Research notebook — Issue 001 · Updated 05 Aug 2026
Volunos Labs
An applied research lab. We take technical questions that are not yet answerable and turn them into evidence.
Put a question to the lab — write to research@volunos.uk
Answered inside 2 working days, by the person who would run the studyBelfastApplied R&DSIC 62012 / 72190
Current lines of enquiry
-
How far can a simulation be coarsened before its decisions stop matching the fine model?
Simulation & surrogate models Opened 06.2026
-
What does a defensible experiment record for an R&D tax claim actually need to contain?
Evidence & research records Opened 06.2026
-
Can edge-side signal filtering remove enough noise to make a cheap sensor behave like an expensive one?
Signal processing on constrained hardware Opened 07.2026
-
Where is the line between a problem that is merely hard and a problem no amount of engineering will rescue?
Algorithmic feasibility Opened 06.2026
Entries
-
How much of a physical process can be replaced by a learned surrogate before the error compounds?
Engineering teams increasingly want a fast approximation of a slow solver — a surrogate that answers in milliseconds what a reference model answers in hours. The interesting question is not whether a surrogate can be fitted; it usually can. It is where the surrogate's error stops being a rounding difference and starts changing the decision the model is being used to make.
Our harness sweeps coarsening levels against a reference solver and reports the point at which decision agreement breaks down, rather than reporting mean error alone.
- Method
- Reduced-order modelling; error-bound sweeps against a reference solver; decision-agreement scoring rather than mean error.
- Measures
- The coarsening level at which the surrogate stops agreeing with the reference model on the decision, and the compute saved at each level.
-
Can filtering at the sensor make cheap hardware behave like expensive hardware?
A great deal of instrument cost buys noise performance rather than capability. Where the noise is structured — thermal drift, mains interference, quantisation, mechanical resonance — some of that gap is recoverable in software, on the device, within a very small compute and power budget.
The honest answer is bounded: filtering cannot recover information the sensor never captured. Our bench establishes where that boundary sits for a given class of measurement, so a hardware decision rests on evidence rather than on a vendor's claim.
- Method
- Characterisation of noise structure; fixed-point filter design; benchmarking against a reference instrument under matched conditions.
- Measures
- How much of the gap between a cheap sensor and a reference instrument is recoverable in software on a constrained microcontroller, and where it stops being recoverable at all.
-
Where does a problem stop being hard and start being intractable?
A recurring pattern in commercial software is a requirement that is quietly impossible at the scale it is written for — a scheduling constraint, an exact-match problem, a combinatorial search that looks tame in a demo and explodes in production. Discovering this after eighteen months of engineering is expensive.
This line of enquiry produces a short, readable feasibility argument early: what the cost curve looks like, which relaxations recover tractability, and what the organisation would have to give up to get them.
- Method
- Complexity framing; empirical scaling curves on representative instances; documented relaxations and their cost in accuracy or coverage.
- Measures
- How the cost of the exact problem grows with the parameters the business actually cares about, and what each tractable relaxation costs in accuracy or coverage.
-
What shape should engineering data take before anyone knows what will be asked of it?
Test rigs, production lines and field equipment generate data long before there is a question to ask of it. The modelling decisions made at capture time — units, provenance, sample identity, what counts as one observation — determine whether that data can answer anything at all five years later.
We are interested in the minimum discipline that keeps an engineering dataset answerable without imposing a schema nobody will maintain.
- Method
- Provenance modelling; unit and uncertainty carrying; retrospective answerability tests against archived datasets.
- Measures
- Which capture-time decisions determine whether an archive can still answer a question years later, tested against public engineering datasets.
-
How much inference can move to the device before latency stops being the binding constraint?
Moving computation to the edge is usually justified on latency. In practice the binding constraint often turns out to be something else — thermal budget, memory bandwidth, model update logistics, or the cost of getting evidence back off the device when something goes wrong.
The enquiry measures which constraint actually binds for a given deployment, so the architecture argument can be settled with numbers.
- Method
- Latency and thermal profiling on target silicon; quantisation sweeps; measurement of update and observability overheads.
- Measures
- Which constraint binds first on the target silicon — latency, sustained thermal budget, memory bandwidth, or the cost of updating and observing a model in the field.
Working with organisations
We are usually brought in before the requirement is written. If you already know what is to be built and need it built, an ordinary software studio is the right supplier. Volunos exists for the stage before that: when the question is whether a thing is possible, at what cost, and with what error.
Feasibility study
You have a technical question with no reliable answer inside the organisation. We frame it precisely, establish what would count as evidence either way, run the smallest experiment that settles it, and write up what we found — including the case where the answer is no.
Typically 3–6 weeks · Written findingsBench build
Where a study is not enough and the answer only appears in something that runs. We build the narrowest artefact that exercises the uncertain part — an instrument for the experiment rather than a product — and report where it holds, where it fails, and what production would demand of it.
4–10 weeks · Working artefactR&D engineering
Laboratory code that has proved itself and now has to survive contact with a production system. We make it maintainable, testable and observable, and we keep the technical documentation and experiment records that an R&D tax relief claim has to be able to stand on.
Ongoing · Code + evidence recordNotes
A result that rules an approach out is still a result, and we price and report it exactly as we would a positive one. It is also why we decline studies that exist to justify a decision already taken.
Our experiment-record tooling captures hypothesis, method, parameters, result and date at the moment the work happens, rather than reconstructing any of it at year end. It runs across our own ledger, and the same discipline carries into the record we keep on a client engagement.
A question is a better thing to send than a specification. One paragraph will do it: what you are trying to establish, which approaches have already been attempted, and what decision the answer would unblock. Put a question to the lab — write to research@volunos.uk and an answer comes back inside two working days.
Colophon — the mark
Two 45° strokes descend to a bottom-centre node. The right-hand stroke carries two units past the vertex — a needle that has overshot its datum and not yet settled. The horizontal tick across the opening is the reference level it is being measured against. Drawn only on grid lines and diagonals; no curves.
Put a question to the lab
Enquiries
A single address, opened by whoever would run the study
Everything arrives at research@volunos.uk, where the person who would run the study is the person who opens it. Two working days is the normal turnaround, and it applies equally to the reply that says a question falls outside what we can usefully take on.
More on how an engagement is scoped and what we would need from you is on the working with us page.