Built to be audited.
Partners at the frontier of AI work from evidence, so do we. Each commitment below names the document we hand over.
Commitments you can check.
Trust in AI infrastructure is usually asked for and rarely shown. A customer is told that capacity is reliable, that data is safe and that models are screened, and is expected to take it on faith. We take the opposite position, and we hold ourselves to it in writing. Every commitment on this page names the document we hand over to support it: a capacity statement that lists what is contracted and what is not, a failover drill report with the method used, a security overview that says what exists and what does not, a custody statement for weights and data, a screening protocol for biosecurity, and a safety case before any robot moves outside the lab. Just as important is the second list, the things we will not claim. We will not claim capacity we do not have, a certification before the audit is done, an endorsement we were not given, a benchmark without a method, a deployment without a safety case or a release without screening. For a partner this is practical. It tells you which statements you can rely on, which are still plans, and what to ask for in due diligence. If something on this site is not backed by one of these documents, treat it as intent and ask us.
Capacity and sourcing
Committed cloud capacity and delivery dates, backed by executed agreements. Providers are named only where the agreement allows.
Reliability
We report goodput and mean time to recover for each training run across capacity pools, with the same definitions every quarter.
Security
Isolated tenants and private networking on every cloud. We inherit the controls of each cloud provider and add our own. SOC 2 Type II and ISO 27001 are audit targets for Sinewgrid itself.
Weights and data custody
Customer weights and proprietary assay data stay in isolated tenants with customer-held keys.
Biosecurity
Sequence-design models ship with output screening, tiered access and logging. Outside experts review capability evaluations before release.
Physical safety
Robot policies are tested on held-out tasks with defined stop conditions before any real-world deployment.
The limits we hold ourselves to.
No capacity we do not have
We state committed capacity only against an executed agreement, and we say which purchasing mode it is.
No certification before the audit
SOC 2 Type II and ISO 27001 are described as targets until an auditor's letter exists.
No endorsement we were not given
Cloud providers, customers and partners are named only with permission. Public documents are linked as references, not as relationships.
No benchmark without a method
Any figure comes with the workload, the configuration and the date, so a reader can repeat it.
No safety case, no deployment
A policy does not move real hardware outside the lab before its safety case and stop conditions are reviewed.
No release without screening
Sequence-design outputs are screened and access is tiered before a model is made available.
Planning a training run?
Tell us the model, the data and the schedule. We reply with a capacity plan and the evidence behind it.