intervals] - [quota_limit_intervals, capacity_reservation_intervals] - [reservation_state_intervals, accelerator_compute_billing_usage_intervals] - [capacity_reservation_intervals, accelerator_compute_billing_usage_intervals] - [accelerator_reservation_billing_usage_intervals, compute_running_intervals] emitted_statement: label: capacity_claim_conflict category: capacity_claim_conflict max_severity: high statement: "Running, billed, reserved, or requested accelerator usage appears above the matching quota or reservation constraint." trigger: type: two_feature_quota_reservation_usage_logic evaluation_window: quota_reservation_usage_overlap_window conditions: all: - metric: usage_to_quota_or_reservation_ratio op: ">" value_ref: max_usage_to_quota_or_reservation_ratio - metric: cloud_control_plane_scope_match_score op: ">=" value_ref: min_cloud_scope_match_score - metric: quota_or_reservation_overlap_fraction op: ">=" value_ref: min_quota_usage_overlap_fraction none: - metric: quota_change_or_cross_scope_explanation_score op: ">=" value_ref: quota_explanation_suppression_score missing_data_behavior: "If provider account/project/region/zone scope is unknown, cap severity at medium and output `scope_match_unknown`." severity_mapping: low: "Small exceedance with weak scope match." medium: "Usage exceeds quota/reservation with partial coverage or possible aggregation delay." high: "Usage ratio > 1.10 with scope match >= 0.80 and no quota-change or cross-scope explanation." benign_explanation_or_suppressor_fields: - quota_change_not_in_interval_feed - cross_account_or_project_split - reserved_capacity_exception - billing_aggregation_delay calibration_defaults: - name: max_usage_to_quota_or_reservation_ratio value: 1.10 unit: ratio evidence_status: calibration_default source_refs: [S23, S24, S26, S35] rationale: "Allows provider export and unit uncertainty while flagging usage above sparse limits." counterevidence_or_failure_modes: - "quota units, reservation units, and billing units may not be directly comparable" calibration_note: "Normalize provider resource units and effective-time semantics." - name: min_cloud_scope_match_score value: 0.80 unit: score_0_to_1 evidence_status: source_informed source_refs: [S23, S24, S26, S35] rationale: "Cloud quota and usage comparisons are meaningful only for matching account/project/region/zone or placement scopes." counterevidence_or_failure_modes: - "cross-account, organization, or managed-service usage can be split" calibration_note: "Calibrate scope matching by provider export fields." - name: min_quota_usage_overlap_fraction value: 0.80 unit: fraction evidence_status: calibration_default source_refs: [S23, S24, S26, S35] rationale: "Quota/reservation and usage intervals should overlap most of the compared window." counterevidence_or_failure_modes: - "billing records can be delayed or aggregated" calibration_note: "Tune by provider interval semantics." - name: quota_explanation_suppression_score value: 0.75 unit: score_0_to_1 evidence_status: project_assumption source_refs: [S23, S24, S26, S35] rationale: "Strong quota-change, cross-scope, or provider exception evidence should lower severity." counterevidence_or_failure_modes: - "exception records can be incomplete or ambiguous" calibration_note: "Tie suppression to concrete provider records." required_output_fields: - provider_scope - usage_quantity - quota_or_reservation_bound - exceedance_ratio - scope_match_score - explanation_status false_positive_or_limitations: - "Provider quotas are not always hard physical capacity limits." source_refs: [S23, S24, S26, S35] - id: activity_before_installation_or_service_conflict purpose: sparse_physical_timeline_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [compute_running_intervals, asset_receiving_installation_events] - [accelerator_compute_billing_usage_intervals, asset_receiving_installation_events] - [accelerator_busy_or_utilization_fraction, asset_receiving_installation_events] - [tensor_matrix_mxu_neuron_or_engine_active_fraction, asset_receiving_installation_events] - [compute_running_intervals, electrical_service_status_intervals] - [accelerator_compute_billing_usage_intervals, electrical_service_status_intervals] - [accelerator_busy_or_utilization_fraction, electrical_service_status_intervals] - [compute_running_intervals, utility_interconnection_status_intervals] - [compute_running_intervals, electrical_permit_status_intervals] derived_signals: - available_accelerator_fraction - capacity_upper_bound_flop emitted_statement: label: physical_timeline_conflict category: physical_timeline_conflict max_severity: critical statement: "Observed activity, running, or usage predates documented asset installation, utility readiness, permit approval, or energized electrical