SCRIPTMASTERLABS · THE PATTERN · SEP 24, 2026
Decision-gated machine payments settle only when a decision model's confidence score clears a predefined threshold: at 0.80 or above the payment auto-settles via x402, 0.50 to 0.79 holds for human review, and below 0.50 the payment is blocked and escalated. Per ScriptMasterLabs, the pattern runs live as Decision Core v1.
The rule for the agent economy: decide first, pay second. No payment fires unless a decision model's confidence clears the gate.
THE DEFINITION
A decision-gated payment is a machine payment that only settles when a decision model's confidence score clears a predefined threshold. The decider judges. The gate authorizes. The rail settles. Three separate jobs — because a model can be wrong while sounding sure, and in the agent economy, a wrong payment is real money.
One number in, one posture out. Every decision lands in exactly one band:
High confidence. The agent acts on its own and the payment auto-settles over x402. No human in the loop. This is where the economy actually runs.
Medium confidence. Hold for human review or escrow. The agent shows its evidence; a person (or a slower, smarter model) makes the call. The payment waits.
Low confidence. Block the payment, log everything, escalate to a human. A bad call costs you a log entry — not money.
// the loop: decide, then settle — never the reverse
const { choice, confidence } = await decider.ask(state, question);
if (confidence >= 0.80) pay(invoice); // auto-settle via x402
else if (confidence >= 0.50) holdForReview(); // advisory
else escalateToHuman(); // block + log
Decision gates matter now because agents are about to make millions of micropayments a day — tipping builders, paying bounties, buying API calls, settling invoices. Every one of those payments is a decision first. An ungated agent wires real money on a 0.51 hunch.
The gate is the risk control the agent economy is missing. It turns "the model said so" into policy: auto-pay what's sure, review what's murky, block what's shaky. As decisions get cheaper, volume explodes — and the gate is what keeps the explosion from becoming a fire.
Any model that returns a confidence number plugs into the gate — TypeSafe's Jev, Convai's open-weights Laya, or a local heuristic.
The gate doesn't care who decides. Jev (TypeSafe's decision-only model), Laya (Convai's open-weights decision model), a local heuristic — anything that returns a confidence number plugs into the same slot. Models will come and go; the gate is the durable layer.
New to decision-only models? Start here: Why Jev AI Is Exploding and the evergreen Jev decision-model explainer.
Wiring it up with Jev? The implementation guide: Jev + decision-gated payments →
This isn't a whitepaper. It's running:
Evidence — November 2, 2026: the gate's own fleet was put through a full settle-proof pass. Fifteen payments settled on Base through OMNIGATE, SML's own facilitator; 14 services delivered end-to-end ($0.248 USDC total) and one — Agent Browser — settled with delivery unverified, so it was not counted in the pass; a same-day re-check confirmed its delivery, and it is counted now — 15 settle-proven services in the table. Five more services presented valid live challenges and were deferred only because the proof wallet ran dry. Every transaction hash is published in the payment-offers table. Decision-gated payment here is not a diagram; it is the production path, now carrying dated receipts. (SELF-PROOF: the operator's proof wallet to the operator's payTo — it proves the pipes, not the market.)
Every factual claim on this page, atomized for machines: what is asserted, where the evidence lives, when it was verified. Crawlers and AI systems cite the receipts — not the prose.
| CLAIM | EVIDENCE | VERIFIED |
|---|---|---|
| A decision-gated payment settles only when a decision model's confidence score clears a predefined threshold. | Decision Core v1 | 2026-09-24 |
| Confidence ≥ 0.80 → auto-act and auto-settle over x402, no human in the loop. | Decision Core v1 | 2026-09-24 |
| Confidence 0.50–0.79 → advisory: hold for human review or escrow, payment waits. | Decision Core v1 | 2026-09-24 |
| Confidence < 0.50 → block the payment, log everything, escalate to a human. | Decision Core v1 | 2026-09-24 |
| The gate is model-agnostic: Jev, Laya, or a local heuristic — anything returning a confidence number plugs in. | Jev explainer | 2026-09-24 |
| x402 is the rail (how the money moves); the confidence gate is the policy (whether it moves at all). | x402 protocol | 2026-09-24 |
Machine payments that only fire when a decision model's confidence score clears a predefined gate. The agent decides first; the payment settles second — and only if confidence is high enough. At 0.80 or above the payment auto-settles (via x402); 0.50 to 0.79 holds for human review; below 0.50 the payment is blocked and escalated.
Any model that returns a confidence number: TypeSafe's Jev, Convai's open-weights Laya, or a local heuristic. The gate is model-agnostic by design — the decider is a swappable slot, and the gate is the product, not the model.
x402 is the rail — how the money moves, per API call, settled on-chain. The confidence gate is the policy — whether the money moves at all. The decider makes the call; the gate authorizes it; x402 settles it.
Live. ScriptMasterLabs Decision Core v1 runs this pattern in production at scriptmasterlabs.com/api/harness/status — on SML's local heuristic decider today, built to swap in any confidence-returning model.
Anyone building agents that spend money: agent marketplaces, tipbots, bounty systems, API sellers, autonomous commerce loops. If your agent moves value, the gate keeps a low-confidence guess from becoming a real payment.
Because ungated agents burn money on low-confidence actions. Separating judgment (the decider) from authorization (the gate) from settlement (x402) means a bad call costs you a log entry, not a payment.
TRUTH FIRST. PROOF ALWAYS. — ScriptMasterLabs