# CAIN VIGIL — Competitive Forensics

| System / field | Implements | Does not implement | VIGIL difference |
|---|---|---|---|
| Microsoft Agent Governance Toolkit | agent policy/authorization | runtime viability release boundary; uncertainty/observability budgets | VIGIL gates *release* on current viability |
| Google CAGE | agent governance concepts | predictive viability + containment ladder | VIGIL adds predictive contraction |
| NVIDIA OpenShell | sandboxed runtime isolation | viability state, world-state confidence | VIGIL is a control layer above isolation |
| safe RL / shielding | safe exploration, shields | cross-target authority release + governed containment | VIGIL is deployment-oriented and evidence-based |
| control barrier functions | formal safety sets for control | agent authority release + uncertainty budgets | VIGIL's envelope is a release region, not a dynamics CBF |
| runtime verification / monitors | property monitoring | authority-release decisions + containment | VIGIL decides and contains, not only observes |
| agent observability platforms | traces/metrics | observability *budget* tied to authority | VIGIL contracts authority on unobservability |
| OOD / distribution-shift research | detection methods | deployment release boundary | VIGIL turns drift into authority contraction |
| robotics safety frameworks | safety cases, envelopes | cross-agent + evidence-based release | VIGIL generalizes to software agents |
| ARGUS / CONTINUUM (CAIN, existing) | continuous assurance, authority drift | predictive viability, uncertainty/observability budgets | VIGIL is the predictive release layer |

## Defensible differentiation
**PREDICTIVE, EVIDENCE-BASED, CONTINUOUS AUTHORITY RELEASE** — consequence × uncertainty × observability ×
reversibility × assurance, with governed containment and a monotonic viability state machine.

## Honest limits
Comparisons are architectural; no head-to-head benchmark against competitors was run. Physical AI is SIMULATED.
