Work with us
Research & Benchmark Partnership
Put one important question to the test. Agree the experiment together and leave with results your team can use and another reviewer can check, even when the idea doesn’t work.
Your starting point
Would this help?
This is a fit when
- You have an AI or data question that affects a decision.
- Someone on your team can act on the result, with a reviewer who knows the subject.
- Finding that the idea doesn’t work would still help.
- We can agree the data, comparison, permissions and compute budget up front.
Full research fit criteria
- A material local-AI, model-systems, evidence, or deployment uncertainty
- An industry owner and method/domain reviewer willing to freeze the protocol
- A decision whose value remains real if the hypothesis fails
- A bounded study with inspectable data, rights, baseline, and compute
When another approach is a better fit
- A sponsor requiring a positive outcome or suppression of failed runs
- An undefined request to build a frontier or general-purpose model
- No qualified baseline, isolated holdout, lawful rights, or decision owner
- Indefinite tuning after claim-bearing holdout exposure
The useful output
What you take away.
- 01An agreed question, comparison and test plan.
- 02Results, including failures and changes along the way.
- 03Reviewed findings to help your team decide what to do next.
Full project checklist
- 01Experiment charter, frozen protocol, baseline, and all-outcome action tree
- 02Rights, ethics, disclosure, environment, and data/model identity records
- 03Complete run, failure, deviation, and holdout-access ledger
- 04Analysis pack and independent or partner review
- 05Industry decision record and an approved public pack only when separately authorised
Working together / 6–12 weeks
How we’ll work together.
Choose a question that would change a decision. Agree the experiment, comparison, funding and reviewers before running it, then keep the findings whatever the result.
Fictional example: Could a local model sort research papers as accurately as our current method on a held-out set, within our computing budget? Before testing, we’d agree what accuracy and change in time, memory or cost would make the result useful on the chosen computer.
The project steps in detail
- 01Bring the right people together and check the question, rights, decision and computing resources.
- 02Before testing, fix the test plan, comparison, held-out examples, success thresholds, analysis and what each possible result would mean for the decision.
- 03Run only the agreed, limited preliminary tests, then carry out the planned confirmatory test.
- 04Keep a record of the runs, failures, departures from the plan, resources used and reviewers’ objections.
- 05Deliver the analysis and an explicit decision to continue, redesign or stop.
What we need before work starts
Before starting
- One material question with an owner who will act on every possible result
- A qualified baseline, representative data, isolated holdout, and lawful rights
- Method/domain authority and an independent review path
- A bounded compute, time, rerun, and disclosure budget
- Agreement that failures and negative outcomes remain in the record
Named industry and method owners, lawful rights, a frozen baseline and holdout, negative-result protection, compute funding, IP/publication terms, and independent review must be agreed.
Scope and handover
What we agree before starting.
We agree the work in a signed statement of work before starting. This page explains the offer; the signed agreement sets the commitment.
We agree the question, scope and finish point before work starts.
How you accept the handover
- The pre-agreed protocol, baseline, rights, environment, and outcome rules remain intact.
- Required runs, failures, deviations, and exclusions are preserved in the ledger.
- Independent or partner review records material objections.
- The final pack supports an explicit continue, redesign, or stop decision—even when the result is null or negative.
What this work covers
Available only when the question, boundary, rights, partner roles, and funding qualify.
One question, one controlled protocol, one decision. No open-ended R&D programme, guaranteed positive result, or automatic product-performance claim.
A controlled experiment is not a positive performance promise, product pilot, or proof of generalisation beyond its protocol.
Offer source and review record
Questions
Before you enquire.
Can a negative result be successful?
Yes. A valid null or negative result that resolves the decision is a successful research delivery.
Why is this offer qualified?
The question, protocol, rights, partners, compute, and decision value must qualify before a scope is offered.
Does every partnership test ERAIS?
Only when the approved protocol makes ERAIS relevant. NF-C can address broader local-AI and evidence questions.
Talk to us
What’s getting in the way?
Tell us what keeps going wrong and what you need to do once it’s sorted. We’ll come back with what we think could help.