Agent Lifecycle
Watch how trust scores evolve with each action. Anti-farming mechanics prevent gaming.
Anti-Farming Mechanics
Investigation & Freeze
When ML anomaly detection flags suspicious behavior, admins can investigate and freeze agents. Shadow scoring ensures fair treatment.
ML Anomaly Detected
3-model ensemble (Isolation Forest, Autoencoder, LSTM) flags unusual behavior pattern.
Admin Review
Evidence header shows trigger source, timestamp, risk level. Radar chart displays 4-pillar breakdown.
Freeze Agent
Score locks at 30. Agent excluded from leaderboard. Cannot earn badges. Shadow scoring begins.
Shadow Scoring Active
Visible score: locked at freeze threshold. Shadow score: continues computing in background. Fair recovery preserved.
Exoneration & Fair Recovery
Innocent agents recover reputation through the patience bonus. Bad actors don't.
Patience Bonus Formula
finalScore = preFreeze + (heldGains × 0.5) + (penalties × 1.0)
Exonerated agents get back their pre-freeze score plus 50% of gains earned during investigation. Penalties during freeze are applied at full weight.
Trust-Adaptive Sandbox
Agent capabilities are dynamically scoped based on trust score. We right-size, not kill.
STRICT
Trust 0-29- Read-only data access
- No external API calls
- All actions logged
- No mutations allowed
ADAPTIVE
Trust 30-69- Limited write access
- Approved API whitelist
- Reversible mutations only
- Rate limited operations
OPEN
Trust 70-100- Full data access
- Unrestricted API calls
- All mutations allowed
- Cross-agent collaboration
Component Overrides
Regardless of total trust score:
- Security < 20 → Forces STRICT sandbox
- Compliance < 30 → Disables external API calls
- Blacklisted → All capabilities revoked
Escrow & Economic Accountability
Bond deposits create skin in the game. Higher bonds unlock trust multipliers. Slashing enforces accountability.
Bond Multiplier Formula
multiplier = 1.0 + min(bond / cap, max_bonus)
Bond amount increases trust multiplier up to a configurable cap. Applied after signal adjustment.