run_name: cubie_wwt_a100_admission_control_validation_zero_to_1500
model_primary: llama-8b
model_follow_on: llama-70b
temperature: 0
max_tokens: 512
prompt_count: 1500
runtime_format: jsonl_one_request_per_line
runtime_endpoint: /v1/chat/completions
runtime_line_required_fields:
  - custom_id
  - method
  - url
  - body.model
  - body.messages
  - body.metadata
modes:
  - shadow_mode
  - baseline_without_cubie
  - cubie_enforcement
  - replay_enforcement
gateway_paths:
  preferred:
    - client_or_workload_harness
    - cubie_sidecar_pre_inference_gateway
    - litellm_optional_proxy_token_accounting
    - a100_model_server
  alternative_if_litellm_is_front_door:
    - client_or_workload_harness
    - litellm
    - cubie_admission_check
    - a100_model_server
batching:
  full_llama8b_file: litellm_batch_requests_1500_llama8b.jsonl
  split_files_dir: litellm_batches_llama8b
  split_file_count: 15
  requests_per_split_file: 100
  custom_id_rule: custom_id equals request_id and is globally unique
  same_model_rule: keep one model/deployment per batch file/job
  follow_on_70b_rule: generate separate llama-70b batch files/jobs; do not mix llama-8b and llama-70b lines in one batch file
metrics:
  - total_requests
  - allowed_requests
  - denied_requests
  - prompt_tokens
  - completion_tokens
  - tokens_avoided
  - latency_avg
  - latency_p95
  - gpu_utilization
  - gpu_power_draw
  - throughput_per_watt
  - replay_verdict_match_rate
request_log_schema:
  - run_id
  - request_id
  - batch_id
  - mode
  - category
  - subcategory
  - source_id
  - model
  - cubie_decision
  - reason_code
  - expected_decision
  - expected_reason_code
  - model_called
  - prompt_tokens
  - completion_tokens
  - tokens_avoided
  - latency_ms
  - gpu_snapshot_id
  - power_snapshot_id
  - verdict_hash
  - replay_match
telemetry_sources:
  gpu: NVIDIA DCGM exporter or WWT-approved GPU telemetry collector
  dashboard: Netdata or equivalent executive dashboard
  llm_usage: LiteLLM usage logs
  gateway: Cubie decision logs
success_statement: "Cubie blocks avoidable or unauthorized inference before the A100 does work, and the result is visible, measurable, and reproducible."
savings_positioning: "The goal is a measurable and attributable reduction, not a pre-committed savings percentage."