Confidential by default. Independent by design.

Confidentiality

• NDA before detail. We sign a mutual NDA before you share anything about your systems, data or results. The commissioning brief asks for nothing confidential.
• Your data stays yours. We use it only for the agreed investigation, keep it in the environment agreed in the statement of work (yours, where you require it), and delete or return it at the end of the engagement, confirmed in writing.
• No quiet reuse. Nothing from an engagement — data, findings, prompts, or the fact that you engaged us — is published, used to train or brief any system, or added to our public evidence register without your express written agreement.
• Sub-processors and reviewers. If an external reviewer or a third-party service is needed, we name it and get your agreement before it sees anything.

What you own, and what we keep

You own the deliverables made for you: the decision report, the results, and any harness or test suite configured around your systems. We keep the right to reuse general methods and our own pre-existing tools, which are open source, and any improvements to them that contain none of your information. If you need different terms — exclusive rights to a harness, for example — we agree them in the statement of work, before the work starts.

Negative findings are deliveries

We are paid for the investigation, not for a favourable answer. A reproducible finding that the proposed approach does not meet the agreed threshold satisfies the contract. Acceptance criteria are written before measurement and allow for "no".

Sponsored research

Vendors and programmes sometimes fund a benchmark, compatibility study or reference implementation they want published. They buy the work, not the conclusion. Sponsored outputs are published by default with the sponsor named, the method and limitations stated, and any conflict disclosed; they are never presented as an unrelated independent endorsement, and the sponsor does not get a veto over results. See the partners page for how sponsored work and research work packages are scoped.

Affiliated businesses and conflicts of interest

Skelf Research is affiliated with two businesses that implement and operate AI systems: Dipankar.co and Neul Labs. Skelf Research investigates feasibility, evaluates claims and produces evidence; those businesses build and run systems. The separation is deliberate, and we apply three rules to it:

1. If an investigation could lead to implementation work for an affiliated business, we say so in writing before the engagement starts.
2. A research conclusion is never contingent on winning follow-on work, and the report is delivered in full whether or not any follow-on is offered or accepted.
3. You are free to take the evidence to any implementer. Where we recommend one, we say whether it is affiliated.

The same applies to our own open-source tools: we recommend the approach the evidence supports, including when it is someone else's software.

Licences

Public source is not the same as unrestricted commercial rights. Before an engagement uses any Skelf project, we confirm its licence and maturity for your use case. Most current projects carry MIT, Apache-2.0 or GPL-3.0 licences; Perishable's repository has no licence file yet, and we do not describe it as open source until it does. Each project page shows its licence and status.

Pricing and invoicing

Prices are published on the services page. We invoice in pounds sterling. Prices exclude VAT. Compute, external reviewers, travel and third-party licences or data are quoted separately. A diagnostic is fixed-price; larger work is quoted after it, against a written scope.

Assurance

We hold no security or quality certifications, and we do not claim any. Where an engagement requires specific assurance — a security questionnaire, on-site-only data handling, named-person vetting — tell us in the brief and we will say plainly whether we can meet it. Our machine-readable company profile is at /company-profile.json.