01 Rust
Kernel correctness
GPUEmu
A single torch.allclose is not enough to trust your custom kernels.
Independent applied research · St Andrews, Scotland
Make the next AI or systems decision with evidence.
Skelf Research runs scoped investigations into reliability, performance, correctness and technical feasibility. We combine reproducible experiments with inspectable software to establish what works, what fails, and what to do next.
Research that compiles: public methods and tools, confidential customer work.
INVESTIGATION / 01 SCOPED
Decision → evidence Negative findings count
What we investigate / 01
Decisions we help
resolve.
Each engagement starts from a decision that is too consequential to make from a demo, a vendor claim or an internal hunch — and ends with evidence you can defend, including when the answer is no.
01 Agent & RAG evaluation Can we release this agent or retrieval change — and what will break? Scope & price 02 LLM cost–quality benchmark Can we cut model cost or latency without an unacceptable loss of quality? Scope & price 03 Systems validation Are these kernels silently wrong — and is NUMA behind our p99? Scope & price 04 By arrangement Private-data AI feasibility · Untrusted code and client-side AI · Constraint-based decisions and ranking · Warehouse-robotics policy evaluation · Generative-media pipeline feasibility Selective
How an engagement runs / 02
Evidence you
can defend.
The same discipline we apply to our own research: a threshold fixed before measurement, a real baseline, controls, and a reproducibility package — so the person who signs off the decision can see exactly what it rests on.
Read how we investigate- 01
Frame the decision
Write down the decision, its owner, the deadline and the threshold that would settle it — before anything is measured.
- 02
Build the instrument
Build a harness around your baseline and the candidates, with versions pinned and the hard cases fixed in advance.
- 03
Test the edges
Probe correctness, cost and failure modes under controls, and classify every failure rather than reporting an average.
- 04
Deliver the evidence
A recommendation, what was measured, inferred and unknown, and a reproducibility package your team can rerun.
The instruments / 03
Open systems behind
the evidence.
We build and maintain these in the open. They make investigations faster and show how we work; whether one belongs in your stack is a question the evidence answers, not a precondition of working with us.
View all 19 systems 02 Rust
Agent protocol
MPL
Your agents call tools and each other and you need contracts, scoring, and provenance.
03 Go
LLM gateway
Route-Switch
You want provider routing plus prompt optimisation grounded in production traces.
04 Rust
NUMA runtime
NumaPerf
You are hunting p99 tail latency on multi-socket NUMA hardware.
05 Rust
Robotics simulation
WareMax
You need a deterministic RMFS dispatching benchmark for RL or policy work.
06 Python
NL to solver
Savanty
You have a discrete constraint problem (scheduling, assignment) to state in plain English.
Published thinking / 04
The argument
behind the artefact.
Working software is stronger when the question, assumptions, and limits are written down. These notes make that reasoning public.
Read all research notesEssay / 01
Open Science in AI: Why We Publish Everything
The case for openness in AI research: reproducible methods, falsifiable claims, and saying plainly which of our own claims have not yet been measured.
Essay / 02
Formalising Prompts as First-Class Research Objects
Why treating prompts as typed, portable artefacts changes how we reason about LLM behaviour, and how promptel and blogus put that idea into practice.
Essay / 03
waremax: Deterministic Warehouse-Robotics Simulation
How waremax models Kiva-style RMFS warehouses deterministically for reinforcement-learning research, and how it differs from RAWSim-O.
Partner with us / 05
Sponsored research and
funded programmes.
Vendors sponsor benchmarks and reference implementations they want published; consortia bring us in to own a defined work package. Either way, you buy the work, not the conclusion.
01
Sponsored research
Reproducible benchmarks and reference implementations, published with sponsor, method and limitations disclosed.
02
Research work packages
A defined work package inside a consortium or funded programme, with named deliverables and dependencies.
03
Technology ecosystems
Platform providers, standards communities, and technical partnership programmes.
04
Open-source pilots
Maintainers and adopters testing published systems against real operational needs.
Questions / 06
Before you brief us.
- What does Skelf Research do?
- Skelf Research is an independent applied-research and technical-validation lab in St Andrews, Scotland. Organisations commission scoped investigations when a technical decision is too consequential to make from a demo, a vendor claim or an internal hunch: whether an agent or retrieval change is ready to release, whether a cheaper model is good enough, whether optimised GPU kernels are silently wrong. We deliver reproducible evidence and a recommendation.
- How much does an investigation cost?
- Prices are published. A fixed-price technical diagnostic is £2,500–£5,000; an evaluation or benchmark sprint is £8,000–£20,000; an applied R&D project is £25,000–£75,000. Prices exclude VAT, and compute and third-party costs are quoted separately.
- Is our data and work kept confidential?
- Yes. Our research methods and tools are public; customer work is confidential unless publication is expressly agreed. We sign a mutual NDA before any detail is shared, and nothing from an engagement is published or reused without written agreement.
- What if the answer is no?
- Then that is the deliverable. A reproducible finding that an approach does not meet the agreed threshold satisfies the contract, and often saves far more than the investigation costs. Nothing we do is priced on a favourable result.
- Do you only recommend your own tools?
- No. We maintain 19 open research systems, and they make some investigations faster, but we recommend whatever the evidence supports, including other people's software. Every claim we make about our own systems is listed, with its source and limits, in the public evidence register.
- Are Skelf Research projects open source?
- 18 of the 19 current public repositories carry a detected MIT, Apache-2.0 or GPL-3.0 licence. Perishable's source is public but its repository has no licence file yet, so we do not describe it as open source. Each project lives at <name>.skelfresearch.com, with source at github.com/Skelf-Research.