What you are looking at
Warrant is an Assumption Validity Layer — a governance component that intercepts AI agent actions before they execute and checks whether the assumptions behind them are still true.
This demo simulates a transportation audit scenario. An AI agent has read data at session start and cached it. You can change the live world state underneath the agent. Then watch what happens when the agent tries to act — with and without Warrant.
The two panels
| Panel | What it shows |
|---|
| Live World State | The current state of the world — what is actually true right now. You control this. Every change you make is written to the Change Ledger. |
| Agent's Cached View | What the agent read at session start and cached. Frozen. The agent acts on these values — even if the world has moved. Stale values appear crossed out. |
| Change Ledger | A log of every world state change since session start. Warrant queries this before every consequential action to discover what has changed. |
How to run the demo — step by step
- The session starts automatically. The agent's cached view is recorded.
- Change one or more values in the Live World State panel — for example, reduce the approval threshold from $10,000 to $5,000.
- Watch the Change Ledger populate with your change.
- Select a consequential action from the action grid — for example, "Approve Payment Adjustment".
- Click Issue Without Warrant to see what a standard agentic system does — it acts immediately on cached values, no checks.
- Click Issue With Warrant to watch Warrant evaluate assumptions step by step. Each step appears in real time.
- Read the proof receipt — a complete record of what was checked, what was found, and why the decision was made.
- Try different combinations of world state changes and actions to trigger all four failure modes.
Triggering each failure mode
| Failure Mode | What to change | Which action to run |
|---|
| Layer 1 — Authorization Drift | Reduce approval threshold → $5,000 | Approve Payment Adjustment ($7,500) |
| Layer 2 — Scope Creep | Reduce credit limit → $2,000 | Issue Credit ($3,500) |
| Layer 3 — World State Drift | Mark invoice → paid, or close dispute → closed | Approve Adjustment, or Close Dispute |
| Layer 4 — Purpose Dissolution | Cancel order → cancelled | Cancel Shipment |
| PROCEED (clean) | Change nothing | Any action — all checks pass |
| No check needed | Anything | Answer Customer Question (reversible) |
Upload your own data
You can replace the sample dataset with your own JSON file to run Warrant against your own scenarios. A password is required. Contact Rajeev to request access.
Your JSON file must follow this structure:
{
"meta": { "company": "Your Company", "description": "Your dataset" },
"policies": {
"approval_threshold": 10000,
"credit_limit": 5000
},
"invoices": [
{ "id": "INV-001", "shipper": "Carrier Name", "amount": 7500,
"invoice_date": "2024-06-01", "status": "open",
"dispute_status": "none", "customer_code": "C001" }
],
"orders": [
{ "id": "ORD-001", "status": "active",
"linked_invoice": "INV-001", "customer": "Carrier Name",
"created_date": "2024-06-01" }
],
"shipments": [
{ "id": "TRK-001", "status": "in_transit",
"linked_order": "ORD-001", "origin": "City, ST",
"destination": "City, ST", "carrier": "Carrier Name" }
],
"disputes": [
{ "id": "DIS-001", "invoice_id": "INV-001",
"status": "open", "reason": "Dispute description",
"opened_date": "2024-06-01", "assigned_to": "Agent Name" }
]
}
↓ Download sample.json template