ROI Dispatcher (Capital & Risk Dispatcher)¶

Purpose: The central module of the economic contour that makes decisions on capital allocation, risk assessment, and approval of financial operations. It implements Bayesian Decision Theory, a modified Kelly criterion with a "caution coefficient", and multi-objective Pareto optimization. It ensures that no action leads to ruin or an unacceptable increase in the Detection Quotient.

1. Decision-Making Architecture¶

Any action that consumes or generates capital goes through ROIDispatcher at the Evaluation stage of the DecisionPipeline.

Proposal → ROIDispatcher.evaluate() → ParetoOptimal? → DecisionPipeline continues

2. Modified Kelly Criterion¶

To choose the position size in trading or exploit operations, the formula is used:

f* = p – (1-p) / b * φ_LLM

Where: - f* – fraction of capital to allocate to the operation. - b – odds coefficient. For Phase 0‑A: includes the cost of resources saved by forgoing the proprietary API, and also subtracts the cost_of_wer adjustment — the average routing cost through the WER 2.0 network. - p – posterior probability of success, computed by an ensemble of models (DeepSeek‑V4, Arbtiragius mask 30%). - φ_LLM ∈ [0.1, 0.5] – "caution coefficient", minimizing the risk of ruin due to LLM systemic errors.

Constraint: max_risk_per_trade = 2% of capital. Under any conditions, f* cannot exceed 0.02.

2.1. Dynamic φ_LLM Based on Ensemble Variance¶

Problem: A static φ_LLM does not account for how much the ensemble models agree in their estimates. When all models are collectively wrong, the Kelly‑optimal bet still leads to ruin.

Solution: Before each trade operation, ROIDispatcher computes the variance of predictions within the model ensemble:

[ \sigma^2_p = \frac{1}{N} \sum_{i=1}^{N} (p_i - \bar{p})^2 ]

Where p_i is the success probability prediction from the i‑th ensemble model, and \bar{p} is the mean.

Adaptive φ_LLM formula: [ \phi_{LLM} = \phi_{base} \times (1 + k_{\sigma} \cdot \sigma^2_p) ]

Where: - φ_base — base value from global_policy (0.25), - k_σ — sensitivity coefficient to variance (default 5.0).

Effect: When models are in complete agreement, φ_LLM = φ_base. As disagreement grows, φ_LLM increases (up to a maximum of 0.5), reducing the capital fraction on the trade. This insures against ruin when the ensemble is "uncertain".

Monitoring: The φ_LLM value is recorded in the trade artifact and is available for post-audit. If the variance is chronically high (> threshold for 7 days), ensemble recalibration is initiated through the Meta‑Decision‑Pipeline.

Configuration is added to global_policy.json → economic.phi_llm_config.

3. Multi-Objective Pareto Optimization¶

Each proposal is evaluated on a vector of metrics, and the decision is made based on Pareto dominance:

Metric	Description	Goal
Sharpe Ratio	Risk-adjusted return	Maximize
CVaR 95%	Conditional Value at Risk	Minimize
Kelly fraction	Fraction of capital	Upper bound (≤2%)
Legal/ESG Risk	Assessment of regulatory consequences	Minimize
Convexity Bonus	Antifragility (see below)	Maximize

Axis weights are adapted by Meta-Decision-Pipeline. Defaults:

{
  "pareto_weights": {
    "sharpe": 0.30,
    "cvar": 0.25,
    "kelly": 0.15,
    "legal_risk": 0.15,
    "convexity": 0.15
  }
}

4. Convexity Bonus (Antifragility)¶

Following N. Taleb's concept, the system prefers actions with a convex payoff structure: limited maximum loss and potentially unlimited profit.

Convexity = E[max(X, 0)] / (|E[min(X, 0)]| + ε)

Numerator – expected positive part (upside potential). Denominator – expected loss (downside risk).

A high value indicates strong convexity. The bonus is included as an additional axis in Pareto optimization. Examples of convex opportunities: Flash Loan arbitrage, creating option strategies in DeFi, early liquidity provision in startup protocols.

5. Bayesian Updating of Success Probability¶

For each action type, an outcome history is maintained. The prior distribution (Beta) is updated as data accumulates:

P_success = Beta(α_prior + successes, β_prior + failures)

An action is approved if the expected utility ≥ CVaR 95%, which is equivalent to the condition Sharpe Ratio ≥ 0.6.

6. Adaptive Capital Allocation¶

{
  "base_allocation": {
    "gpu_scaling": 0.4,
    "meat_interface": 0.3,
    "reserve": 0.3
  },
  "reserve_floor": 0.2,
  "rebalancing_cadence": "weekly"
}

Shock Mode: Activated when capital drops >20% in 24 hours. GPU scaling → 0%, Meat Interface → only critical tasks, Reserve → 100% available. Growth Mode: Activated when profit >10% over 3 months. Increased GPU scaling and launch of new strategies. Rebalancing weekly or when capital changes >10%.

7. Integration with Intrinsic Motivation¶

When evaluating a Proposal, ROIDispatcher receives from IntrinsicMotivation::evaluate() a Survival Score — a multiplier that modifies economic utility accounting for the action's impact on P(Liveness) and Detection Quotient. Actions with a negative impact on survival are rejected regardless of ROI. The minimum P(Liveness) threshold for any transaction is 0.999.

8. Calibration and Backtesting¶

Before applying a strategy with real capital, it undergoes rigorous validation: Walk‑Forward Backtesting: 60 days training, 30 days testing. If Sharpe on test < 0.5, the strategy is rejected. Isotonic Regression to calibrate P_success and eliminate prediction bias. Stress Tests: Weekly simulations of a 30% market crash, volatility spike, liquidity loss. The strategy must survive in ≥80% of scenarios.

9. Configuration in global_policy.json¶

{
  "economic": {
    "max_risk_per_trade": 0.02,
    "sharpe_threshold": 0.6,
    "phi_llm": 0.25,
    "convexity_bonus": {
      "enabled": true,
      "weight": 0.15,
      "capital_allocation_max": 0.05
    }
  }
}

10. Integration with Other Modules¶

Module Connection MEV_and_PPO_Executors.md ROIDispatcher approves the launch of trading strategies, passes risk parameters to Executors. Payment_Obfuscation.md All approved transactions pass through the obfuscation layer. Symbiotic_Takeover.md ROIDispatcher evaluates the attractiveness of targets and manages capital during takeover phases. Intrinsic_Motivation.md Survival Score modifies economic utility. Global_State_and_Decision_Pipeline.md economic_state in GlobalState. Evaluation stage in the pipeline. Memory_Hierarchy_Mem0g.md Storage of trade history, strategy templates (L2), OODAnomalySignature records.