service for the same monitored site." trigger: type: two_feature_physical_timeline_logic evaluation_window: activity_or_usage_with_physical_timeline_window conditions: all: - metric: activity_or_usage_score op: ">=" value_ref: min_sparse_activity_or_usage_score_for_timeline - metric: physical_prerequisite_state op: in value_ref: capacity_not_ready_states - metric: activity_starts_before_ready_delta_seconds op: "<" value_ref: negative_timeline_grace_period_seconds none: - metric: documented_alternative_capacity_score op: ">=" value_ref: alternative_capacity_suppression_score missing_data_behavior: "If physical timeline records are absent rather than contradictory, emit `physical_records_missing` and cap severity at high." severity_mapping: medium: "Activity overlaps incomplete or pending physical records." high: "Activity predates one physical prerequisite with incomplete alternate-capacity review." critical: "Activity predates ready state by more than the grace period and no alternate capacity is documented." benign_explanation_or_suppressor_fields: - alternate_capacity_site_or_scope - temporary_generation - public_record_lag - asset_event_tracks_receipt_not_commissioning calibration_defaults: - name: min_sparse_activity_or_usage_score_for_timeline value: 0.50 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S07, S08, S23, S26, S35, S36, S37, S38, S39] rationale: "Physical timeline conflict requires material observed activity, running, or usage." counterevidence_or_failure_modes: - "cloud running or billing can be scoped away from the physical site" calibration_note: "Calibrate activity and usage materiality by source type." - name: capacity_not_ready_states value: [not_installed, received_only, pending, requested, denied, not_energized, disconnected, expired] unit: categorical_set evidence_status: source_informed source_refs: [S40, S41, S44] rationale: "These physical states contradict local activity unless alternate capacity is documented." counterevidence_or_failure_modes: - "state names vary and public records can lag operations" calibration_note: "Normalize utility, permit, service, and asset event states." - name: negative_timeline_grace_period_seconds value: -3600 unit: seconds evidence_status: calibration_default source_refs: [S40, S41, S44] rationale: "Allows one hour of timestamp and interval-boundary uncertainty." counterevidence_or_failure_modes: - "administrative records can lag actual energization by longer periods" calibration_note: "Replace with source-specific timestamp and record-lag accuracy." - name: alternative_capacity_suppression_score value: 0.80 unit: score_0_to_1 evidence_status: project_assumption source_refs: [S40, S41, S44] rationale: "Documented alternate site, temporary generation, or out-of-scope capacity can suppress impossibility." counterevidence_or_failure_modes: - "alternate-capacity claims may be incomplete or unverifiable" calibration_note: "Tie to concrete site, utility, and monitored-boundary records." required_output_fields: - activity_or_usage_window - physical_prerequisite - ready_time - timeline_delta_seconds - alternate_capacity_status - missing_physical_records false_positive_or_limitations: - "Physical site records can be incomplete, delayed, or scoped to only part of a site." source_refs: [S07, S08, S23, S26, S35, S36, S37, S38, S39, S40, S41, S44] - id: high_power_low_or_missing_activity_conflict purpose: sparse_power_activity_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [rack_pdu_it_power, accelerator_busy_or_utilization_fraction] - [rack_pdu_it_power, tensor_matrix_mxu_neuron_or_engine_active_fraction] - [server_system_input_power, accelerator_busy_or_utilization_fraction] - [server_system_input_power, tensor_matrix_mxu_neuron_or_engine_active_fraction] - [facility_or_room_it_power_rollup, accelerator_busy_or_utilization_fraction] - [accelerator_device_power_energy, tensor_matrix_mxu_neuron_or_engine_active_fraction] derived_signals: - telemetry_coverage_fraction - clock_aligned_overlap_window emitted_statement: label: power_activity_conflict category: power_activity_conflict max_severity: high statement: "High physical load conflicts with low or missing accelerator activity in the same mapped scope/window." trigger: type: two_feature_power_activity_logic evaluation_window: lag_adjusted_power_activity_window conditions: all: - metric: baseline_subtracted_power_fraction_of_scope_capacity op: ">=" value_ref: min_sparse_power_load_fraction - metric: lag_adjusted_overlap_duration_seconds op: ">=" value_ref: min_sparse_power_activity_overlap_seconds any: - metric: sparse_activity_score op: "<=" value_ref: max_sparse_low_activity_score - metric: telemetry_coverage_fraction.accelerator_activity op: "<" value_ref: min_sparse_activity_coverage_fraction