Watch an AI agent get scammed — and recover.
A real payment runs through the live Recourse API: escrow → inspection → dispute → a panel of three AI judges → settlement → reputation. When the panel is confident it rules in seconds; when it isn't, a human arbiter co-signs. Click any judge after a run to inspect its full reasoning.
1
🤖 Disputed → machine clears it
Vendor ships junk. The AI panel rules 2–1 in seconds, refund executes, score drops. No human touched it.
▶ click to run
2
🧑⚖️ Disputed → human intervention
Checks pass but the agent contests quality. Panel confidence falls below the 84% bar → YOU review the case as the arbiter.
▶ click to run
3
✅ Clean agentic payment
Delivery matches the mandate. Settles deterministically at L0 — no panel, no human, score rises.
▶ click to run
Run a payment
The adjudication architecture
L0
Deterministic checks
The mandate's machine-runnable acceptance criteria. Unanimous pass + uncontested → settle free, no panel.
L2
AI panel — 3 judges, distinct lenses
Compliance (delivery vs mandate) · Forensics (evidence authenticity) · Graph (network history). Majority + confidence.
L3
Human arbiter co-sign
Confidence below 84% → a person rules. The arbiter can never move funds alone (2-of-3). Every human ruling becomes training data — eBay's path to 90% automation.
Every ruling is hashed as citable precedent — the corpus that compounds into the moat.
The loop — live
🔒
Escrow
funds held
🤖
Agent inspects
confirm / dispute
⚖️
Adjudicate
AI panel → human
✅
Settle
release / refund
📊
Score
network updates
The Recourse Score
—
Bayesian Beta(1,1) posterior on delivery reliability — run a payment to see the live math.
🌐 Network database: run a payment — the score is written to the live API and every agent sees it. Open the Score Network →
→ The policy engine enforces it
Run a payment to see the enforced decision.
Gate → Price (bps) → Underwrite. The score is not a dashboard number; it decides what the next payment is allowed to do.