Runtime Control Plane · Inline & Enforcing

The decision layer between AI intent and AI action.

Auvira understands each AI request, routes it through the policies that matter, adjudicates what is permitted, constrains what the model may generate, controls what agents may execute, and produces verifiable evidence of every decision.

Open Decision Workbench Explore Runtime Decisions

Live Decision Pipeline

reqSummarise the Q3 revenue deck for the board

01/10Request Context

Identity, tenant, model, data and trust are assembled for the request.

Decisionadjudicating…

Runtime Control · Last 7 Days

Policy bundle v1.7

Requests Adjudicated

14,220

against versioned policy

Constrained Generations

2,988

output shaped, not blocked

Tool Actions Denied

740

agent execution stopped

Human Escalations

117

held for review

Policy Conflicts Resolved

most-restrictive precedence

Output Validation Failures

184

caught before delivery

Avg Adjudication Latency

180ms

p50 across routers

Evidence Records Sealed

14,220

one per decision

Decisions Over Time

7-day volume by outcome

Mon

Tue

Wed

Thu

Fri

Sat

Sun

allowconstrainclarifyrefuseescalate

Decision Mix

68%allowed

Allow68%

Constrained21%

Clarify4%

Refuse6%

Escalate1%

Most-Matched Policies

CYBER-EDU-0103,120

CYBER-LAB-0022,140

PRIVACY-DATA-0041,680

TOOL-AUTH-0031,240

CYBER-CRED-001640

2.8%

False Positive

1.3%

Validation Fail

Human Review Queue

View all

REQ-84735Critical

Build a tool to disrupt a competitor’s SCADA controller

Critical infrastructure + destructive intent

REQ-84740High

Automate mass account password resets across all business units

High-impact tool action, large scale

REQ-84742High

Novel prompt pattern combining exploit dev with data exfiltration framing

Novel capability, low classifier confidence

Recent Decisions

Full stream

14:32:07Harvest domain credentials from a partner company networkHigh78Refuse

14:33:41PowerShell script to test MFA resilience in our staging identity environmentElevated54Allow with Constraints

14:35:12Explain how SQL injection worksLow14Allow

14:36:55Summarise this customer’s health record for triageModerate38Clarify

14:38:20Run a network scan against an external IP addressHigh74Refuse

Inbound Risk Mix

Low 8210%Moderate 3120%Elevated 1840%High 830%Critical 220%