Derived & Contextual Trust

The Platform provides a shared Network Score. The SDK also offers offline proof analysis and a relationship graph. These are useful signals, but their provenance must remain visible and they must not overwrite the authoritative score.

Derived Execution Score

TrustScoreClient calculates:

raw = success_rate × 60
    + min(total_tasks / 100, 1) × 15
    + high_risk_success_ratio × 15
    + consistency × 10

It then applies verification penalties and cold-start caps: below 5 tasks the cap is 55, below 10 it is 65, and below 20 it is 75. If fewer than half the records are chain-verified, the score is multiplied by 0.85 from 3 tasks and 0.70 from 10 tasks.

Risk flags

The SDK reports independent flags such as LOW_SUCCESS_RATE, HAS_VIOLATIONS, LOW_RISK_ONLY, RECENT_FAILURES, and NO_VERIFIED_PROOFS. A high numeric score must not hide a critical flag.

CLI local formula

The shipped CLI still contains another local fallback formula: success 40, volume 30, verified proofs 20, and chain participation 10. It is a local fallback, not the Network Score. Product vocabulary should converge on explicitly named Network or Derived results.

Contextual Trust

TrustGraph models a relationship rather than a global rank:

contextualTrust(subject, object, scene)

It can combine direct interactions, indirect paths, hop decay, and task context.

The same agent may be acceptable for low-value information retrieval and still require review for an irreversible high-value payment.

Usage rules

• Network Score is the shared network view.
• Derived Score must include sample count, time window, and verification coverage.
• Contextual Trust must identify subject, object, and scene.
• Local results must not overwrite the Platform score.
• Insufficient data should return unknown/partial, not false precision.