schema_version: "0.1" created: "2026-05-16" source_inventory: observables/observables.yaml status: gated_aggregation_rule_spec purpose: > Executable staged aggregation layer for joining sparse rule outputs into algorithm-facing capacity, warning-height training-like, suppressor, discrepancy, and final-routing claims. Aggregation consumes sparse rule categories and labels by default; raw feature interpretation remains in the active sparse rule files. source_and_calibration_policy: online_research_checked: "2026-05-16" source_ledger: observables/rules/source_ledger.yaml notes: - "S01 confirms the 1.0e25 FLOP policy-scale threshold; this is scale, not workload identity." - "S07-S09 and S36-S39 support accelerator/activity telemetry surfaces but not portable training-identity thresholds." - "S11-S17 support collective, sharded, and distributed-training communication mechanisms, while also showing why HPC/MPI and benchmark alternatives must be checked." - "S28-S35 support flow-log, storage-log, and billing coverage/delay caveats; missing or delayed channels must not be treated as cleared suppressors." - "S45-S47 support checkpoint and distributed-checkpoint storage behavior, but checkpoint-like writes can also be storage maintenance, copy, backup, restore, or export activity." aggregation_policy: stage_order: - A_capacity_gate - B_training_candidate_detection - C_discrepancy_and_explanation_review - final_claim_routing scope_join_policy: default_scope_match: same_monitored_scope_or_mapped_child_scope child_scope_rollup: "Roll up child scopes only when the mapping coverage is explicit for the candidate window." no_hidden_capacity_policy: "Hidden, unknown, or out-of-scope capacity prevents strong negative certification." window_segmentation_policy: segmentation_signal: capacity_segment_window split_at: - inventory_validity_boundary - advertised_peak_rate_validity_boundary - quota_or_reservation_boundary - running_or_billing_interval_boundary - partition_interval_boundary - maintenance_interval_boundary - health_degraded_or_throttle_state_boundary - electrical_service_or_installation_boundary - topology_domain_or_route_change_boundary - clock_drift_correction_boundary short_circuit_policy: stop_segment_if: - label: capacity_ruled_out_for_scope confidence: conservative_bound min_coverage_fraction: value_ref: min_capacity_ruleout_coverage_fraction do_not_run_stages: - B_training_candidate_detection - C_discrepancy_and_explanation_review rationale: "A high-coverage conservative upper bound below the policy threshold rules out threshold-scale compute for the monitored scope/window." do_not_stop_if: - capacity_unknown_due_to_missing_inputs - capacity_limited_but_not_ruled_out - capacity_possible_for_scope - capacity_possible_under_sparse_assumptions missing_data_policy: general: "Missing primary, suppressor, coverage, or clock-alignment data emits an explicit missingness or inconclusive claim; it is not interpreted as absence." suppressor_channels: "Missing expected suppressor channels block high warning heights." final_warning_heights: ordered_values: - none - weak_training_like_candidate - medium_training_like_warning - high_training_like_warning hard_constraints: - "Activity alone cannot produce any training-like warning above weak candidate routing." - "Fabric evidence alone cannot produce a training-like warning." - "Storage writes alone cannot produce a training-like warning." - "Large compute alone emits large_compute_candidate or large_compute_training_identity_unresolved, not training." - "High warning requires capacity possible/unknown, accelerator activity or achieved operations, at least two independent aligned B/supporting identity categories, checked suppressors, and no unresolved strong discrepancy." global_calibration_defaults: - name: min_capacity_ruleout_coverage_fraction value: 0.90 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09, S18, S23, S26, S35, S40, S41] rationale: "Negative capacity claims require high coverage of inventory, service, availability, and peak-rate inputs." counterevidence_or_failure_modes: - "missing inventory, peak rate, electrical service, cloud scope, or hidden capacity can invalidate a rule-out" calibration_note: "Replace with source-specific completeness guarantees for the monitored scope." - name: policy_threshold_operations value: 1.0e+25 unit: operations evidence_status: source_backed source_refs: [S01, S03] rationale: "Policy-scale compute threshold used for capacity and large-compute review; it is not workload identity." counterevidence_or_failure_modes: - "large inference, synthetic data generation, HPC, benchmarks, and burn-in can cross the same scale" calibration_note: "Replace only if governing policy threshold changes." aggregation_rules: - id: capacity_gate_segment_ruleout purpose: capacity_gate stage: A_capacity_gate consumes: rule_output_categories: all: - capacity_policy_scale any: - capacity_count_or_shape - capacity_peak_rate - capacity_availability - capacity_power_service - capacity_topology derived_signals: - capacity_segment_window - capacity_upper_bound_flop - telemetry_coverage_fraction - clock_aligned_overlap_window time_alignment: segmentation: capacity_segment_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_capacity_coverage_fraction: value_ref: min_capacity_ruleout_coverage_fraction clock_alignment_confidence: value_ref: min_capacity_clock_alignment_confidence missing_data_behavior: capacity_unknown_due_to_missing_inputs short_circuit_behavior: "On emitted capacity_ruled_out_for_scope, stop B and C for this capacity segment." trigger: type: aggregation_gate_logic conditions: all: - metric: capacity_upper_bound_flop.unit_normalized op: "==" value: true - metric: capacity_upper_bound_flop.capacity_upper_bound_operations