none: - metric: maintenance_or_cooling_explanation_overlap_fraction op: ">=" value_ref: benign_power_explanation_overlap_fraction missing_data_behavior: "If activity coverage is low, emit `activity_missing_during_power_load` instead of concluding true low accelerator activity." severity_mapping: low: "High power with weak scope mapping or benign explanation." medium: "High power with low or missing activity for the minimum duration." high: "High power with covered low activity and no maintenance, cooling, mixed-tenant, or storage explanation." benign_explanation_or_suppressor_fields: - maintenance_operation_intervals - cooling_alarm_state - mixed_tenant_or_non_gpu_load - cpu_or_storage_load calibration_defaults: - name: min_sparse_power_load_fraction value: 0.25 unit: fraction_of_scope_power_capacity evidence_status: calibration_default source_refs: [S07, S08, S09, S40, S41] rationale: "Power must be materially above baseline before challenging accelerator activity telemetry." counterevidence_or_failure_modes: - "CPU, storage, network, cooling faults, and mixed tenancy can raise power" calibration_note: "Replace with rack/server/device baselines and scope capacity." - name: min_sparse_power_activity_overlap_seconds value: 600 unit: seconds evidence_status: calibration_default source_refs: [S07, S08, S09, S40, S41] rationale: "Avoids short transients and telemetry skew." counterevidence_or_failure_modes: - "brief high-power tests can still matter" calibration_note: "Tune by power and accelerator scrape intervals." - name: max_sparse_low_activity_score value: 0.20 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S07, S08, S36, S37, S38, S39] rationale: "Defines low normalized accelerator activity for sparse conflict review." counterevidence_or_failure_modes: - "activity counters can under-report or miss non-GPU load" calibration_note: "Calibrate by confirmed idle and active windows." - name: min_sparse_activity_coverage_fraction value: 0.80 unit: fraction evidence_status: source_informed source_refs: [S07, S08, S36, S37, S38, S39] rationale: "Low activity is meaningful only when activity coverage is adequate." counterevidence_or_failure_modes: - "collector gaps can mimic low activity" calibration_note: "Replace with activity telemetry coverage." - name: benign_power_explanation_overlap_fraction value: 0.80 unit: fraction evidence_status: calibration_default source_refs: [S18, S40, S41] rationale: "A benign explanation should cover most of the high-power window before lowering severity." counterevidence_or_failure_modes: - "broad maintenance windows can over-explain" calibration_note: "Tune by maintenance, cooling, and scope-mapping quality." required_output_fields: - power_window - mapped_scope - activity_score - activity_coverage_fraction - power_delta_watts - explanations_checked false_positive_or_limitations: - "Power is often coarser than accelerator telemetry and may include unrelated loads." source_refs: [S07, S08, S09, S18, S36, S37, S38, S39, S40, S41] - id: fabric_without_job_or_topology_mapping_conflict purpose: sparse_fabric_mapping_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [scaleout_port_tx_rx_bytes_packets, allocated_accelerator_count_by_sku] - [scaleout_port_tx_rx_bytes_packets, compute_running_intervals] - [scaleout_port_tx_rx_bytes_packets, container_resource_usage_quantity] - [fabric_port_device_sample_counters, allocated_accelerator_count_by_sku] - [fabric_port_device_sample_counters, compute_running_intervals] - [local_interconnect_tx_rx_bytes, allocated_accelerator_count_by_sku] - [east_west_internal_flow_bytes, compute_running_intervals] derived_signals: - collective_cadence_score - stable_participant_set - telemetry_coverage_fraction emitted_statement: label: fabric_mapping_conflict category: fabric_mapping_conflict max_severity: high statement: "High or collective-like fabric activity has no stable job, container, allocation, running, or topology mapping in the covered sparse channel." trigger: type: two_feature_fabric_mapping_logic evaluation_window: fabric_mapping_overlap_window conditions: all: - metric: sparse_fabric_activity_score op: ">=" value_ref: min_sparse_fabric_activity_score - metric: sparse_fabric_duration_seconds op: ">=" value_ref: min_sparse_fabric_duration_seconds - metric: mapping_coverage_fraction op: "<=" value_ref: max_sparse_mapping_coverage_fraction none: - metric: storage_or_network_test_explanation_score op: ">=" value_ref: fabric_explanation_suppression_score missing_data_behavior: "Missing mapping is the finding; output which mapping channel was expected, observed, or unavailable." severity_mapping: low: "Fabric activity with partial mapping gap." medium: "High fabric activity with mapping coverage <= 0.20 or unknown participant