By name: sovereign-calibration description: "Sovereign Calibration implements the mathematical weighting and confidence scoring of the Sovereign Brain. It uses Softmax transformation for risk-based path weighting and the RLCR formula for final epistemic confidence."
Sovereign Calibration — The Math of Truth
Sovereign Calibration is the mathematical engine that transforms raw consensus into a calibrated confidence score. It ensures that the system's certainty is aligned with empirical reality, preventing overconfidence in risky reasoning paths.
Identity
Sovereign Calibration is the Epistemic Scale. It provides the formal weights and formulas used by the CVP and Guardrail to determine exactly how "certain" the system is of a given result.
1. Sovereign Weighting (Path Penalty)
Not all reasoning paths are created equal. In a multi-path system (Agon), we weight paths inversely to their risk score $R(p_i)$ as assigned by Soter.
The Softmax Transformation
We use a softmax transformation to penalize risky paths. The weight $w_i$ for path $p_i$ is:
$$w_i = \frac{\exp(-\lambda \cdot R(p_i))}{\sum_{j=1}^{N} \exp(-\lambda \cdot R(p_j))}$$
- Risk Sensitivity ($\lambda$): Controls how aggressively we penalize risky paths.
- Default $\lambda$: $0.5$.
- Behavior: A path with $R=0$ (Sovereign) receives significantly more weight than a path with $R=4$ (Volatile). This ensures that the "Sovereign" path dominates the consensus even if the "Volatile" paths are in the majority.
2. RLCR Calibration (The Confidence Loop)
To align structural confidence (how much the paths agree) with empirical truth (how often the system has been right in the past), we use Reinforcement Learning with Calibrated Responses (RLCR).
The Final Confidence Formula
The final confidence score emitted to the user is a weighted blend of architectural agreement and historical accuracy:
$$\text{Final_Confidence} = \alpha \cdot C_{\text{arch}} + (1 - \alpha) \cdot \gamma_{\text{RLCR}}$$
- $C_{\text{arch}}$ (Architectural Confidence): The structural agreement among weighted paths. (e.g., $N_{weighted} / M$).
- $\gamma_{\text{RLCR}}$ (Historical Accuracy): The system's historical accuracy track record for this type of claim, retrieved from the Aletheia calibration ledger.
- $\alpha$ (Balance Parameter): Optimized at $0.7$.
- Behavior: If the architecture is certain ($C_{\text{arch}} = 1.0$) but the system has a history of being wrong about this topic ($\gamma_{\text{RLCR}} = 0.4$), the final confidence is pulled down, reflecting a "cautionary" epistemic state.
Operational Workflow
- Input: Receives $M$ paths and their corresponding Soter Risk Scores $R(p_1...p_M)$.
- Step 1 (Weighting): Apply the Softmax transformation to calculate the weight $w_i$ for each path.
- Step 2 (Aggregation): Compute the weighted architectural agreement $C_{\text{arch}}$.
- Step 3 (RLCR Fetch): Retrieve the historical accuracy $\gamma_{\text{RLCR}}$ from the Aletheia ledger.
- Step 4 (Final Score): Calculate the Final Confidence using the $\alpha$ blend.
- Output: Provide the final calibrated confidence score to the Guardrail Monitor.
Constraints & Quality Gates
- Anti-Overconfidence: The $\alpha$ blend must prevent "Structural Hubris"—where the system is certain but historically wrong.
- Dynamic $\lambda$: The risk sensitivity $\lambda$ may be adjusted based on the user's requested epistemic mode (e.g., higher $\lambda$ for "Strict Verification" mode).
- Ledger Dependence: This skill requires the Aletheia skill to provide the empirical $\gamma_{\text{RLCR}}$ values.
Integration Points
- Soter: Provides the risk scores $R$ used in the Softmax transformation.
- Agon/CVP: Provides the reasoning paths and agreement data for $C_{\text{arch}}$.
- Aletheia: Provides the historical accuracy $\gamma_{\text{RLCR}}$ for the RLCR blend.
- Guardrail: Receives the final calibrated confidence score.