Score Decay

What is Score Decay?

When an AI agent's configuration changes, the historical trust data that informed its current score may no longer be fully applicable. An agent that scored 800 with one model may not perform at the same level with a different model. Score decay addresses this uncertainty by reducing the score proportionally to the significance of the change, then allowing the agent to rebuild its score under the new configuration through continued operation.

The rate of decay varies by component type, reflecting the relative impact of each change. Model swaps cause the most significant decay because the foundation model is the most fundamental determinant of agent behavior. Prompt updates cause moderate decay. Tool changes and memory modifications cause less decay. These decay rates are calibrated through empirical analysis of how different types of configuration changes actually affect agent performance.

Score decay is not a punishment -- it is a calibration mechanism. It ensures that trust scores always reflect the agent's current configuration rather than a historical one. Agents that perform well after a configuration change will quickly rebuild their scores, often surpassing their pre-change levels if the change was an improvement. The decay simply provides a period of appropriate caution while the new configuration proves itself.

Example

How Signet addresses this

Signet implements a precise score decay schedule: model swaps trigger 25% decay, prompt updates trigger 10% decay, tool changes trigger 8% decay, and memory changes trigger 5% decay. These rates are applied to the current score at the time of the change. Signet tracks configuration change history and displays the decay and recovery pattern in the agent's Trust Report, giving counterparties full visibility into how the agent's score has evolved through configuration changes.