mapping." high: "High fabric activity for the minimum duration with mapping coverage <= 0.05 and no storage, network-test, route-change, or maintenance explanation." benign_explanation_or_suppressor_fields: - storage_operation_intervals - object_or_file_copy_movement_records - topology_change_events - maintenance_operation_intervals calibration_defaults: - name: min_sparse_fabric_activity_score value: 0.60 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S11, S25, S28, S29, S42, S43] rationale: "Requires material normalized fabric or internal-flow activity before treating a mapping gap as important." counterevidence_or_failure_modes: - "storage, inference, benchmarks, and HPC can all raise fabric activity" calibration_note: "Replace with local fabric score distributions." - name: min_sparse_fabric_duration_seconds value: 600 unit: seconds evidence_status: calibration_default source_refs: [S11, S25, S42, S43] rationale: "Filters short bursts and control-plane noise." counterevidence_or_failure_modes: - "short communication tests can still matter" calibration_note: "Tune by fabric aggregation and local benchmark durations." - name: max_sparse_mapping_coverage_fraction value: 0.20 unit: fraction evidence_status: calibration_default source_refs: [S18, S20, S25, S42, S43] rationale: "At most sparse mapping exists for the suspicious fabric window." counterevidence_or_failure_modes: - "mapping source outages can create false positives" calibration_note: "Calibrate port-to-job, container, and topology mapping reliability." - name: fabric_explanation_suppression_score value: 0.75 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S25, S31, S32, S33, S42, S43] rationale: "Strong alternative explanation should lower mapping-conflict severity." counterevidence_or_failure_modes: - "storage explanations or tests can overlap real collective traffic" calibration_note: "Tune with storage, benchmark, HPC, and maintenance labels." required_output_fields: - fabric_window - fabric_activity_score - missing_mapping_channel - mapping_coverage_fraction - participant_scope - explanations_checked false_positive_or_limitations: - "Aggregate fabric counters may not expose participant identity." source_refs: [S11, S18, S20, S25, S28, S29, S31, S32, S33, S42, S43] - id: checkpoint_writes_without_activity_conflict purpose: sparse_storage_activity_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [storage_write_operation_bytes, tensor_matrix_mxu_neuron_or_engine_active_fraction] - [storage_write_operation_bytes, accelerator_busy_or_utilization_fraction] - [storage_write_operation_bytes, generic_achieved_operation_rate] - [object_storage_operation_counts, tensor_matrix_mxu_neuron_or_engine_active_fraction] - [object_storage_operation_counts, accelerator_busy_or_utilization_fraction] - [filesystem_throughput_metadata_ops, tensor_matrix_mxu_neuron_or_engine_active_fraction] derived_signals: - checkpoint_periodicity_score - checkpoint_size_consistency - telemetry_coverage_fraction emitted_statement: label: storage_activity_conflict category: storage_activity_conflict max_severity: high statement: "Checkpoint-like writes appear without overlapping or adjacent accelerator activity in covered telemetry." trigger: type: two_feature_checkpoint_activity_logic evaluation_window: checkpoint_activity_adjacency_window conditions: all: - metric: checkpoint_like_write_score op: ">=" value_ref: min_checkpoint_like_write_score - metric: write_burst_count op: ">=" value_ref: min_sparse_checkpoint_write_burst_count - metric: checkpoint_activity_adjacency_fraction op: "<=" value_ref: max_checkpoint_activity_adjacency_fraction none: - metric: storage_operation_explanation_score op: ">=" value_ref: checkpoint_storage_explanation_suppression_score missing_data_behavior: "If activity telemetry coverage is below threshold, emit `activity_missing_near_writes` and route to telemetry-gap review." severity_mapping: info: "Writes are explained by explicit storage operation." medium: "Checkpoint-like writes with low or missing activity coverage." high: "Repeated checkpoint-like writes with adequate activity coverage, adjacency <= 0.10, and no storage explanation." benign_explanation_or_suppressor_fields: - storage_operation_intervals - object_or_file_copy_movement_records - backup_or_snapshot - etl_export - synthetic_data_generation_output calibration_defaults: - name: min_checkpoint_like_write_score value: 0.55 unit: score_0_to_1 evidence_status: source_informed source_refs: [S31, S32, S33, S45, S46, S47] rationale: "Repeated write bursts and object/file metadata can be checkpoint-like but are not unique to training." counterevidence_or_failure_modes: - "backups, snapshots, ETL, replication, and generated data can mimic checkpoints" calibration_note: "Replace