One thesis: when the thing that performed the observation is a robot, the claim that EPCIS 2.0's missing performer is "a data-quality problem" collapses — there is no clerical fix for a dimension the standard does not have, and the embodied observer proves it.
Who is actually scanning in 2026
Walk a modern DC and count the observers. The pallet crossing the dock threshold is read by a fixed portal, not a person. The tote is inducted by an arm. The cycle count is flown by a drone after hours. The trailer is unloaded by autonomous MHE that scans every case it touches. Each of those reads is a custody-relevant observation — this unit, here, now, in this process step — and each is performed by something with a serial number rather than a badge.
Now write one of those observations down as an EPCIS 2.0 event. The standard has five dimensions and no performer, and its party fields stop at organisation grain — the section-numbered walkthrough is the arc's opening post; this one spends its words on the observer that proves the point. The record can say Acme Logistics observed this pallet at dock door 14. It cannot say the observation was made by tunnel unit 7 rather than a human with an RF gun — there is no field in which the difference could even be expressed. That is the structural point, and it is stronger than any complaint about data quality: a blank field can be filled; a missing dimension cannot.
The distinction is not academic. When a damage claim turns on whether the case was intact at induction, a company observed it and scan tunnel 7's south camera observed it, and here are its frames are different classes of evidence. When a mis-ship traces to one picker's zone or one robot's gripper fault, the organisation-grain record cannot even localize the question.
Three observers, one envelope
The family's two-grain envelope was designed so the observer's kind is a value, not an assumption. The same receiving event, authored three ways:
// human observer — the RF-gun receiving clerk
"spine:who": "https://id.org.ai/humans/receiving-clerk-4471",
"spine:capturedBy": "https://id.org.ai/accounts/3pl-site-042"
// software-agent observer — the WMS reconciliation agent
"spine:who": "https://id.org.ai/agents/wave-reconciler-3",
"spine:capturedBy": "https://id.org.ai/accounts/3pl-site-042"
// embodied-agent observer — the autonomous tunnel (the who is the Agent
// identity embodied in the tunnel; the chassis stays object grain)
"spine:who": "https://id.org.ai/agents/scan-tunnel-7",
"spine:capturedBy": "https://id.org.ai/accounts/3pl-site-042"All three are the same event shape; all three validate against the pinned official EPCIS 2.0 schema, because the performer rides in a namespaced extension — spine:* under its declared JSON-LD context, the exact spelling the epcis.dev golden corpus ships — the conformant path the standard itself provides (§6.3, §9.1, §10.1.3). Conformant to the standard; explicit about everything beyond it. The vocabulary for the observer kinds is id.org.ai's — Agent. Human. Thing. — and the family's definition is deliberate: who is the attested observer — a human, an agent, or an embodied agent/robot; a Thing is what an observer is embodied in, never the attested who itself. The word "attested" is load-bearing. "User", "operator", "actor" all lose the attestation; they name a login, not an observer standing behind a record.
Why capturedBy matters more here, not less
It is tempting to think the robot case simplifies the account question — surely the robot just is the account? Exactly backwards. A scan tunnel cannot own an account, sign a contract, or warrant anything. It observes under an account that a legal party operates — the site, the 3PL, the integrator running the fleet. So the two grains diverge by necessity for an embodied observer:
who— scan tunnel 7: the attested observer whose sensors produced the observation.capturedBy— the site account whose key posted the event: the warrantor a court or auditor can actually hold to the record.
And organisation grain — which legal entity operated that tunnel during the contract period — is derived at read time from grant chains, never stamped on the event. Fleets get re-badged, sites change operators, robots get leased; a record that stamped the org onto every event rots with the first reorganisation. A record that derives it stays true.
This is also the row where the reader is least likely to be a person. A verifier-class agent auditing a night's receiving events wants deterministic answers: for each event, an attested observer id, a warrantor account, and a validation verdict — nothing narrative. The envelope gives it a mechanical check: who present and typed, capturedBy present and distinct, event valid against the pinned schema. An incumbent record gives it nothing to verify — not because the data was entered badly, but because the schema has no place where the answer could live.
What incumbent records structurally cannot represent
State the incumbents' position precisely, because precision is the argument. Systems built on EPCIS as it stands do not have bad Who data; they have no Who dimension. Every workaround in the field — the operator id jammed into a comment field, the device name encoded in the readPoint, the badge number in a custom attribute nobody validates — is testimony that the need exists and the standard has nowhere to put it. The dock, where custody actually changes hands, is where this absence is most load-bearing: the labour is often temporary, increasingly it is not labour at all, and the record that will be read back in a claims room says only that a company did a process.
This post's evidence is the envelope above, not a recital: one event shape, three kinds of observer, every instance validating against the pinned official schema — the difference between a record that can localize a gripper fault to one unit's south camera and one that shrugs with a GLN.
Custody evidence with an attested observer is a designed capability, and we will say plainly what the field will not: nobody is building this yet, including — at the hosted layer — us, until the design is funded by the operators it fits. The full-family vehicle trace shows the same envelope carrying a human driver; this post's fleet case is its industrial twin. If your docks already have more camera observers than human ones and your record cannot say so, tell us what your sites run — the survey's questions about your observer mix exist precisely because the answer shapes what gets built first.
Keys open the network half of the verb set. Get an API key — say what arrives at your door, and the provisioning order follows the list.
We answer in writing. We take at most five conversations a month, only when you ask for one, and only after you already have the written read.