Agree the terms before the results exist

This page is a scoping discussion aid. It is not a legal agreement, not a template contract and not legal advice. It sets out what needs agreeing and where I start from; the binding terms come from an agreement reviewed by appropriate legal advisers.

The checklist

1. Background IP. Each party lists what it brings: the sponsor’s models, data and code; my existing tools, corpora and harnesses. Background stays with its owner. Open-source material I bring keeps its existing licence.
2. Foreground IP. Agree who owns or may use what the study creates — the protocol, harness changes, results and report. Separate sponsor-specific material from general methods that either side may reuse.
3. Data access and deletion. State what is shared, in what form, where it is processed and who can see it. My starting position: sponsor data is used only for the study, and at the end it is returned or deleted with written confirmation, except anything both sides agreed to retain for reproducibility.
4. Publication and review window. Decide before any results exist whether the work is published. If it is, the sponsor gets a fixed review window, with its length agreed at scoping, to flag confidential information and factual errors. Confidential material is removed or abstracted; factual errors are corrected. Findings are not vetoed or reworded.
5. Unfavourable results. Agree now that a negative or null result is reported in the same form as a positive one. A publication plan that applies only to favourable outcomes turns a study into marketing. I hold my own work to the same standard: How Reproducible Are Evaluation Conclusions? is a self-audit reporting where an evaluation’s ranking was less reliable than a single table suggested.
6. Authorship and acknowledgement. Authorship follows intellectual contribution; funding and access are acknowledged. Agree the funding statement wording in advance.
7. Result-independent fee. The fee is fixed at scoping and paid against deliverables, never against the direction of the result.
8. Conflicts. See below.

Conflicts and independence

I founded Neul Labs, which builds AI agent infrastructure in Rust; my research affiliation is Skelf Research. I may have built, advised or hold interests in systems that a sponsor wants evaluated, or in their competitors. I disclose these at intake, before confidential material is shared. Depending on the conflict, the outcome may be to proceed with the disclosure stated in the report, narrow the scope to avoid the conflicted component, arrange independent review of the analysis, or decline.

Running the work under a different domain or brand does not make the reviewer independent. Independence is a property of the person and their interests, and the report says so plainly.

Illustrative example — not a client engagement. A sponsor wants two agent frameworks compared, one of which builds on infrastructure from a company I founded. Disclosed at intake, the options are to decline, to restrict the study to a component where I have no interest, or to have a reviewer with no connection to either framework re-run the analysis from the published artefacts. The sponsor chooses the third, and the report names the conflict and the reviewer.

Before any of this, the study itself should be specified; start with the study brief.