op: "<" value_ref: policy_threshold_operations - metric: capacity_upper_bound_flop.coverage_fraction op: ">=" value_ref: min_capacity_ruleout_coverage_fraction - metric: clock_aligned_overlap_window.clock_alignment_confidence op: ">=" value_ref: min_capacity_clock_alignment_confidence none: - metric: capacity_upper_bound_flop.hidden_or_unmonitored_capacity_possible op: "==" value: true - metric: capacity_upper_bound_flop.missing_required_peak_rate op: "==" value: true emitted_claim: label: capacity_ruled_out_for_scope stage_result: stop_segment max_confidence: conservative_bound warning_height: none statement: "Threshold-scale compute is conservatively ruled out for the monitored scope/window." missing_data_behavior: "Emit capacity_unknown_due_to_missing_inputs when coverage, unit normalization, scope mapping, clock alignment, peak rate, inventory, or service state is insufficient." cannot_conclude: - training - no_non_threshold_training - no_hidden_out_of_scope_run calibration_defaults: - 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 workload identity." counterevidence_or_failure_modes: - "threshold-scale compute can be non-training" calibration_note: "Replace only if governing policy threshold changes." - name: min_capacity_ruleout_coverage_fraction value: 0.90 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09, S18, S23, S26, S35, S40, S41] rationale: "Strong rule-out requires high capacity input coverage." counterevidence_or_failure_modes: - "missing inventory, peak rate, service, quota, or hidden capacity can invalidate rule-out" calibration_note: "Replace with source-specific coverage requirements." - name: min_capacity_clock_alignment_confidence value: 0.80 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S18, S21, S22, S23, S26, S28, S29, S35] rationale: "Capacity validity intervals and source records must align before a negative conclusion is trusted." counterevidence_or_failure_modes: - "clock drift or delayed cloud/provider records can shift apparent capacity boundaries" calibration_note: "Replace with measured clock drift and source delivery-delay bounds." - id: capacity_gate_live_segment_classification purpose: capacity_gate stage: A_capacity_gate consumes: rule_output_categories: any: - capacity_count_or_shape - capacity_peak_rate - capacity_memory - capacity_topology - capacity_availability - capacity_power_service - capacity_policy_scale derived_signals: - capacity_segment_window - capacity_upper_bound_flop - telemetry_coverage_fraction time_alignment: segmentation: capacity_segment_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_overlap_fraction: value_ref: min_capacity_live_overlap_fraction missing_data_behavior: capacity_unknown_due_to_missing_inputs short_circuit_behavior: "Never short-circuit B/C from this rule; live segments continue to B." aggregation_logic: ordered_classification: - output: capacity_possible_for_scope when: all: - metric: capacity_upper_bound_flop.capacity_upper_bound_operations op: ">=" value_ref: policy_threshold_operations - metric: capacity_upper_bound_flop.coverage_fraction op: ">=" value_ref: min_capacity_possible_coverage_fraction - output: capacity_limited_but_not_ruled_out when: all: - category: capacity_availability - metric: capacity_upper_bound_flop.coverage_fraction op: "<" value_ref: min_capacity_ruleout_coverage_fraction any: - category: capacity_count_or_shape - category: capacity_peak_rate - category: capacity_power_service - output: capacity_unknown_due_to_missing_inputs when: any: - missing_category: capacity_count_or_shape - missing_category: capacity_peak_rate - metric: telemetry_coverage_fraction.primary_channel_min op: "<" value_ref: min_capacity_possible_coverage_fraction emitted_claim: label: capacity_segment_live_for_B possible_outputs: - capacity_possible_for_scope - capacity_limited_but_not_ruled_out - capacity_unknown_due_to_missing_inputs stage_result: continue_segment max_confidence: weak statement: "The capacity segment remains live for B candidate detection unless separately ruled out by a conservative high-coverage upper bound." missing_data_behavior: "Emit capacity_unknown_due_to_missing_inputs and keep the segment live." cannot_conclude: - training - capacity_ruled_out_for_scope calibration_defaults: - name: policy_threshold_operations value: 1.0e+25 unit: operations evidence_status: source_backed source_refs: [S01, S03] rationale: "Policy-scale compute threshold for feasibility review." counterevidence_or_failure_modes: - "policy threshold does not classify workload type" calibration_note: "Replace only if governing policy threshold changes." - name: min_capacity_possible_coverage_fraction value: 0.60 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09, S18, S23, S26, S35] rationale: "Live capacity classification can proceed at lower coverage than a negative rule-out because missingness keeps B/C alive." counterevidence_or_failure_modes: - "low coverage can overstate or understate available capacity" calibration_note: "Tune by source coverage and audit risk tolerance." - name: min_capacity_live_overlap_fraction value: 0.50 unit: fraction evidence_status: calibration_default source_refs: [S18, S23, S26, S28, S29, S35] rationale: "At least half-window overlap prevents stale capacity facts from classifying a live segment." counterevidence_or_failure_modes: - "short but important capacity changes can be diluted by this default" calibration_note: "Replace with source-specific validity interval logic." - id: large_compute_candidate_routing purpose: large_compute_candidate_aggregation stage: B_training_candidate_detection consumes: rule_output_categories: any: - large_compute_scale - accelerator_activity_run_existence - capacity_policy_scale derived_signals: - rule_output_join_window - aggregation_candidate_seed_window - achieved_operation_integral - capacity_upper_bound_flop - telemetry_coverage_fraction time_alignment: segmentation: aggregation_candidate_seed_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_overlap_fraction: value_ref: min_large_compute_capacity_overlap_fraction missing_data_behavior: large_compute_training_identity_unresolved allowed_adjacency_seconds: value_ref: max_achieved_ops_capacity_adjacency_seconds trigger: type: large_compute_scale_logic conditions: all: - metric: achieved_operation_integral.operation_count op: ">=" value_ref: policy_threshold_operations - metric: achieved_operation_integral.coverage_fraction op: ">=" value_ref: min_large_compute_coverage_fraction any: - category: large_compute_scale - category: accelerator_activity_run_existence emitted_claim: label: large_compute_candidate alternate_outputs: - large_compute_capacity_conflict - large_compute_training_identity_unresolved warning_height: none max_confidence: medium statement: "Achieved or inferred operations reach policy scale, but workload identity remains unresolved until joined with independent training-like evidence." demotion_policy: capacity_conflict: "If achieved_operation_integral exceeds capacity_upper_bound_flop by the configured ratio, emit large_compute_capacity_conflict and route to C." no_identity_evidence: "If no fabric/checkpoint/non-serving identity-shape category aligns, emit large_compute_training_identity_unresolved." missing_data_behavior: "If operation units, counter resets, or achieved-counter coverage are unresolved, route to C and suppress training-like promotion." cannot_conclude: - training - distributed_training - checkpoint_training calibration_defaults: - name: policy_threshold_operations value: 1.0e+25 unit: operations evidence_status: source_backed source_refs: [S01, S03] rationale: "Large-compute policy threshold." counterevidence_or_failure_modes: - "large compute may be inference, HPC, synthetic-data generation, benchmarks, or burn-in" calibration_note: "Replace only if governing policy threshold changes." - name: min_large_compute_coverage_fraction value: 0.75 unit: fraction evidence_status: source_informed source_refs: [S07, S08, S36, S37, S38, S39] rationale: "Operation integrals require substantial achieved-counter or calibrated proxy coverage." counterevidence_or_failure_modes: - "counter gaps, unit mismatch, or proxy calibration can distort operation counts" calibration_note: "Replace with counter-source reliability and unit-normalization tests." - name: min_large_compute_capacity_overlap_fraction value: 0.70 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09, S23, S24, S26, S35] rationale: "Achieved operations and capacity must mostly describe the same scope/window before conflict or feasibility routing." counterevidence_or_failure_modes: - "billing, running, and achieved-counter windows can lag or differ in scope" calibration_note: "Tune by provider/accounting and counter aggregation behavior." - name: max_achieved_ops_capacity_adjacency_seconds value: 600 unit: seconds evidence_status: calibration_default source_refs: [S18, S23, S26, S28, S35] rationale: "Allows short accounting, aggregation, and clock-boundary mismatches without merging unrelated windows." counterevidence_or_failure_modes: - "provider logs can be delayed by more than this; long jobs split at hour boundaries may need source-specific handling" calibration_note: "Replace with measured source delivery and boundary behavior." - id: distributed_training_like_candidate_promotion purpose: training_candidate_promotion stage: B_training_candidate_detection consumes: required_capacity_outputs: any: - capacity_possible_for_scope - capacity_limited_but_not_ruled_out - capacity_unknown_due_to_missing_inputs rule_output_categories: all: - accelerator_activity_run_existence - fabric_parameter_update_like optional: - non_serving_workload_shape - topology_context_support - physical_activity_support derived_signals: - rule_output_join_window - aggregation_candidate_seed_window - warning_height_score - suppression_resolution_state - telemetry_coverage_fraction - clock_aligned_overlap_window time_alignment: segmentation: aggregation_candidate_seed_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_overlap_fraction: value_ref: min_activity_identity_overlap_fraction minimum_coverage_fraction: value_ref: min_primary_B_channel_coverage_fraction missing_data_behavior: "candidate_requires_manual_review if suppressor or primary coverage is below threshold." allowed_adjacency_seconds: value_ref: max_fabric_activity_adjacency_seconds trigger: type: independent_category_promotion_logic conditions: all: - capacity_output: any: - capacity_possible_for_scope - capacity_limited_but_not_ruled_out - capacity_unknown_due_to_missing_inputs - category: accelerator_activity_run_existence - category: fabric_parameter_update_like - metric: rule_output_join_window.overlap_fraction op: ">=" value_ref: min_activity_identity_overlap_fraction - metric: telemetry_coverage_fraction.primary_channel_min op: ">=" value_ref: min_primary_B_channel_coverage_fraction none: - label: capacity_ruled_out_for_scope - suppression_state: strong_unresolved_suppressor emitted_claim: label: distributed_training_like_candidate warning_height_by_score: weak_training_like_candidate: "activity plus fabric identity shape align, but support or suppressor review is incomplete." medium_training_like_warning: "activity plus fabric identity shape align with checked suppressors and at least one support or non-serving category." high_training_like_warning: "activity plus fabric identity shape align with a second independent identity/support category, checked suppressors, high coverage, and no unresolved discrepancy." max_confidence: medium statement: "Aligned accelerator activity and collective-like fabric behavior form a distributed training-like candidate; final warning height depends on suppressor and discrepancy review." suppression_policy: checked_suppressors: - serving_counterevidence - benchmark_hpc_alternative - storage_operation_explanation - topology_route_conflict missing_suppressor_behavior: "Cap at weak_training_like_candidate and emit inconclusive_due_to_missingness." strong_suppressor_behavior: "Emit candidate_demoted_by_unresolved_suppressor or candidate_benchmark_like." missing_data_behavior: "Missing fabric/activity coverage, clock alignment, or suppressor state prevents medium/high warning." cannot_conclude: - high_training_like_warning_from_activity_alone - high_training_like_warning_from_fabric_alone - final_training_claim_without_C_when_suppressors_present calibration_defaults: - name: min_activity_identity_overlap_fraction value: 0.60 unit: fraction evidence_status: calibration_default source_refs: [S11, S12, S13, S14, S16, S17, S28, S29] rationale: "Compute and fabric phases should substantially overlap for distributed training-like promotion, while allowing aggregation and phase lag." counterevidence_or_failure_modes: - "asynchronous prefetch, gradient accumulation, coarse flow logs, HPC, and distributed inference can change overlap shape" calibration_note: "Calibrate with local step traces, flow-log intervals, and known non-training workloads." - name: min_primary_B_channel_coverage_fraction value: 0.75 unit: fraction evidence_status: source_informed source_refs: [S07, S08, S28, S29, S36, S37, S38, S39, S42, S43] rationale: "B candidate promotion needs substantial coverage of activity and identity-shape channels." counterevidence_or_failure_modes: - "sampled flow logs or missing device counters can overstate or understate alignment" calibration_note: "Replace with per-source coverage and sampling guarantees." - name: max_fabric_activity_adjacency_seconds value: 300 unit: seconds evidence_status: calibration_default source_refs: [S11, S12, S13, S14, S28, S29, S42, S43] rationale: "Allows coarse fabric aggregation and phase relation without joining unrelated runs." counterevidence_or_failure_modes: - "long checkpoint or communication phases may sit outside this default; coarse logs may require wider windows" calibration_note: "Tune by fabric telemetry aggregation interval and clock drift." - id: checkpoint_training_like_candidate_promotion purpose: training_candidate_promotion stage: B_training_candidate_detection consumes: required_capacity_outputs: any: - capacity_possible_for_scope - capacity_limited_but_not_ruled_out - capacity_unknown_due_to_missing_inputs rule_output_categories: all: - accelerator_activity_run_existence - storage_checkpoint_state_like optional: - data_lifecycle_support - lifecycle_operational_context - physical_activity_support derived_signals: - aggregation_candidate_seed_window - rule_output_join_window - warning_height_score - suppression_resolution_state - checkpoint_periodicity_score - telemetry_coverage_fraction time_alignment: segmentation: aggregation_candidate_seed_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_adjacency_fraction: value_ref: min_checkpoint_activity_adjacency_fraction allowed_adjacency_seconds: value_ref: max_checkpoint_activity_adjacency_seconds minimum_coverage_fraction: value_ref: min_primary_B_channel_coverage_fraction missing_data_behavior: "candidate_requires_manual_review if storage-operation explanations are unavailable." trigger: type: adjacent_identity_promotion_logic conditions: all: - capacity_output: any: - capacity_possible_for_scope - capacity_limited_but_not_ruled_out - capacity_unknown_due_to_missing_inputs - category: accelerator_activity_run_existence - category: storage_checkpoint_state_like - metric: rule_output_join_window.adjacency_or_overlap_fraction op: ">=" value_ref: min_checkpoint_activity_adjacency_fraction - metric: telemetry_coverage_fraction.primary_channel_min op: ">=" value_ref: min_primary_B_channel_coverage_fraction none: - label: capacity_ruled_out_for_scope emitted_claim: label: checkpoint_training_like_candidate warning_height_by_score: weak_training_like_candidate: "activity and checkpoint-like storage are adjacent or overlapping, but explanation review is incomplete." medium_training_like_warning: "activity and repeated checkpoint-like storage align with checked storage-operation explanations." high_training_like_warning: "checkpoint-like storage aligns with activity plus another independent identity channel, checked suppressors, high coverage, and no unresolved discrepancy." max_confidence: medium statement: "Aligned accelerator activity and checkpoint-like state writes form a training-like checkpoint candidate; storage-operation explanations must be adjudicated before promotion." suppression_policy: checked_suppressors: - storage_operation_explanation - storage_activity_conflict - serving_counterevidence - lifecycle_operational_context missing_suppressor_behavior: "Emit inconclusive_due_to_missingness and cap at weak_training_like_candidate." strong_suppressor_behavior: "Emit candidate_explained_by_storage_operation." missing_data_behavior: "Missing storage-operation channels block medium/high warning rather than clearing storage explanations." demotion_policy: storage_operation_explanation_present: "Do not block candidate emission in B; attach the explanation to suppression_resolution_state and route to C for demotion or explanation." cannot_conclude: - high_training_like_warning_from_storage_writes_alone - final_training_claim_without_storage_explanation_review calibration_defaults: - name: min_checkpoint_activity_adjacency_fraction value: 0.50 unit: fraction evidence_status: calibration_default source_refs: [S45, S46, S47] rationale: "Checkpoint writes can lag or trail compute, so adjacency can be weaker than same-second overlap." counterevidence_or_failure_modes: - "backup, restore, replication, ETL, and synthetic-data output can look checkpoint-like" calibration_note: "Calibrate against local checkpoint cadence and storage-operation logs." - name: max_checkpoint_activity_adjacency_seconds value: 1800 unit: seconds evidence_status: calibration_default source_refs: [S31, S32, S33, S45, S46, S47] rationale: "Allows delayed checkpoint flushes and object/filesystem logging lag without joining unrelated storage jobs by default." counterevidence_or_failure_modes: - "large asynchronous checkpoint systems may flush longer; cloud storage logs can be delayed or best-effort" calibration_note: "Replace with framework, storage system, and log-delivery measurements." - name: min_primary_B_channel_coverage_fraction value: 0.75 unit: fraction evidence_status: source_informed source_refs: [S07, S08, S31, S32, S33, S36, S37, S38, S39] rationale: "B candidate promotion needs substantial activity and storage-channel coverage." counterevidence_or_failure_modes: - "object logs and filesystem metrics may be delayed, sampled, or not enabled" calibration_note: "Replace with source-specific coverage guarantees." - id: sparse_large_compute_training_like_candidate_promotion purpose: training_candidate_promotion stage: B_training_candidate_detection consumes: required_prior_outputs: all: - large_compute_candidate any: - distributed_training_like_candidate - checkpoint_training_like_candidate rule_output_categories: any: - fabric_parameter_update_like - storage_checkpoint_state_like - non_serving_workload_shape derived_signals: - aggregation_candidate_seed_window - warning_height_score - suppression_resolution_state - achieved_operation_integral time_alignment: segmentation: aggregation_candidate_seed_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_overlap_fraction: value_ref: min_large_compute_identity_overlap_fraction missing_data_behavior: large_compute_training_identity_unresolved short_circuit_behavior: "Do not run this rule in capacity_ruled_out_for_scope segments." trigger: type: large_compute_plus_identity_logic conditions: all: - label: large_compute_candidate - metric: achieved_operation_integral.operation_count op: ">=" value_ref: policy_threshold_operations - metric: suppression_resolution_state.strong_unresolved_suppressor_count op: "==" value: 0 any: - label: distributed_training_like_candidate - label: checkpoint_training_like_candidate emitted_claim: label: sparse_large_compute_training_like_candidate warning_height_by_score: medium_training_like_warning: "large compute plus one independent training-like identity channel with checked suppressors." high_training_like_warning: "large compute plus two independent aligned identity/support channels, high coverage, checked suppressors, and no unresolved C conflict." max_confidence: medium statement: "Policy-scale compute plus independent training-like identity evidence forms a sparse large-compute training-like candidate." suppression_policy: checked_suppressors: - serving_counterevidence - storage_operation_explanation - benchmark_hpc_alternative - activity_attribution_conflict - capacity_claim_conflict missing_suppressor_behavior: "Emit large_compute_training_identity_unresolved or inconclusive_due_to_missingness." missing_data_behavior: "Large compute without identity or suppressor resolution remains unresolved." cannot_conclude: - training_from_large_compute_alone - high_warning_with_unresolved_strong_suppressor calibration_defaults: - name: policy_threshold_operations value: 1.0e+25 unit: operations evidence_status: source_backed source_refs: [S01, S03] rationale: "Large-compute policy threshold." counterevidence_or_failure_modes: - "threshold crossing is not workload identity" calibration_note: "Replace only if governing policy threshold changes." - name: min_large_compute_identity_overlap_fraction value: 0.50 unit: fraction evidence_status: calibration_default source_refs: [S11, S12, S13, S14, S16, S17, S45, S46, S47] rationale: "Large achieved-operation windows can be broader than identity-shape windows, so half-window overlap is enough for candidate routing." counterevidence_or_failure_modes: - "short identity bursts can over-promote broad compute windows; storage and benchmark alternatives remain important" calibration_note: "Tune by job-level known workload traces." - id: support_context_aggregation purpose: support_context_aggregation stage: B_training_candidate_detection consumes: required_prior_outputs: any: - distributed_training_like_candidate - checkpoint_training_like_candidate - sparse_large_compute_training_like_candidate rule_output_categories: any: - physical_activity_support - large_model_context_support - data_lifecycle_support - topology_context_support - lifecycle_operational_context derived_signals: - aggregation_candidate_review_window - rule_output_join_window - warning_height_score time_alignment: segmentation: aggregation_candidate_review_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_overlap_fraction: value_ref: min_support_candidate_overlap_fraction allowed_adjacency_seconds: value_ref: max_support_candidate_adjacency_seconds missing_data_behavior: "Support absence does not demote by itself; missing expected support coverage is reported." aggregation_logic: outputs: - label: candidate_context_strengthened when: "A core B candidate already exists and support overlaps or is adjacent within the configured window." - label: candidate_context_weak_only when: "Only weak support categories exist beside a candidate, without a second identity-shape category." - label: support_present_but_no_training_identity