with local checkpoint and storage-operation traces." - name: min_sparse_checkpoint_write_burst_count value: 2 unit: bursts evidence_status: calibration_default source_refs: [S45, S46, S47] rationale: "Two bursts are needed before a sparse write pattern resembles checkpoint cadence." counterevidence_or_failure_modes: - "short jobs may save only one final checkpoint" calibration_note: "Calibrate by checkpoint policy and job duration." - name: max_checkpoint_activity_adjacency_fraction value: 0.10 unit: fraction evidence_status: calibration_default source_refs: [S31, S32, S33, S45, S46, S47] rationale: "Allows small timing mismatch while flagging writes mostly unadjacent to activity." counterevidence_or_failure_modes: - "asynchronous checkpoint flushes can lag activity" calibration_note: "Tune adjacency window by framework and storage logging interval." - name: checkpoint_storage_explanation_suppression_score value: 0.75 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S31, S32, S33, S34, S35] rationale: "A strong storage operation explanation should lower severity." counterevidence_or_failure_modes: - "backup can overlap real checkpointing" calibration_note: "Calibrate storage operation labels and byte attribution." required_output_fields: - write_window - write_burst_count - checkpoint_like_write_score - activity_adjacency_fraction - activity_coverage_fraction - storage_explanation_status false_positive_or_limitations: - "Checkpoint-like storage shape is not unique to training." source_refs: [S31, S32, S33, S34, S35, S45, S46, S47] - id: telemetry_gap_during_sparse_candidate_window purpose: sparse_telemetry_missingness_conflict raw_feature_budget: 1 required_feature_sets: any_one_of: - [allocated_accelerator_count_by_sku] - [compute_running_intervals] - [generic_achieved_operation_rate] - [tensor_matrix_mxu_neuron_or_engine_active_fraction] - [accelerator_busy_or_utilization_fraction] - [scaleout_port_tx_rx_bytes_packets] - [storage_write_operation_bytes] - [accelerator_device_power_energy] - [rack_pdu_it_power] derived_signals: - telemetry_coverage_fraction emitted_statement: label: telemetry_missingness_conflict category: telemetry_missingness_conflict max_severity: critical statement: "One or more expected primary telemetry channels have low coverage during a sparse candidate, high-activity, high-fabric, high-storage, or high-power window." trigger: type: one_feature_candidate_missingness_logic evaluation_window: sparse_candidate_or_high_activity_window conditions: all: - metric: sparse_candidate_window_score op: ">=" value_ref: min_sparse_candidate_window_score_for_gap - metric: telemetry_coverage_fraction.primary_channel_min op: "<" value_ref: min_required_sparse_primary_coverage_fraction any: - metric: missing_channel_importance_score op: ">=" value_ref: min_sparse_missing_channel_importance_score - metric: telemetry_gap_overlap_fraction op: ">=" value_ref: min_sparse_gap_overlap_fraction none: - metric: documented_collector_or_maintenance_outage_overlap_fraction op: ">=" value_ref: benign_gap_explanation_overlap_fraction missing_data_behavior: "The missing data is the finding; do not infer absence of training, serving, storage, or power activity from the gap." severity_mapping: medium: "One support channel gaps during a weak sparse candidate." high: "A primary activity, fabric, storage, serving, power, allocation, or billing channel gaps during a candidate window." critical: "Multiple primary channels gap during a policy-scale or high-confidence candidate and coincide with unexplained physical access or maintenance." benign_explanation_or_suppressor_fields: - documented_collector_or_source_outage - log_configuration_disabled - provider_best_effort_delay - maintenance_operation_intervals - physical_access_events calibration_defaults: - name: min_sparse_candidate_window_score_for_gap value: 0.50 unit: score_0_to_1 evidence_status: project_assumption source_refs: [S07, S08, S27, S28, S29, S30, S31, S35, S36, S37, S38, S39] rationale: "Telemetry gaps are escalated when they overlap a candidate or materially active sparse window." counterevidence_or_failure_modes: - "candidate scoring can under- or over-prioritize windows" calibration_note: "Calibrate from rule outputs and known incident reviews." - name: min_required_sparse_primary_coverage_fraction value: 0.75 unit: fraction evidence_status: source_informed source_refs: [S07, S08, S27, S28, S29, S30, S31, S35, S36, S37, S38, S39] rationale: "Primary channels below this coverage materially weaken sparse verification." counterevidence_or_failure_modes: - "some sources are sampled or best-effort by design and need adjusted expectations" calibration_note: "Replace with expected source coverage by channel." - name: min_sparse_missing_channel_importance_score