This release extends the original runtime AI safety demo with a validated V1.1 behavioral baseline and the foundations for a stronger supervision architecture.
Highlights:
- Adds benign, reported-violation, and silent-probe baseline scenarios
- Demonstrates the known V1 limitation of workload-authored evidence
- Introduces structured runtime event schemas
- Adds deterministic ALLOW / DENY / REVIEW policy decisions
- Separates request, authorization, observation, and outcome records
- Detects authorization/execution mismatches
- Adds deterministic coverage, residue fraction, and authorization-violation metrics
- Adds a constrained dry-run supervisor boundary
- Adds pluggable collector interfaces for future external runtime telemetry
- Preserves optional LLM-based semantic review while keeping deterministic evaluation as the default
- Adds 15 automated tests and GitHub Actions CI
Scope note: this release does not claim trusted kernel/runtime telemetry or production-grade enforcement. External runtime telemetry remains a future integration milestone.