AI agents are entering production faster than auditors can certify them. Today we're opening the AgentAuditLabs (AAL) protocol — a signal-based, append-only trust scoring layer for AI agents that anyone can read, verify, and contribute to.
WHY ANOTHER SCORE?Existing “trust scores” are opinions in disguise
Most agent ratings today are bucketed into categories — “safety,” “capability,” “reliability” — and combined with weights chosen by whoever runs the leaderboard. The buckets and weights are policy decisions dressed up as math. Change the weights, change the leaderboard.
AAL is built differently. There are no categories. There are no buckets. There is only the primitive of a signal: a signed, timestamped event with open-vocabulary tags emitted by some source we know how much to trust. Anyone can emit signals. The engine sums them weighted by source trust and source diversity. That's it.
FORMULAThe whole thing fits in one line
score = round(weighted_mean × coverage)
Where weighted_mean is on the 0–100 scale and coverage is on 0–1. No magic multipliers. No fudge factors. The score is reproducible from the signal event log; if you don't trust the score, you can replay it yourself.
Every signal is an Ed25519 receipt
Each signal is signed at the moment of emission by the source. The source's public key lives in our open registry. If a source is compromised or its weight changes, every historical signal still verifies — the chain of custody is permanent and auditable.
WHAT'S NEXTRead the methodology, run the protocol
Browse the scoring methodology, check the registered agents, or read the REST API reference to start integrating AAL trust scores into your own systems.
— The AAL Team