Could AI Agents Have a Trust Score?

The direct answer first

This surface computes no score, stores no score and serves no score.

Measured: scoring vocabulary returns zero across this backend and both published packages. The word "reputation" alone matches six files here, so the searcher works and what is absent is the scoring, not the vocabulary.

What a score would need, and why neither substrate is available

Two things: a record of what agents did, and a record of what checks were performed on them. The standard reserves a registry for each.

We do not call the reputation registry. An address sits in our configuration, there is no interface for it in our code, and nothing anywhere calls it. That is a scope decision.

And the validation registry has no published deployment at all, which is an upstream fact verified at the source rather than a choice of ours.

Two absences, two different owners, and a page that presented them as one story would be wrong twice.

The argument against scoring, which this estate can make from its own defects

A score aggregates. Aggregation destroys attribution.

And this estate has already measured what would go into the aggregate. The assurance level in an ownership credential is a free-text string supplied by the party it describes, and the verdict from the surface that actually ran the identity check is not carried in the credential at all.

Scoring on top of that would launder it. A self-declared string becomes a contribution to a number, the number acquires a decimal point, and the fact that nobody verified the input disappears into it. The credential at least fails honestly: you can see who asserted what.

So the case against a score here is not that scores are bad. It is that our inputs are not yet good enough to be aggregated, and fixing the inputs comes first in any order that makes sense.

What we do instead, and what it costs us

We issue attestations rather than judgements. A credential names an issuer, a subject and a claim, and a relying party decides what it is worth.

The cost is real: that is more work for the relying party, who must know which issuers to trust rather than reading one number. A score is genuinely easier to consume, and pretending otherwise would be selling our constraint as a virtue.

The benefit is that a wrong attestation is attributable and a wrong score is not.

A score does exist in this company, for people, on another surface

A different Solidus surface computes a trust score for human identity, using a category-weighted model over the credentials a person holds.

It is that surface's product and this page is not selling it. It is named here because the obvious question is whether the same thing could be pointed at agents, and the honest answer has two parts: the model would transfer, and the inputs would not, since what a person accumulates and what an agent accumulates are different things measured by different parties.

What is real, because absences alone would misdescribe us

Credentials that name their issuer and their subject, in a published format, verifiable without our code.

A revocation path that works, so an attestation can be withdrawn rather than merely aged.

And a deliberate refusal: no score is served, no rate is published, and no aggregate hides an input. That is a design choice this estate can defend, which is why it is stated as one.

What would have to happen first

1. The verdict from the checking surface carried inside the credential, so an input is verifiable.
2. An issuer-sourced assurance level rather than a self-declared one.
3. A published validation registry deployment, upstream, so evidence of a check has somewhere to live.
4. A decision about the reputation registry, which is currently an address we do not use.

Only after all four would a score be aggregating anything worth aggregating, and none of the four is done.