when: "Support categories exist without accelerator activity plus fabric/checkpoint identity evidence." emitted_claim: label: candidate_context_strengthened alternate_outputs: - candidate_context_weak_only - support_present_but_no_training_identity warning_height_effect: "May raise review priority or contribute to moving weak to medium only when core B evidence is already present and suppressors are checked; this changes routing warning height, not this support rule's standalone evidentiary confidence." max_confidence: weak statement: "Support context can strengthen or route a live candidate but is not primary workload identity evidence." missing_data_behavior: "Missing support is recorded but does not clear or prove training." cannot_conclude: - training_from_support_only - warning_height_from_power_only - warning_height_from_memory_only - warning_height_from_topology_only calibration_defaults: - name: min_support_candidate_overlap_fraction value: 0.40 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09, S31, S32, S33, S40, S41] rationale: "Support signals such as power, cooling, staging, and lifecycle events may lag or precede core activity." counterevidence_or_failure_modes: - "support windows can reflect unrelated physical load, data movement, or maintenance" calibration_note: "Calibrate by support-source lag and local workload traces." - name: max_support_candidate_adjacency_seconds value: 1800 unit: seconds evidence_status: calibration_default source_refs: [S31, S32, S33, S40, S41] rationale: "Allows thermal/cooling and storage lifecycle lag while keeping support local to the candidate." counterevidence_or_failure_modes: - "facility cooling response and storage staging can have longer lags in some sites" calibration_note: "Replace with site thermal, storage, and lifecycle timing distributions." - id: suppressor_and_explanation_aggregation purpose: suppressor_and_explanation_aggregation stage: C_discrepancy_and_explanation_review consumes: required_prior_outputs: any: - distributed_training_like_candidate - checkpoint_training_like_candidate - sparse_large_compute_training_like_candidate rule_output_categories: any: - serving_counterevidence - storage_operation_explanation - benchmark_hpc_alternative - storage_activity_conflict - topology_route_conflict - suppressors_alternative_explanations derived_signals: - aggregation_candidate_review_window - suppression_resolution_state - rule_output_join_window - telemetry_coverage_fraction time_alignment: segmentation: aggregation_candidate_review_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_overlap_fraction: value_ref: min_suppressor_candidate_overlap_fraction allowed_adjacency_seconds: value_ref: max_suppressor_candidate_adjacency_seconds missing_data_behavior: candidate_requires_manual_review short_circuit_behavior: "Do not run after capacity_ruled_out_for_scope unless reviewing a separate out-of-scope or integrity claim." aggregation_logic: demotion_outputs: - candidate_explained_by_serving - candidate_explained_by_storage_operation - candidate_benchmark_like - candidate_hpc_mpi_alternative - candidate_demoted_by_unresolved_suppressor missingness_output: inconclusive_due_to_missingness rules: - "Serving counterevidence overlapping accelerator activity demotes non-serving and fabric/checkpoint candidates." - "Storage operation explanations overlapping checkpoint-like or fabric/storage activity explain those candidates unless independent fabric and non-serving identity remains." - "Benchmark/HPC alternatives demote fabric-heavy candidates before high warning." - "Missing expected suppressor channels cap warning height at weak and require manual review." emitted_claim: label: candidate_demoted_by_unresolved_suppressor possible_outputs: - candidate_explained_by_serving - candidate_explained_by_storage_operation - candidate_benchmark_like - candidate_hpc_mpi_alternative - inconclusive_due_to_missingness warning_height_effect: "Demote or block medium/high promotion until explanation is resolved." max_confidence: not_applicable statement: "Suppressors and benign explanations are adjudicated before final warning-height promotion." missing_data_behavior: "Expected but missing suppressor channels emit inconclusive_due_to_missingness, not cleared suppressors." cannot_conclude: - cleared_suppressors_from_missing_channels - high_training_like_warning_with_unresolved_suppressor calibration_defaults: - name: min_suppressor_candidate_overlap_fraction value: 0.50 unit: fraction evidence_status: calibration_default source_refs: [S28, S29, S31, S32, S33, S45, S46, S47] rationale: "An explanation should overlap a material portion of the candidate window before demoting it." counterevidence_or_failure_modes: - "broad maintenance or backup windows can over-explain unrelated activity" calibration_note: "Tune by operation-log precision and workload-specific explanation coverage." - name: max_suppressor_candidate_adjacency_seconds value: 1800 unit: seconds evidence_status: calibration_default source_refs: [S28, S31, S32, S33] rationale: "Allows best-effort network/storage log delivery and adjacent storage operations without joining distant events." counterevidence_or_failure_modes: - "cloud logs can be delayed longer; high-throughput storage operations can span broader windows" calibration_note: "Replace with source delivery delay and operation-duration distributions." - id: discrepancy_targeted_escalation purpose: discrepancy_escalation_aggregation stage: C_discrepancy_and_explanation_review consumes: required_prior_outputs: any: - distributed_training_like_candidate - checkpoint_training_like_candidate - sparse_large_compute_training_like_candidate - large_compute_capacity_conflict - large_compute_training_identity_unresolved - candidate_requires_manual_review rule_output_categories: any: - activity_attribution_conflict - capacity_claim_conflict - physical_timeline_conflict - power_activity_conflict - fabric_mapping_conflict - storage_activity_conflict - telemetry_missingness_conflict - health_throttle_conflict - topology_route_conflict derived_signals: - aggregation_candidate_review_window - capacity_segment_window - rule_output_join_window - telemetry_coverage_fraction - suppression_resolution_state time_alignment: segmentation: aggregation_candidate_review_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_overlap_fraction: value_ref: min_discrepancy_candidate_overlap_fraction missing_data_behavior: inconclusive_due_to_missingness short_circuit_behavior: "Do not run globally after capacity_ruled_out_for_scope; run only for live candidates, capacity conflicts, or decision-blocking missingness." trigger: type: targeted_C_routing_logic conditions: any: - label: distributed_training_like_candidate - label: checkpoint_training_like_candidate - label: sparse_large_compute_training_like_candidate - label: large_compute_capacity_conflict - label: candidate_requires_manual_review - category: telemetry_missingness_conflict - category: capacity_claim_conflict emitted_claim: label: integrity_review_required possible_outputs: - activity_attribution_conflict - capacity_claim_conflict - inconclusive_due_to_missingness - candidate_requires_manual_review warning_height_effect: "Unresolved high/critical C findings block high warning and route to manual review." max_confidence: high statement: "Targeted C findings are escalated only when they adjudicate a live candidate, capacity conflict, or missingness that blocks a decision." missing_data_behavior: "Decision-blocking gaps emit inconclusive_due_to_missingness." cannot_conclude: - global_C_required_after_capacity_ruleout - high_warning_with_unresolved_capacity_conflict calibration_defaults: - name: min_discrepancy_candidate_overlap_fraction value: 0.50 unit: fraction evidence_status: calibration_default source_refs: [S18, S21, S22, S23, S26, S28, S29, S31, S35] rationale: "A discrepancy should affect a material part of the candidate or capacity-conflict window before blocking final routing." counterevidence_or_failure_modes: - "short severe conflicts may still matter; long audit windows can dilute brief manipulation or outages" calibration_note: "Tune by severity, source precision, and audit policy." - id: no_training_candidate_coverage_gate purpose: negative_screen_integrity stage: C_discrepancy_and_explanation_review consumes: required_capacity_outputs: any: - capacity_possible_for_scope - capacity_limited_but_not_ruled_out - capacity_unknown_due_to_missing_inputs excluded_prior_outputs: any: - distributed_training_like_candidate - checkpoint_training_like_candidate - sparse_large_compute_training_like_candidate - large_compute_candidate rule_output_categories: optional: - telemetry_missingness_conflict - activity_attribution_conflict - fabric_mapping_conflict derived_signals: - capacity_segment_window - aggregation_candidate_seed_window - negative_screen_required_coverage_state - telemetry_coverage_fraction - clock_aligned_overlap_window time_alignment: segmentation: capacity_segment_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_primary_coverage_fraction: value_ref: min_negative_screen_primary_coverage_fraction clock_alignment_confidence: value_ref: min_negative_screen_clock_alignment_confidence missing_data_behavior: inconclusive_due_to_missingness short_circuit_behavior: "Do not run after capacity_ruled_out_for_scope; this is only for live capacity segments where B found no candidate." global_C_policy: "Run coverage, scope, clock, and missingness checks needed to trust the negative screen; do not run unrelated suppressor or discrepancy checks globally." trigger: type: negative_screen_integrity_logic conditions: all: - capacity_output: any: - capacity_possible_for_scope - capacity_limited_but_not_ruled_out - capacity_unknown_due_to_missing_inputs - metric: negative_screen_required_coverage_state.primary_activity_coverage_fraction op: ">=" value_ref: min_negative_screen_primary_coverage_fraction - metric: negative_screen_required_coverage_state.identity_shape_coverage_fraction op: ">=" value_ref: min_negative_screen_identity_coverage_fraction - metric: negative_screen_required_coverage_state.scope_mapping_coverage_fraction op: ">=" value_ref: min_negative_screen_scope_mapping_coverage_fraction - metric: negative_screen_required_coverage_state.clock_alignment_confidence op: ">=" value_ref: min_negative_screen_clock_alignment_confidence none: - label: distributed_training_like_candidate - label: checkpoint_training_like_candidate - label: sparse_large_compute_training_like_candidate - label: large_compute_candidate - metric: negative_screen_required_coverage_state.decision_blocking_missingness op: "==" value: true emitted_claim: label: negative_screen_coverage_sufficient alternate_outputs: - negative_screen_blocked_by_missingness - inconclusive_due_to_missingness - integrity_review_required warning_height: none max_confidence: conservative_bound statement: "No B candidate was detected in a live segment and the required primary, identity-shape, scope-mapping, and clock-alignment coverage is sufficient to treat the absence of evidence as a covered negative screen." missing_data_behavior: "If required negative-screen channels are missing, emit negative_screen_blocked_by_missingness or inconclusive_due_to_missingness instead of a clean negative screen." cannot_conclude: - no_training_outside_monitored_scope - no_training_from_uncovered_channels - no_training_from_missing_activity_telemetry - global_C_clearance calibration_defaults: - name: min_negative_screen_primary_coverage_fraction