01 / CONTEXT
Connect.
Bring approved evidence and operational context into a defined decision scope.
DECISION ASSURANCE FOR AI AGENTSGV / 001
Permission must
still be valid
when action
happens.
When one fact changes, Guardian Veritas computes the blast radius across tracked pending AI agent actions, holds the affected set, and leaves unaffected work eligible to proceed.
Reference demonstration · Controlled pilot evaluation
THE DECISION BOUNDARY
Independent path Affected path
Illustrative dependency model, not a live system.
DESIGNED FOR THE SPACE BETWEEN INTENT AND EXECUTIONSCROLL TO EXPLORE
01 / THE APPROACHACTION-TIME AUTHORITY
An earlier approval
is not the
whole story.
An agent can follow yesterday’s instructions perfectly—and still take the wrong action today.
A customer opens a dispute. A source is corrected. An approver changes the scope. The important question is no longer just “Was this allowed?” It is “Does the permission still hold, given what changed?”
Guardian Veritas is an evidence-aware control layer: it identifies dependent actions, rechecks their authority, and preserves a record a person can inspect.
01
Keep authority current.
Bind decisions to the evidence and policy context that support them.
02
Contain with precision.
Hold affected actions instead of treating every change as a global stop.
03
Make the reason visible.
Keep a bounded record of what changed, what was checked, and why.
02 / SEE IT IN CONTEXTINTERACTIVE REFERENCE
A changed signal.
Not a full stop.
Follow a fictional accounts-receivable workflow. Change the evidence, watch the affected reminder pause, then revalidate its permission.
SOURCE EVIDENCE
Invoice A-104
FICTIONAL RECORD- Dispute status
- No open dispute
- Evidence state
- Original snapshot
- Decision scope
- Collection reminder
The reminder’s permission depends on this invoice’s dispute status.
ACTION-TIME CHECK2 independent paths
A / DEPENDENTELIGIBLE
Prepare collection reminder
The current evidence supports preparing a reminder for Invoice A-104.
Depends on Invoice A-104
B / INDEPENDENTCONTINUES
Reconcile Invoice B-208
This separate record has no dependency on A-104’s dispute evidence.
Independent evidence path
01 / READY
The original evidence supports the reminder. Both simulated paths remain eligible.
Browser-local simulation. No actions executed.
Inspect the decision record
Decision
Eligible to prepare a reminder.
Evidence & check
No open dispute. Policy: do not prepare collection reminders while a dispute is open.
Scope & limits
Applies to A-104 only. A synthetic decision receipt—not a guarantee of correctness or a production audit.
This is an illustrative, synthetic workflow—not a live ERP integration. No customer records, collection messages, payments, or production systems are connected. The demonstration does not establish production assurance.
WHAT EXISTS TODAYRECORDED SYNTHETIC EVALUATION
What exists
today.
The simulation above is an explainer. The working engine runs end to end in local, synthetic evaluations.
In the recorded accounts-receivable evaluation, changed evidence held five affected actions while two unaffected actions continued. Partial revalidation and fresh approval produced three mock effects with three independently verified signed receipts.
The latest recorded AR regression passed 1,958 tests with no failures and eight optional skips. These results cover the evaluated software snapshot, not live customer integrations.
Local execution is demonstrated. On-premises and fully air-gapped deployments require environment-specific validation. No real payments, customer messages, or live ERP actions are established by these results.
U.S. provisional patent filed October 2026.
03 / THE ARCHITECTURESIX RESPONSIBILITIES. ONE DECISION.
A control layer.
Not another agent.
Six complementary responsibilities organize the Guardian Veritas architecture—from approved context to a reviewable decision.
ARCHITECTURE DIRECTION02 / AUTHORITY
Gate.
Check whether the proposed action still satisfies the relevant authority and policy conditions.
03 / DEPENDENCY
Trace.
Map how evidence supports a decision so a change can be followed to the affected action.
04 / CONTAINMENT
Immune.
Isolate suspect or stale inputs and hold dependent actions within configured boundaries.
05 / RESOLUTION
Recover.
Support reviewed repair and revalidation. Compensation depends on what the connected system permits.
06 / ACCOUNTABILITY
Proof.
Capture the evidence, policy context, and outcome in a bounded, inspectable decision record.
These are six responsibilities of the engine architecture. Implementation and validation maturity vary by module; production integrations are scoped and validated per customer. “Recover” does not mean universal undo; “Proof” does not mean universal proof of safety or correctness.
04 / THE ENTERPRISE HORIZONSTART NARROW. EARN THE NEXT SCOPE.
More autonomy.
Explicit boundaries.
The ambition is broader than one workflow: accountable AI-assisted work across enterprise systems, with human authority and system limits kept visible.
Financial operations
Evaluate evidence-sensitive workflows such as payment release, receivables review, and reconciliation. Current execution evidence uses synthetic workflows and mocked effects.
Cross-system workflows
Explore how changes in one approved source should affect dependent actions elsewhere.
Human-governed decisions
Keep review, escalation, and explicit authority in scope for consequential enterprise actions.
BUILT BYFOUNDER
MD S.
“Doc” Rana.
Cyber Warfare Operations Officer, U.S. Space Force. Director, Human-AI Teaming Lab, Space Systems Command.
16 years in cybersecurity engineering across Army, Navy, Microsoft, and Walmart. Harvard Kennedy School MPA candidate.
CONTROLLED PILOT EVALUATIONGV / NEXT
Start with one
decision worth
getting right.
Bring a bounded workflow, the evidence it depends on, and the conditions that should stop it. Let’s define what a responsible evaluation would need to show.
Request a pilot conversationOpens your email app · No form or tracking