Private Evaluations
Confidential checkpoint testing.
MineBench runs blind checkpoint evaluations for approved organizations, using the same Arena surface while keeping private identities, builds, votes, and reports scoped to authorized members.
The workspace
From checkpoint to evidence.
- 01
Approved organizations
MineBench approves an organization once. Admins can then invite teammates, and Members can run evaluation work without repeated MineBench approval.
- 02
Private checkpoints
Each evaluation can contain related confidential checkpoints while it is a draft. Once activated, the checkpoint set is fixed so results stay coherent.
- 03
Endpoint or upload
Teams can generate a cohort from a supported private endpoint or provide a complete cohort for validation. Accepted builds remain immutable.
- 04
Blind Arena sampling
Private matchups place one checkpoint against one public model. Identities stay hidden before a vote, and public rankings remain unchanged.
- 05
Organization reports
Authorized members can inspect builds, generation diagnostics, aggregate results, uncertainty, prompt performance, and approved exports.
- 06
Retention window
Closed evaluations remain available for review and export during the agreed retention period, then private data is removed from active systems.
Request access
Private evaluations are available for model providers, research labs, developers, and organizations with unreleased checkpoints or confidential model variants. Access starts with organization approval, then the organization operates its workspace directly.
Confidentiality, export, disclosure, and retention terms may be governed by a separate evaluation agreement or data processing addendum.