Building the Eval Gate

The Eval Gate is the heart of SOVR—the component that decides whether an AI action should proceed, require approval, or be blocked.

Architecture Overview

Every AI action passes through the Eval Gate before execution:

AI Intent → Eval Gate → [Approved | Pending | Denied] → Action

AI Intent → Eval Gate → [Approved | Pending | Denied] → Action

The Five-Factor Value Function

SOVR evaluates each action using five weighted factors:

1. Risk Factor (0-1)

Measures potential negative impact:

• Data sensitivity level
• Reversibility of action
• Blast radius if something goes wrong

2. Budget Factor (0-1)

Tracks resource consumption:

• API call limits
• Cost thresholds
• Rate limiting compliance

3. Approval Factor (0-1)

Determines human oversight needs:

• Action classification (read/write/delete)
• Historical approval patterns
• Policy-defined approval requirements

4. Compliance Factor (0-1)

Ensures regulatory adherence:

• GDPR data handling rules
• HIPAA PHI protections
• SOC 2 audit requirements

5. Context Factor (0-1)

Evaluates situational appropriateness:

• Time of day restrictions
• User role permissions
• Environmental conditions

Value Score Calculation

valueScore = Σ(factor_i × weight_i) / Σ(weight_i)

valueScore = Σ(factor_i × weight_i) / Σ(weight_i)

Default weights can be customized per tenant:

• Risk: 0.25
• Budget: 0.15
• Approval: 0.20
• Compliance: 0.25
• Context: 0.15

Decision Thresholds

• Score ≥ 70: Auto-approve
• Score 40-69: Require human approval
• Score < 40: Auto-deny

Thresholds are configurable and can be adjusted based on risk tolerance.

Performance

The Eval Gate is designed for speed:

• P50 latency: 8ms
• P99 latency: 45ms
• Throughput: 10,000+ evaluations/second

Conclusion

The Eval Gate transforms AI from a black box into a transparent, auditable system. Every decision is explainable, every action is traceable.