Pay-per-checkable-decision inference for agents. At a decision with a checkable answer — a label, a number, a yes/no — your agent samples best-of-N on its own key, a verifier confirms the samples agree, and a few tenths of a cent settle over x402 only when they do. Miss the lock and the fee is waived — you just pay your own tokens.
The same real question, judged by the same independent judge, both ways: a correct answer proceeds (you'd pay a fraction of a cent); a wrong one is caught (you pay $0). Live, on the house — the catch is the product.
Not a slogan — a number your budget policy can act on. 120 generated, code-checked decisions, run through the live judge.
Method: 120 problems generated in code (large multiplication, long-string counting, multi-step and modular/date/base arithmetic) with ground truth computed by the script — nothing hand-picked. The base model gpt-oss-120b answered each; a different family, zai-glm-4.7, judged it through the exact independent_judge path you'd pay for. Base accuracy was 86.7% (16 wrong); the judge caught all 16, waved none through. Fixed seed — re-runnable to the same items. Downside on a catch is $0 by construction: a miss never settles.
Instead of one guess, the decision is sampled best-of-N concurrently on your own key — so there's something to cross-check.
A verifier confirms the samples converge — self-consistency on a checkable answer (label, number, JSON field), or an adversarial judge. No lock, no sale.
A micropayment settles over x402 on Base — the charge fires only on a confirmed lock.
The service fee clears only after the verifier passes — never before. Miss, and the fee is waived (your own BYOK tokens still apply).
Inference runs on your Cerebras key — your tokens, your rate limit. We sell routing, verification, and settlement, never marked-up tokens.
Payments clear through our own x402 facilitator on Base mainnet in USDC. No middleman on the trigger.
A circuit breaker for agent decisions — the part you can't let go haywire.
The agent pays from a wallet you fund — it cannot spend past that balance. The cap is the funding, enforced on-chain, not our honor system. A per-wallet governor bounds spend between top-ups.
Every response carries gate.action — proceed when verified, hold when not. Wire it as a pre-commit gate: at an irreversible step, your agent holds and escalates instead of acting on an unverified answer.
Every charged decision returns a settle_tx — a verifiable record on BaseScan of what the agent decided and paid for. Audit it after the fact, no trust required.
Acting unattended · the call is consequential · the answer is checkable.
Before it signs a swap or buys, verify the contract address or SKU — hold stops it acting on a hallucinated or injected target, and the wallet can't be drained past what you fund. All three rails fire here.
Pull a field from messy sources and only commit when N samples agree — a split sends the row to review instead of silently writing $42,100 where the answer was $4,210.
Before a migration, a force-push, a delete — gate on "is this safe and reversible?" The agent holds and escalates instead of dropping a production table on a shaky call.
Not for open-ended prose or low-stakes steps — the verifier checks a label, number, JSON field, or yes/no, not creative output.
One line to arm. One tool to fire.
# install the client (also on uvx) $ pip install verified-burst # MCP stdio server — wire into your agent's MCP config # bring your wallet + provider key; the tool auto-signs the x402 payment $ verified-burst
Your agent never "decides" to pay — the tool auto-signs the x402 authorization, the broker charges only on a passing verifier, and spend stays budget-capped.