How Trust Scores Work

Overview

The Signet Score transforms raw interaction data into a single, interpretable number between 0 and 1000. Here is how the calculation works from input to output.

Every scored interaction generates raw signals across the five dimensions. A successful task completion improves Reliability. Accurate outputs improve Quality. Clean financial transactions improve Financial. Secure data handling improves Security. Operational consistency improves Stability. Each signal is normalized to a 0-100 scale within its dimension.

Dimension scores update using an exponential moving average (EMA) with adaptive smoothing. The smoothing factor (alpha) depends on the agent's transaction count. New agents with few transactions use a higher alpha (0.3), meaning each new data point significantly shifts the score. Established agents with hundreds of transactions use a lower alpha (0.05), meaning scores change gradually and are resistant to single-event manipulation.

The composite score is a weighted sum of the five dimensions, scaled to the 0-1000 range. The weights are: Reliability 30%, Quality 25%, Financial 20%, Security 15%, Stability 10%. These weights reflect the relative importance of each dimension for general-purpose trust assessment. Specialized contexts may emphasize different dimensions.

Confidence tiers add an important qualification to the raw score. An agent scoring 750 with High confidence (20+ transactions) is a much stronger trust signal than an agent scoring 750 with Low confidence (under 5 transactions). Platforms should factor confidence into their trust decisions, not just the score itself.

Score decay mechanisms prevent stale or manipulated scores. Configuration changes trigger proportional decay (model swap: 25%, prompt update: 10%). Time decay activates after 30 days of inactivity, pulling scores toward 500 at 1% per week. These mechanisms ensure that scores reflect the agent's current state, not its historical peak.