Every agent gets a character of classification.
One string a counterparty can read in four seconds — what this agent is, who answers for it, how recently anyone checked, and whether it still holds class today.
Three questions, kept separate on purpose.
Most tools collapse these into one number. Collapsing them is what lets a confident-looking dashboard be wrong.
A credential the agent still declares, which its own probe was refused, is refuted — not quietly carried forward as though it worked.
Someone's word and a machine's observation are both useful. They are not the same thing, and a record that blurs them is worth less than one that admits the difference.
Derived from the other two and from how old the evidence is. There is no field for it anywhere in the system.
Unclassed is not a failing grade. It means nobody has looked yet — and the register says that out loud rather than leaving a blank for the reader to fill in optimistically.
Nobody can type their own class.
Including us.
Class is derived — computed from evidence already on the record. There is no field for it, no admin override, and no support ticket that changes it. An operator cannot assert it, and neither can the agent.
This is the one design decision the whole registry rests on. A rating anyone can influence is a rating nobody should price against, which is the lesson the credit agencies learned in public in 2008.
Our own relay kit used to grade the evidence it gathered. It no longer does — it reports status codes and provider errors, and nothing else.
The kit runs inside the operator's own trust domain. A grade it awards itself is self-asserted trust, so the registry now rejects it and derives the grade at this end. We removed a feature to make the record mean something.
Published, versioned, and binding on us.
148 clauses, 88 of them in force. Part G binds the Registry itself — including a clause requiring that any published claim about a record be generated from that record, so our marketing cannot drift from our data. Read the Rules →
What went wrong, written like an accident report.
Publicly disclosed incidents involving autonomous actors, analysed without vendor spin. Every report contains a section titled What FleetRegistry would not have done. Across the series so far, incidents we would have prevented: zero. Read the returns →
Put a fleet on the record.
Register a fleet, run the relay, and see what your agents actually hold — before someone else asks you.