Fieldnote
A sensor is not its reading
Device, deployment, calibration, observation, and evidence are five different things
A soil-moisture probe sits in a field and sends a number every fifteen minutes. In most farm software, the probe, its location, and its data stream share one identifier: the device ID on the telemetry feed. That is convenient until the first ordinary event in the device's life — a recalibration, a move, a replacement — and then the single identifier starts telling small lies. This note argues that a sensor and its readings are different referents, and that keeping them apart is what lets evidence survive the life of the instrument.
Five things are easily collapsed. The device is the physical probe, with its serial number and its history. The deployment is the fact that this device was in this place, at this depth, during this period. A calibration is an event that changed how the device's raw signal is interpreted. An observation is a recorded result: a value, a unit, a time, a method, a device, a place. Derived evidence — a seasonal dataset, a model run, a chart in a report — is built from many observations by someone else, for some purpose. Each has a different author, a different lifetime, and a different reason to be trusted.
Why the distinction pays for itself
Consider what each collapse costs. If the device identifier stands for its readings, a probe that is moved to another field carries the first field's history with it, and a dashboard that queries "readings of device 17" silently mixes two places. If the deployment is folded into the device, there is no record of where a reading was taken except by inference from dates. If calibration is not an event of its own, an analyst cannot tell which readings a correction applies to. And if a dataset is treated as the observations themselves, revising the dataset appears to change what the probe measured. Separating them costs a few more records; collapsing them costs the ability to answer questions an auditor, a scientist, or a regulator will eventually ask.
A worked case
Soil Probe 17, 451.EPT1QX0PDW4YF6KG, is installed in North Field in spring 2031 by the cooperative's operations team. The Valley Soil Observatory, which owns the calibration protocol, recalibrates it in June after a drift is noticed. In September the cooperative redeploys it to another field. Over the season it produced thousands of readings, which the observatory publishes as one dataset referring to Soil Probe 17 and to North Field by RID. Anika, observing soils independently, cites that dataset in a fieldnote about the dry spring.
Each step leaves its own trace. The probe's RID never changes: recalibration and redeployment are events in its history, not new identities. The deployment record says which field the probe was in, and when. The calibration event says which readings were adjusted. The dataset is a knowledge artifact with its own identity and revisions, related to the device and the field. Anika's fieldnote is yet another artifact, making claims that cite the dataset. If the observatory later corrects a batch of readings, the dataset is revised; the probe is untouched, and the fieldnote still points at the dataset it cited.
Now suppose the probe fails in 2033 and a new probe is installed in the same hole with the same label painted on it. That is a new device. It needs a new RID, and the old probe's identity is retired rather than reused, so the 2031 readings keep pointing at the device that actually produced them. A label on a stake is a name; it is not an identity.
The integrity question
Readings are often protected by checksums or signatures so that tampering is detectable. That is valuable and orthogonal. Document 08 keeps cryptographic content digests separate from the RID and distinguishes representation integrity from persistent identity (§9.3). A digest proves that these bytes have not changed; the RID says which referent they are about. A signed dataset of Soil Probe 17's readings still needs to say, by RID, which probe and which field it describes.
What remains open
Document 08 does not say how fine-grained observation identity should be: whether every reading earns an RID, whether only datasets do, or whether readings are records inside a space that never need registration. That is a granularity decision for each space, and the related question of whether events such as calibrations are named or kept as relationship records is open (SPEC-05). When a repaired device counts as the same device, and when a replacement must be a new one, also needs continuity rules that governance has not yet written (SPEC-04). What is settled is the boundary this note defends: the device that measures, and the record of what it measured, are never the same identity.