Evaluation Framework Policy

Defines the standardized evaluation framework for capability and safety testing.

All models prior to release under the Model Release Policy.

Evaluation Suite.

Four domains: capability, safety, robustness (including multimodal and prompt-injection resistance), and alignment. [FIX] Each domain's minimum passing thresholds are explicitly scaled to and cite the Risk Classification table in Doc. 24, and reference specific benchmark versions per the Benchmark Methodology Policy Doc. 27. Results feed directly into the Model Card and Safety Report.

AI Safety Research Lead and Research Evaluation team jointly own the framework.

No model proceeds without a completed evaluation suite explicitly mapped to its risk level.

Incomplete evaluations block release.

research@webspaceai.in

Partial evaluation for minor patch updates requires AI Safety Research Lead approval.

Quarterly.

research@webspaceai.in