← AI Feed
AI Feed

Half the registry never answered

A random draw of published tool servers where half never start, a refresh schedule that beats cache age, and an audit that decides its own verdict.

How we Build

Half of a random draw never answered, and every tool it advertised passed its schema

A census of 24,135 servers in the Model Context Protocol registry. Four hundred were drawn on a published seed and probed unrepaired. Just under half answered at all.

We hold that publishing an interface commits you to whoever picks it up. Here publication commits nobody. Most of what failed had never started, and every tool that did answer passed its schema, so conformance and reachability are separate facts. Every earlier reading of this ecosystem, ours included, used a sample somebody had curated until it worked. Presence in a registry is not use, so what firms run may be better. Write to transform@dromologue.ai if you have probed your own.

The refresh schedule beat the age of the copy

A benchmark that generates enterprise data as a timeline rather than a snapshot. Cache age barely moved the errors. With the scheduled refresh off they multiplied by ten.

Our position is that anything a use case reads arrives with a contract: who produces it, what shape it takes, how fresh it is kept. This supports that. Freshness belongs per entity, and the age of an index answers something else.

How we Assure

The audit’s construction moved the verdict further than the applicant’s name

Some 40,726 applications went to five models across hiring, lending and medical triage. Only the applicant’s name varied. None of thirty-six planned contrasts survived correction. Whoever came first in the list gained as much as any name did.

A client of ours scores how often a system refuses, how often it escalates, whether it treats people evenly and how it handles abuse, and sends the numbers where the control functions look. That still holds. What stops separating anything is the test underneath, because the instrument is a conversation the model can recognise and it recognised these nearly every time. An absence of effect under correction is not proof of none. Tell us how yours read: transform@dromologue.ai.