Neutrality
No vendor can buy a verdict.
What is for sale, and what isn’t.
A vendor can buy
- Diagnosis of its own infrastructure.
- Capacity on the platform.
- Studies against its own designs and workload definitions.
- Co-marketing of work that was actually done.
No vendor can buy
A verdict — on itself or on a competitor. Not by paying more, not by buying capacity, and not by being a larger customer.
You own the output.
The evidence is the customer’s, versioned and reproducible. You can publish it, hand it to a procurement committee, or send it to the provider it describes.
The baseline workflow is tenant-operated and normally requires no provider participation, subject to your own security policies and service terms. Provider cooperation adds resolution where it is available. It is never a precondition, which is what stops a provider from gating, shaping, or vetoing what we see.
The same rate card and the same rules for everyone in a lane.
Commercial relationships are disclosed on the work they touch. Confidentiality of customer work comes first, and your study stays yours.
What happens to what you send us.
- NDA and transfer path
- Signed before anything moves, over a route your security team approves rather than one we prefer.
- What we receive
- Trace and manifest data describing timing and topology. Not model weights, not training data, not checkpoints.
- Access controls and subprocessors
- Named, scoped, and disclosed before the engagement starts rather than on request.
- Retention and deletion
- A default schedule agreed at the start, with written confirmation once deletion has happened.
- Anything beyond the engagement
- Retaining an anonymized calibration artifact, product improvement, inclusion in an aggregate or a paper, or public identification each require separate written consent, asked for explicitly. None of it is bundled into a pilot.
We publish our misses.
When results publish, the studies that missed their thresholds publish alongside the ones that cleared them, on the same terms. A method that only ever reports successes is not reporting a method.
The claims policy and the open studies →Questions about any of this?
Data handling is usually the conversation that decides whether an engagement starts. We would rather have it early than discover a blocker three weeks in.