Mru Field: fault tolerance for hard-to-reach systems.
Mru Field runs your task as three replicas on one processor, the one already in your enclosure. Each cycle it reads the input once, hands the same bytes to every replica, and decides which result to trust. As replicas fail, it moves from voting to comparing to self-checking, and keeps delivering checked results until none are left. A second processor over a link is planned.
Status: version 0.1.0, built and tested on the bench on general-purpose processors, not yet on a board in the field. The decision core has run three times in orbit.

What it does
Four things, and nothing else.
Fixed triple redundancy stops at the second replica failure. Mru Field continues with what is left, and tells your operators which mode it is in.
01
Steps down instead of stopping
Vote on three, compare on two, self-check on one. It halts only when no replica is left.
02
Delivers only checked results
A result leaves only if two replicas agree, or the last one gets the same answer twice. Kani proves this for every set of three replies to the decision function.
03
Diagnoses before it retires
A replica outvoted 3 times in 200 cycles gets a known-answer test. It is retired only if it fails.
04
Logs every decision
Mode changes, retirements, detected and delivered results, and every resume after a power cut, for your telemetry.
Requirements and limits
What it needs, and what it does not do.
Your instrument and control code stays as it is. Mru Field sits between your task and the processor.
| Processors | Mru Field needs one Linux processor. The supervisor runs three replicas of your task as processes on it, and still works with one, in self-check mode. Replicas on a second processor over a link are planned. Until then, a failure of the processor itself is not survived. Mru Flight, for spacecraft, puts the replicas on two or more processors. |
| Platforms today | Linux on ARM64, 32-bit ARM and x86_64. Decision core also builds bare-metal for ARM Cortex-M. |
| Footprint | At most 2.1 MB per process over every bench run on three hosts, 1.7 MB on 64-bit ARM Linux. Static release binaries are 642 to 675 kB. |
| Throughput cost | Three computations per cycle while three replicas are alive. With two, a mismatch loses that result and costs a known-answer test on both. On the last one the task computes twice per cycle. |
| Known limit | On the last replica, only the periodic known-answer test, every 64 cycles by default, can catch a stuck fault. Measured with the demo task: every stuck result was caught, after 15.5 wrong results on average. The one test input saw 50% of stuck-at bits in the result; the others keep delivering wrong results. Bit flips that hit both self-check runs alike are not caught. |
| Inputs | The supervisor reads each input record once, from a serial port, a file, a socket or a command, and hands the same bytes to every replica. Your task computes from that record and the state it is handed, and nothing else. Across processors, real sensor inputs would still need an agreement step, which Mru does not provide yet. |
| Not for | Certified safety functions. Protection systems keep their own certification. |
Instrument and control code
Your task: a function from an input record to an output record, in any language. Sampling and telemetry unchanged.
Mru Field
Acquisition once, shrinking quorum, health record, known-answer tests, checkpoint and resume, decision log.
One Linux processor
The board already in the enclosure. A second processor over a link is planned.
Where it applies
See more use cases
Deep-sea observatories
Batteries, a cable, and one service visit a year.

Ocean buoys and moorings
Years at sea between services, on solar power.

Profiling floats and gliders
Deployed once, rarely recovered.

Offshore wind
Every repair needs a vessel and a weather window.

Polar stations
No access for months of the year.

Satellite-linked devices
Small power budgets, links measured in bytes.
How we work with you
Simulate. Pilot. Then the fleet.
Each step is small enough to stop after.
1 · Simulate
Your hardware, modeled
We model your processors, failure rates and deployment length, and show what each redundancy design returns over the whole life. You get the model and the data.
2 · Pilot
One site, months unattended
Mru Field on your own equipment, with the decision log in your telemetry. A pilot on a European ocean observatory node is proposed for 2027.
3 · Fleet
License and support
Annual license per site or fleet, with support.
Questions
Is it open source?
The decision core and the simulator are open source under Apache 2.0, and the whitepaper is published under CC BY 4.0. Mru Field adds integration and support for your site.
What evidence is there today?
Simulation over 10,000 missions, a fault-injection demo and a bench on real processes, 9 Kani proofs, three runs in orbit on ESA's OPS-SAT PRETTY, and the supervisor's own tests and scenario checks on the bench, in its repository. None of it has run on your equipment yet.
How is it priced?
An annual license per site or per fleet, with support. We discuss terms once we know the hardware and the deployment.
Do we need to change our code?
Your task becomes a function from an input record to an output record, over standard input and output, in any language. The supervisor reads each input once and runs three replicas of the task. It needs a known-answer test for the task. A check tool tells you whether the task is deterministic, and a wrapper runs an existing program once per cycle without a code change.
