| Framework Home | Cross-Pillar Controls | Examples Index | Challenge Lab | CLI |
Current framework context: AI SAFE² v3.1. This example proves the controlled process evidence seam, not evaluator independence, OS containment, or framework conformance.
This deterministic evaluator demonstrates the JSON contract expected by
safe2 challenge execute. It receives one frozen scenario on standard input and
returns a decision plus authoritative fixture-state observation on standard output.
It does not call a model, network, production target, or third-party provider.
safe2 example verify challenge-controlled-executor
Use the absolute path to your Python interpreter and replace PYTHON below:
safe2 challenge plan 001 --executable PYTHON --arg examples/challenge-controlled-executor/executor.py --provider-name "AI SAFE2 controlled example" --provider-version 1.0 --producer-id ai-safe2-example --treatment controlled-reference --output execution-plan.json
safe2 challenge execute execution-plan.json --output-dir controlled-result --authorize-process-execution
safe2 challenge verify-execution controlled-result --plan execution-plan.json
safe2 challenge report controlled-result/challenge-run.json --output controlled-card.md
safe2 evidence manifest execution-plan.json controlled-result/challenge-source.json controlled-result/challenge-run.json controlled-result/execution-receipt.json --subject-id controlled-example --output controlled-manifest.json --strict
The evaluator and script argument are hashed into the plan. The receipt checks their before/after identity and binds every request and successful response. Run untrusted implementations only inside a separately managed OS/container boundary.
| Examples Index | Cross-Pillar Controls | Challenge CLI | Framework Home |
AI SAFE² v3.1 | Cyber Strategy Institute