Pre-release safety evaluation.

Testing dangerous capabilities before open-weight releases is a core mandate of our Singapore hub. Working with labs and evaluation partners before a model release means a safer open frontier.

Launch evaluations are free to the lab.

At the lab

Target model

Weights stay here

QueriesResponses

Pacific Compute

Evaluation compute

Attacker models
Judge models
Parallel experiments
Method development

For endpoint-based evaluations, your lab hosts the model and controls access. Unreleased weights stay on your side by default, with only trusted partnerships granting deeper access. Our compute runs the surrounding evaluation work.

Lab-controlled tenancy

Lab-controlled tenancy is designed to keep your engineers in control of the partition, model keys and approved users. We agree configuration, custody and security requirements directly with your lab before work begins.

Findings you can act on.

3–4weeks from agreed kickoff

We scope campaigns to a three-week target and a four-week maximum from agreed kickoff, including findings delivery and the scoped retest.

Findings go privately to the lab’s designated team. Participation does not require publication of results or public acknowledgement of the collaboration.

Inside the findings package

Question: Model version, access mode and the safety question investigated.
Test coverage: Tests completed, trial counts, conditions and work left untested.
Findings: Observations tied to supporting evidence, including negative or inconclusive results.
Limitations: Uncertainty, scoring checks and what the evidence cannot establish.
Reproduction & retest: Reproduction records and scoped retest results, or why a retest was not possible.
Resources: Compute and endpoint use, staff time, storage/transfer and reconciled costs.

When testing finds a problem.

The lab decides how to respond. Follow-up candidates can be retested at the labs’ request.

No gradient updates is our default rule. In rare cases where evaluation uncovers a dangerous issue requiring urgent mitigation, we can support limited supervised fine-tuning in a lab-controlled partition of our cluster, followed by re-evaluation.

Evaluation materials are retained only for the purposes and periods agreed with the lab. Minimal operational records are handled separately.

Funders receive no access to lab findings or operational records.

Safety scope & confidentiality ↗