ai-safe2-framework

Stranger Acceptance

A first-time evaluator can check an installed release without cloning the source repository, granting credentials, enabling telemetry, or executing an untrusted skill:

python -m venv .venv
# Activate the environment, then install the released wheel from your trusted source.
safe2 self-check --strict
safe2 acceptance run ./safe2-acceptance --strict
safe2 acceptance verify ./safe2-acceptance

The bundle contains the installation report, one fixed benign static-control fixture, one fixed hostile static-control fixture, their expected and observed decisions, exact hashes, a machine report, and a human card. Verification checks the report seal, fixture bytes, and replayed decisions. Neither fixture is executed, and the workflow uses no network.

This makes the first evaluation easy to reproduce; it does not make the result independent. The CLI authors selected the controls and the CLI evaluates them. Independent assurance begins when a separate evaluator runs the released artifact in an environment and test corpus they control, preserves the resulting bundle, and reports both successes and failures. A passing bundle is not proof of general security, scanner precision, runtime enforcement, or AI SAFE² conformance.

After the controls replay, run safe2 assess against an explicitly authorized project and use task receipts, claim audit, Challenge Lab, and an independent reviewer for the assurance depth appropriate to the decision.