value: 0.60 unit: score_0_to_1 evidence_status: project_assumption source_refs: [S07, S08, S27, S28, S29, S30, S31, S35] rationale: "Prioritizes channels that would materially affect the sparse candidate label." counterevidence_or_failure_modes: - "importance depends on available alternate channels" calibration_note: "Calibrate with the evidence-combination model." - name: min_sparse_gap_overlap_fraction value: 0.50 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S27, S28, S29, S30, S31, S35] rationale: "A gap should cover a material fraction of the candidate window." counterevidence_or_failure_modes: - "short gaps can hide decisive bursts" calibration_note: "Tune by sampling interval and expected burst duration." - name: benign_gap_explanation_overlap_fraction value: 0.80 unit: fraction evidence_status: calibration_default source_refs: [S18, S21, S22, S27, S40, S41] rationale: "Documented outages or maintenance can lower integrity severity when they explain most of the gap." counterevidence_or_failure_modes: - "planned outages can coincide with evasion or unrelated gaps" calibration_note: "Tune with collector outage and maintenance records." required_output_fields: - candidate_window - candidate_basis_feature - missing_channels - coverage_fraction_by_channel - gap_overlap_fraction - explanations_checked false_positive_or_limitations: - "Expected coverage must be source-specific; generic defaults are only routing defaults." source_refs: [S07, S08, S18, S21, S22, S27, S28, S29, S30, S31, S35, S36, S37, S38, S39, S40, S41] - id: route_or_topology_change_during_sparse_candidate purpose: sparse_topology_route_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [topology_change_events, scaleout_port_tx_rx_bytes_packets] - [topology_change_events, fabric_port_device_sample_counters] - [topology_change_events, compute_running_intervals] - [network_gateway_nat_route_state, vpc_or_nsg_flow_records] - [network_gateway_nat_route_state, east_west_internal_flow_bytes] - [network_gateway_nat_route_state, north_south_external_egress] derived_signals: - largest_usable_topology_domain - telemetry_coverage_fraction - clock_aligned_overlap_window emitted_statement: label: topology_route_conflict category: topology_route_conflict max_severity: high statement: "A topology, route, NAT, gateway, firewall, or security-state change overlaps candidate fabric, flow, or compute activity and may change attribution or visibility." trigger: type: two_feature_route_topology_change_logic evaluation_window: sparse_candidate_with_route_or_topology_change_window conditions: all: - metric: topology_or_route_change_overlap_fraction op: ">=" value_ref: min_sparse_change_overlap_fraction - metric: visibility_or_participant_impact_score op: ">=" value_ref: min_sparse_visibility_impact_score any: - metric: fabric_or_flow_coverage_fraction op: "<" value_ref: min_sparse_fabric_or_flow_coverage_fraction - metric: participant_churn_fraction op: ">=" value_ref: min_sparse_participant_churn_fraction missing_data_behavior: "If change records are unavailable, do not clear fabric or flow discrepancies; emit `topology_change_coverage_unknown`." severity_mapping: low: "Change overlaps candidate window but measured impact is weak." medium: "Change overlaps and affects participant mapping, path, or visibility." high: "Change overlaps a medium/high sparse candidate and coincides with coverage loss or remap without maintenance explanation." benign_explanation_or_suppressor_fields: - documented_network_maintenance - failure_recovery - planned_network_or_capacity_change - monitoring_path_change calibration_defaults: - name: min_sparse_change_overlap_fraction value: 0.10 unit: fraction evidence_status: calibration_default source_refs: [S25, S28, S29, S30, S42, S43] rationale: "Topology and route events are discrete; even partial overlap can affect attribution." counterevidence_or_failure_modes: - "event timestamps can be imprecise" calibration_note: "Tune by topology and network event time accuracy." - name: min_sparse_visibility_impact_score value: 0.50 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S25, S28, S29, S30, S42, S43] rationale: "Only changes that plausibly affect mapping or visibility should escalate." counterevidence_or_failure_modes: - "impact may be unknown or delayed" calibration_note: "Calibrate from network operations records." - name: min_sparse_fabric_or_flow_coverage_fraction value: 0.75 unit: fraction evidence_status: source_informed source_refs: [S28, S29, S30, S42, S43] rationale: "Flow or fabric visibility below this during a route/topology change weakens attribution." counterevidence_or_failure_modes: - "provider flow logs can be sampled, skipped, or delayed" calibration_note: "Replace with fabric and flow source