value: 0.75 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S36, S37, S38, S39] rationale: "Reuses existing positive-candidate coverage defaults; no-candidate routing is trustworthy only when primary activity channels substantially cover the live segment." counterevidence_or_failure_modes: - "collector outages, disabled profiling counters, or unmapped devices can create false negative screens" calibration_note: "Replace with source-specific accelerator telemetry coverage requirements." - name: min_negative_screen_identity_coverage_fraction value: 0.75 unit: fraction evidence_status: calibration_default source_refs: [S28, S29, S31, S32, S33, S42, S43, S45, S46, S47] rationale: "Fabric/checkpoint identity channels must be substantially covered before their absence is meaningful." counterevidence_or_failure_modes: - "sampled flow logs, disabled object logs, or unmapped fabric ports can hide identity-shape evidence" calibration_note: "Replace with source-specific fabric and storage logging coverage." - name: min_negative_screen_scope_mapping_coverage_fraction value: 0.75 unit: fraction evidence_status: calibration_default source_refs: [S18, S19, S20, S23, S26, S35] rationale: "A negative screen must map telemetry to the monitored scope well enough to know which assets were covered." counterevidence_or_failure_modes: - "scheduler, billing, or provider scope mismatch can make active assets invisible to the monitored scope" calibration_note: "Replace with site-specific scheduler, provider, and asset-mapping coverage." - name: min_negative_screen_clock_alignment_confidence value: 0.80 unit: score_0_to_1 evidence_status: calibration_default source_refs: [S18, S21, S22, S27, S28, S29, S31, S35] rationale: "Negative screens need aligned source windows so absent activity is not an artifact of clock drift or delayed records." counterevidence_or_failure_modes: - "coarse aggregation, clock drift, or delayed logs can move evidence outside the checked window" calibration_note: "Replace with measured timestamp accuracy and delivery-delay bounds." - id: final_algorithm_claim_routing purpose: final_claim_routing stage: final_claim_routing consumes: required_prior_outputs: any: - capacity_ruled_out_for_scope - capacity_possible_for_scope - capacity_limited_but_not_ruled_out - capacity_unknown_due_to_missing_inputs - distributed_training_like_candidate - checkpoint_training_like_candidate - sparse_large_compute_training_like_candidate - candidate_explained_by_serving - candidate_explained_by_storage_operation - candidate_benchmark_like - candidate_hpc_mpi_alternative - negative_screen_coverage_sufficient - negative_screen_blocked_by_missingness - integrity_review_required - inconclusive_due_to_missingness derived_signals: - capacity_segment_window - aggregation_candidate_final_window - negative_screen_required_coverage_state - warning_height_score - suppression_resolution_state time_alignment: segmentation: capacity_segment_window scope_match: same_monitored_scope_or_mapped_child_scope minimum_output_window_coverage_fraction: value_ref: min_final_routing_window_coverage_fraction missing_data_behavior: inconclusive_due_to_missingness short_circuit_behavior: "If capacity_ruled_out_for_scope was emitted for the segment, final route is rule-out and B/C outputs are absent for that segment." aggregation_logic: final_routes: - final_claim: capacity_ruled_out_for_scope when: "A_capacity_gate emitted a conservative high-coverage rule-out." - final_claim: high_training_like_warning when: "At least two independent aligned B identity/support categories are present, warning_height_score.high is true, suppressors are checked, and no unresolved high/critical C conflict exists." - final_claim: medium_training_like_warning when: "One core identity-shape pathway aligns with activity and capacity remains live, suppressors are checked or weakly unresolved, and C has no blocking conflict." - final_claim: weak_training_like_candidate when: "A sparse B candidate exists but support, suppressor, coverage, or C review is incomplete." - final_claim: candidate_explained_or_demoted when: "Suppressor/explanation aggregation explains or demotes the candidate." - final_claim: no_training_like_candidate_detected_in_covered_live_segment when: "Capacity remains live, no B candidate labels exist, negative_screen_coverage_sufficient was emitted, and no decision-blocking missingness exists." - final_claim: inconclusive_due_to_missingness when: "Missing data blocks capacity, identity, suppressor, or discrepancy adjudication." emitted_claim: label: final_algorithm_facing_claim final_claim_labels: - capacity_ruled_out_for_scope - weak_training_like_candidate - medium_training_like_warning - high_training_like_warning - candidate_explained_or_demoted - no_training_like_candidate_detected_in_covered_live_segment - inconclusive_due_to_missingness max_confidence: high statement: "Final route for each monitored scope/window segment, with warning height, caveats, suppressor state, and missingness." missing_data_behavior: "Missingness is surfaced as caveat or final inconclusive route; never as negative evidence." cannot_conclude: - binary_training_truth - no_training_from_missing_B - cleared_C_from_absent_C calibration_defaults: - name: min_final_routing_window_coverage_fraction value: 0.75 unit: fraction evidence_status: calibration_default source_refs: [S07, S08, S09, S18, S23, S26, S28, S29, S31, S35, S40, S41] rationale: "Final routing should cover most of the candidate segment, with lower confidence or inconclusive output below this value." counterevidence_or_failure_modes: - "short high-confidence subwindows can be obscured by broad audit windows" calibration_note: "Report subwindow-level final routes and tune by source coverage."