Identity
Machine control is disabled by default and route-authorized; interactive access is a hidden break-glass path.
Security model
Authority is a property of the system, not the prompt.
The research platform is designed to fail closed when identity, scope, approval, evidence or cost controls are missing.
Scope
Every target, redirect and tool action is validated against active authorization.
Approval
Sensitive actions require one-time approval bound to exact parameters.
Evidence
Artifacts are stored privately, hashed and linked to the source action.
Models
Providers propose structured actions but cannot grant themselves authority.
Cost
Paid calls reserve budget atomically and stop at the configured hard ceiling.
Memory
Model output remains a candidate until an accountable, evidence-bound promotion decision exists.
Audit
Control decisions create sanitized, append-only application events.
Research boundary
Viridis does not test systems without authorization.
We do not perform denial of service, credential theft, phishing, persistence, malware, lateral movement, mass exploitation or out-of-scope testing. Sensitive actions stop for human review.