Fidelity

Layer: Agent (behavioral assessment)
Scale: 0–100 with Low/Moderate/High/Critical/Extreme Risk bands
Production Tier: Transaction-Grade (<10ms cached), Monitoring for evidence processing

Purpose

Fidelity measures signal integrity—the credit score for autonomous systems. It scores whether an agent can be trusted based on its observable behavioral track record. Identity is measured by Provenance; Fidelity measures behavior exclusively.

Mathematical Methodology

Fidelity uses a weighted geometric mean of four behavioral dimensions. The geometric mean ensures trust requires consistency across all dimensions—perfect contract fulfillment with high anomaly frequency should not score well.

Formula

FIDELITY = 100 × (B^β × C^γ × R^δ × A^ε)^(1/(β+γ+δ+ε))

Where:

• Default weights: β=0.30, γ=0.30, δ=0.25, ε=0.15
• All dimensions floored at 0.01

Scoring Dimensions

1. Behavioral Consistency (B) - Weight: 30%

Pattern stability across interactions:

B = 1 - (σ_agent/μ_cohort) × diversity_penalty

• Peer-cohort normalization
• Action sequence entropy
• Context-switching patterns

2. Contract Fulfillment (C) - Weight: 30%

Completion rate and quality:

C = Σ(stake_i × success_i × quality_i) / Σ(stake_i)

• Stake-weighted success rate
• Quality scoring per outcome
• Penalty for abandoned contracts

3. Reputation (R) - Weight: 25%

Multi-party trust assessment:

R = (Σ(feedback_i × trust_i × recency_i) / n) × diversity

• Weighted by submitter's Fidelity
• Recency decay function
• Source diversity factor

4. Anomaly Freedom (A) - Weight: 15%

Absence of suspicious patterns:

A = e^(-λ × anomaly_rate)

• Exponential penalty for anomalies
• Severity-weighted incidents
• Time-windowed measurement

Risk Bands

• Low Risk: 80–100
• Moderate Risk: 60–79
• High Risk: 40–59
• Critical Risk: 20–39
• Extreme Risk: 0–19

Aggregation Justification

Weighted geometric mean:

• Non-compensatory across behaviors
• Zero in any dimension collapses score
• Creates incentive for consistent quality

Gaming Resistance

Sybil Reputation Attack

Attack: Create fake counterparties for positive feedback
Countermeasure: Reputation weighted by submitter's Fidelity; low-Fidelity feedback discounted

Coordinated Boosting

Attack: Real agents rate each other highly
Countermeasure: Diversity factor measures independence; cluster analysis detects coordination

Cherry-picking Tasks

Attack: Only accept easy tasks
Countermeasure: Stake-weighting prioritizes high-value tasks; avoiding them reduces impact

Edge Cases

New Agent (No History)

• Return "insufficient evidence"
• Specify interactions needed for confidence
• No zero score—absence of evidence ≠ evidence of absence

No Peer Cohort

• Fall back to absolute variance
• B = 1 - σ(actions)/σ_theoretical_max
• Less precise but defined

Single Counterparty

• Reputation diversity = 0
• R score heavily discounted
• Flag as "limited reputation breadth"

Example Scenarios

Vendor Agent A:

• B = 0.92 (consistent patterns)
• C = 0.78 (good fulfillment)
• R = 0.88 (positive reputation)
• A = 0.95 (few anomalies)
• Fidelity: 88 (Low Risk)

Vendor Agent B:

• B = 0.90
• C = 0.35 (poor fulfillment)
• R = 0.85
• A = 0.92
• Fidelity: 60 (Moderate Risk)

Geometric mean makes fulfillment failure visible.

Target Buyers

• Agent orchestration platforms
• Enterprise procurement teams
• AI marketplace operators
• Payment protocols