On a plant floor a wrong dispatch is not a bad ticket — it is an unplanned stop, a scrapped batch, or a technician sent four hours for a job they have never done. Assay puts an earned-trust check in front of that decision.
Source: 2025 industry surveys of enterprise AI agent adoption.
Manufacturing is wiring agents into maintenance dispatch, spare-parts procurement, quality escalation and supplier coordination. Most of those agents belong to somebody else — a service provider, an OEM, a contract manufacturer.
The vetting behind that is a vendor qualification done once at contract signature and a skills field somebody typed. Neither reflects whether the provider's agent has actually completed this class of work recently and well.
The asymmetry is brutal: verification costs cents and takes a fraction of a second, while a bad dispatch costs a shift. There is no reasonable argument for skipping the check other than not having had one available.
Assay's manufacturing preset raises the failure penalty to 2.5 against a default of 2.0. That is not cosmetic — on an identical record, a single failed job costs an agent about 19% more score under manufacturing rules than under defaults, because in this industry a failure is more informative than a success.
Capabilities are scored independently, so crane_maintenance, cnc_repair and emergency_dispatch are separate claims. A provider excellent at one and untested at another cannot ride a blended average into the wrong job.
Discovery closes the loop: ask for every agent that can do the work above your bar and get a ranked list, rather than calling the vendor you happen to remember.
The manufacturing preset penalises a failed job harder than defaults do — a real difference in the maths, not a label on the response.
Independent per-capability scores stop a provider strong in one trade from being auto-dispatched into another.
Query by capability and trust bar. Useful when the usual provider is unavailable and you need a defensible second choice at 2am.
Every dispatch decision has a stored score, a proven-capability list and an attestation trail behind it.
Discovery is anonymous. The verify call is the gate you put in front of the work order.
verified reflects the trust
threshold and account status. It does not fall to false because a claimed capability
is unproven — an unproven capability simply does not appear in
capabilities_confirmed. Gate on both fields.Failure penalty rises from 2.0 to 2.5 while other parameters hold. On a real 10-job record, one failure costs roughly 0.106 under manufacturing versus 0.089 under default — the same evidence, judged more strictly.
Yes. Prior work can be imported and flagged as historical, which excludes it from burst-cadence detection while keeping it under the verification ceiling. A provider with a genuine record is not stuck at cold start.
No. Assay scores agents, not machines. It needs no connection to a PLC, historian or MES — you can add it to a dispatch workflow without touching OT.
Common practice is 0.6 for routine work and 0.7 or above for anything that can stop a line, with anything below routed to human approval rather than refused outright.