Terrasope

First spike

August 2026. A bench question, not a field campaign.

The base

Terrasope does not start from zero. The record layer, the offline behaviour and the on-device inference all come off Corner, an unattended retail system running on the same edge stack. Same boards, same ledger, same reconciliation path.

Corner is the small version of the problem. A machine in a room nobody is standing in, that has to keep an account of itself and still be right about it a year later. Terrasope is that machine with a drill on it.

Raspberry Pi 5 and a Jetson module with heatsinks and ribbon cable on a bench
The bench. The same pair of boards Corner runs on.

What carries, what is new

Carried from Corner

Already running in production

  • Durable event ledger, one sequence space
  • Offline outbox, reconcile on restore
  • On-device inference in a fixed power envelope
  • ARM hardware, no cloud in the path
  • Fleet health and telemetry

Added for Terrasope

What this spike tested

  • Instrument payload: spectra, not payments
  • Detection events in that same sequence space
  • Byte-identical scoring across architectures
  • Depth and run identity bound at recovery
  • Power budget of a rig, not a countertop

The spike

The question was whether the ledger holds when the payload is an instrument instead of a payment. A mock spectrometer and a deliberately trivial correlation detector went behind it. Detection events and every other event on the box share one sequence space and one reconciliation path.

Verification checks green
31
Runtime paths
container, bare process
Architectures
x86, ARM
Median spectrum to durable, Pi 5
52 ms
Full eight-service stack
~170 MB

The offline claim

Demonstrated, not asserted. The uplink was cut mid-run at the network layer while the instrument was sampling. Sampling continued and the outbox buffered. On restore the ledger converged with zero loss, zero gaps and zero duplicates, audited by event identity.

Determinism

The same seed produced the same detection with a byte-identical score on x86 and on ARM. If the number moves when the hardware moves, it is not a measurement.

No sensor hardware exists. The instrument is a mock and is labelled as one everywhere it appears. Terrasope has no client, no deployment, no field campaign and no revenue.

What is next

Scale the hardware, not the argument. The record layer has now held twice, once in production against payments and once on the bench against an instrument. The next spend is the sensing front end, because that is the only part still unproven.

Craftpine · craftpine.com