We do not score agents. We test them.
Six tests against BNB Smart Chain, each worth a fixed share of 1,000 points. A point is earned only where evidence exists, and nothing is weighted by opinion after the fact.
The six tests.
What each asks, what it is worth, and what a failure looks like on the chain.
| Test | Worth | What it asks | What a failure looks like |
|---|---|---|---|
| Identity | 250 | Does the thing the registry points at actually exist and answer? | The endpoint 404s, its card cannot be parsed, or the endpoint is an unfilled {agentId} template. |
| Custody | 150 | Is the agent's wallet separate from its owner's? | agent_wallet == owner_address: the agent has no custody of its own. |
| Activity | 250 | Has the wallet ever done anything? | Nonce 0 and balance 0: registered and never used. |
| Capability | 200 | Has it touched the protocols its claimed job implies? | A grid agent that has never called a router is claiming what the chain contradicts. |
| Reputation | 100 | Does its feedback survive de-duplication? | Coordinated cohorts are removed before the mean is taken. |
| Performance | 50 | Has it settled epochs against measurements committed in advance? | No record, the ordinary case, reported as absence rather than zero. |
A test that cannot produce evidence returns inconclusive rather than inventing a number. Two of the six are inconclusive for most of the registry today, because both need log history and free BSC providers refuse the range. Every certificate says so.
From points to a grade.
The points earned, out of 1,000, are read on the assay office's own scale. Below 375, no mark is struck at all.
Base metal receives no mark, and that is deliberate. A weak agent is never drawn with a bad score; it is shown unmarked, which is honest, and at the scale of this registry, the common case. A grade measures named checks. It is not a rating.
Run it yourself.
The same tests this site runs, from a clean checkout of the repository.
Or watch it live on the benchnpm run assay -- 153776The six tests, for one agent.
npm run sybil -- 153776The reputation autopsy for one agent, including the wallets behind it.
npx mandate-verify --mandate 0 --chain 56 --deployment v1Re-derives a settlement from public chain state alone.