ClawlasDocs
Open the app
Cost & economics

Cost-control engine

Per-scope budgets, a 80/90/100% degradation policy separated by operational importance, and the metrics that decide when a local model replaces a paid one.

Budgets are enforced and separated by operational importance. Critical position monitoring and reconciliation must never stop because the research budget was spent.

Controls

Every provider, model, agent, strategy and trade candidate has a budget and limits. They fall into three groups — spend caps, request-shape limits, and quality/fallback guards:

Spend caps

  • daily / monthly LLM budget
  • budget per model / agent / strategy / candidate
  • provider daily / monthly budget
  • max RPC compute units
  • max social-data spend
  • cloud escalation budget
  • hard stop enabled

Request shape

  • max agent steps
  • max tool calls
  • max input / output tokens
  • max retries
  • min seconds between analyses
  • max router escalations

Quality & fallback

  • fallback model
  • local-model max queue depth
  • local-model max latency
  • min quality score
  • min schema success rate
  • shadow mode enabled

Degradation policy

Budgets are enforced and separated by operational importance — critical position monitoring and reconciliation must never stop because the research budget was spent.

  1. 80%
    Warn; route non-critical tasks to local models; reduce context/output limits
  2. 90%
    Disable optional research & premium sources; escalate only high-value candidates
  3. 100% soft
    Only execution safety, reconciliation & critical alerts
  4. 100% hard
    Block all non-essential paid calls

When a local queue saturates or exceeds its latency objective, the router uses a budget-approved cloud fallback or defers the non-critical task — never letting report generation delay reconciliation or risk controls.

Cost metrics

Profitability is measured per candidate, per trade and per strategy:

  • technical_cost_per_candidate
  • technical_cost_per_executed_trade
  • technical_cost_per_profitable_trade
  • llm_cost_per_strategy
  • rpc_cost_per_chain
  • net_profit_after_all_costs
  • profit_to_total_cost_ratio
  • cost_per_valid_ai_result
  • local_vs_cloud_cost_per_valid_result
  • router_escalation_rate
  • router_fallback_rate
  • model_schema_success_rate
When a local model replaces a paid one
Only when all hold: validated local quality ≥ threshold; local cost per valid result < cloud cost per valid result; local latency ≤ deadline; local availability ≥ requirement. “Open source” is never itself an acceptance criterion.

Cost categories

CategoryIncludes
Fixed / semi-fixedShared VPS & domain allocation, premium data/RPC plans, paid swap APIs, backups, future GPU
Usage-basedLLM tokens & web search, X calls, RPC compute units, market-data units, object-storage ops, CI minutes, GPU seconds
Trade-levelMaker/taker fees, spread, slippage, price impact, gas, priority fees, approvals, failed-tx gas, withdrawal/bridge fees, MEV

Net profit definition

Net profit
  • +gross realized trading profit
  • -exchange fees · spread · slippage · price impact
  • -gas & priority fees · failed-transaction costs
  • -withdrawal & bridge fees
  • -LLM costs · external API costs
  • -infrastructure allocation
= net profit after every cost