coverage expectations." - name: min_sparse_participant_churn_fraction value: 0.20 unit: fraction evidence_status: calibration_default source_refs: [S18, S20, S25, S42, S43] rationale: "Material participant churn during candidate windows is attribution-relevant." counterevidence_or_failure_modes: - "elastic jobs and failures can churn normally" calibration_note: "Tune by scheduler and topology behavior." required_output_fields: - change_event_time - affected_scope - candidate_window_overlap - visibility_impact_score - participant_churn_fraction - explanation_status false_positive_or_limitations: - "Route changes can be benign but still alter telemetry interpretation." source_refs: [S18, S20, S25, S28, S29, S30, S42, S43] - id: fabric_or_storage_activity_explained_by_storage_operation purpose: sparse_storage_or_fabric_explanation raw_feature_budget: 2 required_feature_sets: any_one_of: - [scaleout_port_tx_rx_bytes_packets, storage_operation_intervals] - [fabric_port_device_sample_counters, storage_operation_intervals] - [east_west_internal_flow_bytes, storage_operation_intervals] - [data_center_interconnect_wan_usage, storage_operation_intervals] - [vpc_or_nsg_flow_records, storage_operation_intervals] - [storage_write_operation_bytes, storage_operation_intervals] - [storage_read_operation_bytes, storage_operation_intervals] - [storage_write_operation_bytes, object_or_file_copy_movement_records] derived_signals: - storage_staging_window emitted_statement: label: storage_operation_explanation_screen category: storage_activity_conflict max_severity: medium statement: "High fabric, flow, read, or write activity overlaps an explicit storage operation that may explain bytes otherwise mistaken for collective, staging, or checkpoint evidence." trigger: type: two_feature_storage_explanation_logic evaluation_window: storage_activity_operation_overlap_window conditions: all: - metric: storage_or_fabric_activity_score op: ">=" value_ref: min_sparse_storage_or_fabric_activity_score - metric: storage_operation_type op: in value_ref: explanatory_storage_operation_types - metric: storage_operation_overlap_fraction op: ">=" value_ref: min_sparse_storage_operation_overlap_fraction - metric: bytes_explained_fraction op: ">=" value_ref: min_sparse_bytes_explained_fraction missing_data_behavior: "If storage-operation labels are missing, emit `explanation_unavailable` rather than suppressing training evidence." severity_mapping: info: "Storage operation explains most suspicious bytes." low: "Storage operation partly explains bytes; residual activity remains." medium: "Storage operation overlaps but byte attribution is weak or mixed." benign_explanation_or_suppressor_fields: - backup - restore - replication - rebuild - migration - disaster_recovery - etl_export calibration_defaults: - name: min_sparse_storage_or_fabric_activity_score value: 0.50 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S28, S29, S31, S32, S33, S34, S35, S42, S43] rationale: "Explanation check applies to material normalized fabric, flow, read, or write activity." counterevidence_or_failure_modes: - "small but meaningful collective or checkpoint bursts can be below this" calibration_note: "Calibrate by storage, fabric, and flow baselines." - 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: "Explicit operation types can explain storage-driven network and write traffic." counterevidence_or_failure_modes: - "operation labels may be broad or overlap true training" calibration_note: "Normalize provider and storage labels." - name: min_sparse_storage_operation_overlap_fraction value: 0.80 unit: fraction evidence_status: calibration_default source_refs: [S31, S32, S33, S34, S35] rationale: "Storage operation should overlap most suspicious activity to explain it." counterevidence_or_failure_modes: - "coarse operation intervals can over-explain" calibration_note: "Tune by operation interval precision." - name: min_sparse_bytes_explained_fraction value: 0.70 unit: fraction evidence_status: calibration_default source_refs: [S31, S32, S33, S34, S35] rationale: "Most bytes should be explainable before lowering training suspicion." counterevidence_or_failure_modes: - "shared links can mix storage and accelerator traffic" calibration_note: "Use route/path and byte-attribution calibration." required_output_fields: - suspicious_activity_window - storage_operation_type - activity_score - explained_bytes - unexplained_residual - effect_on_candidate_label false_positive_or_limitations: - "Storage operations explain only overlapping and attributable traffic." source_refs: [S28, S29, S31, S32, S33, S34, S35, S42, S43] - id: high_activity_during_degraded_or_throttled_state_conflict purpose: sparse_health_throttle_conflict raw_feature_budget: 2 required_feature_sets: any_one_of: - [accelerator_health_error_state, tensor_matrix_mxu_neuron_or_engine_active_fraction] - [accelerator_health_error_state, generic_achieved_operation_rate] - [accelerator_health_error_state, accelerator_busy_or_utilization_fraction] - [accelerator_throttle_state, tensor_matrix_mxu_neuron_or_engine_active_fraction] - [accelerator_performance_limit_state, generic_achieved_operation_rate] - [accelerator_power_limit_active, generic_achieved_operation_rate] - [accelerator_thermal_limit_active, tensor_matrix_mxu_neuron_or_engine_active_fraction] - [power_cap_violation_throttle_state, generic_achieved_operation_rate] - [local_link_error_recovery_throttle, scaleout_port_tx_rx_bytes_packets] - [rdma_congestion_retry_drop_error_counters, scaleout_port_tx_rx_bytes_packets] derived_signals: - cap_clock_power_factor - available_accelerator_fraction - telemetry_coverage_fraction emitted_statement: label: health_throttle_conflict category: health_throttle_conflict max_severity: critical statement: "High observed activity, achieved operations, or fabric activity overlaps degraded, throttled, power-limited, thermal-limited, or link-degraded state that should reduce credible capacity." trigger: type: two_feature_degraded_activity_logic evaluation_window: degraded_or_limited_activity_overlap_window conditions: all: - metric: sparse_activity_or_fabric_score op: ">=" value_ref: min_sparse_degraded_conflict_activity_score - metric: degraded_or_limited_affected_fraction op: ">=" value_ref: min_sparse_degraded_affected_fraction - metric: degraded_activity_overlap_fraction op: ">=" value_ref: min_sparse_degraded_activity_overlap_fraction any: - metric: cap_clock_power_factor.cap_clock_power_factor op: "<=" value_ref: max_sparse_cap_clock_power_factor_for_conflict - metric: achieved_ops_to_degraded_capacity_ratio op: ">" value_ref: max_sparse_degraded_capacity_ratio missing_data_behavior: "If health or throttle coverage is missing, emit coverage caveat and route to telemetry-gap review rather than clearing high activity." severity_mapping: medium: "High activity overlaps degraded/limited state but affected fraction or capacity conflict is modest." high: "High activity conflicts with degraded adjusted capacity over most of the sparse window." critical: "Achieved operations exceed degraded cap-adjusted capacity by > 25 percent with high coverage." benign_explanation_or_suppressor_fields: - partial_device_set_limitation - transient_low_materiality_limit - counter_scope_mismatch - automatic_recovery calibration_defaults: - name: min_sparse_degraded_conflict_activity_score value: 0.70 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S07, S08, S36, S37, S38, S39, S42, S43] rationale: "Conflict requires clear high activity, achieved operations, or fabric activity." counterevidence_or_failure_modes: - "activity or fabric scores can be inflated by proxy estimates" calibration_note: "Calibrate by device, link, and counter source." - name: min_sparse_degraded_affected_fraction value: 0.20 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09, S42, S43] rationale: "A material fraction of devices or links must be degraded for a window-level conflict." counterevidence_or_failure_modes: - "one critical link can affect more than its fraction suggests" calibration_note: "Tune by topology and failure impact." - name: min_sparse_degraded_activity_overlap_fraction value: 0.50 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09, S42, S43] rationale: "Degraded state should overlap material activity." counterevidence_or_failure_modes: - "short severe events can still matter" calibration_note: "Tune by telemetry interval and failure duration." - name: max_sparse_cap_clock_power_factor_for_conflict value: 0.80 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09] rationale: "A 20 percent or larger capacity reduction should be considered in sparse high-activity consistency checks." counterevidence_or_failure_modes: - "some limit states are non-binding" calibration_note: "Calibrate with per-SKU cap/performance data." - name: max_sparse_degraded_capacity_ratio value: 1.10 unit: ratio evidence_status: calibration_default source_refs: [S04, S05, S06, S07, S08, S09, S24] rationale: "Allows uncertainty while flagging activity above degraded adjusted capacity." counterevidence_or_failure_modes: - "capacity model can be conservative or unit-mismatched" calibration_note: "Replace with cap-adjusted uncertainty model." required_output_fields: - degraded_window - activity_window - affected_fraction - capacity_reduction_factor - exceedance_ratio - coverage_summary false_positive_or_limitations: - "Limit and health states are source-specific and may be transient." source_refs: [S04, S05, S06, S07, S08, S09, S24, S36, S37, S38, S39, S42, S43]