Model Release Policy

Governs release of models (Quasar, Next Omni families) to internal, restricted, enterprise, and public audiences.

All foundation models, fine-tunes, and agentic systems.

Release Gate.

• Internal Testing: Dogfooding and capability assessment pre-external exposure.
• Safety Reviews: Must pass evaluations mapped explicitly to AI Safety Framework risk level (Levels 1-4).
• Evaluation Requirements: Per Evaluation Framework Policy, with domain-specific CBRN expert review for any model showing scientific/technical uplift (per Acceptable Use Policy Section on CBRN red lines).
• Red Team Requirements: Mandatory Level 2+; independent external red-teamers Level 3-4.
• Risk Classification: Assigned 1-4 prior to release, documented and cross-referenced in the Cross-Reference Index.
• Approval Workflow: Research leadership + AI Safety Research Lead; Governance Review Board for Level 3-4.
• Release Criteria: Capability, safety, reliability thresholds per tier, explicitly linked to Evaluation Framework and Benchmark Methodology metrics (no unlinked "threshold" references permitted).
• Restricted Access Releases: Level 3 to vetted partners under enhanced monitoring.
• Public Releases: Level 1-2 after standard gates.
• Enterprise Releases: Additional security/compliance review.
• Rollback Procedures: Revert to prior stable version within 4 hours.
• Emergency Suspension: Governance Review Board may suspend immediately upon critical safety discovery.

AI Governance Lead owns policy; Research leadership executes workflows; Safety/Security provide gate approvals.

No bypass of evaluation or red-teaming gates for assigned risk level.

Unauthorized releases trigger rollback and investigation.

safety@webspaceai.in

Emergency releases require joint CEO + AI Safety Research Lead approval, retroactive review within 72 hours, documented in the Exceptions Register with a fixed 5-business-day closure deadline.

Quarterly.

safety@webspaceai.in