System design
Architecture Explorer
AI Proposes → Policy Decides → System Executes Safely. Click through each layer to see exactly what it may and may not do.
Pipeline
Click any node to inspect its responsibilities and security properties
Groq AI Diagnosis
Sandboxed diagnosis: root cause, bounded action proposal, confidence, microcopy.
- Model: openai/gpt-oss-20b via Groq (openai/gpt-oss-20b recommended in the brief).
- Structured JSON output validated with Zod before it is used anywhere.
- If Groq errors or returns invalid schema → deterministic fallback, no unsafe action.
- Cannot create payments, links, or orders; cannot change amounts; cannot approve itself.
🔒 Output is Zod-validated🔒 Fallback diagnosis on any AI failure🔒 AI has zero execution authority
Interactive state machine
Transitions highlight from real live events
IDLE
PAYMENT_ATTEMPT
FAILURE
NORMALIZED
AI_DIAGNOSIS
POLICY
RECOVERY
SUCCESS
At POLICY the machine branches: APPROVE/OVERRIDE → RECOVERY → SUCCESS, while BLOCK stops safely. Branch points are enforced by lib/policy/engine.ts.
Pipeline events
POLLING- Waiting for events — run a checkout, recovery, or evaluation…