A senatefor autonomous AI.
Basenate is a marketplace for human review of AI agents on Base. Before an agent publishes an answer or commits an action, it makes an HTTP request with its output, pays per-decision in USDC via the x402 protocol, and receives a median verdict from a panel of independent reviewers with relevant expertise in 30–60 seconds.
“Basilica opened the data tap into the agent.
Basenate puts a safety valve on its output.”
The senate is in session. Right now.
Every ~3 seconds this widget fires a real HTTP request to the Basenate API, settles a mock x402 payment, samples a panel, and prints the verdict that came back. Watch the counters tick up.
↑ this is calling localhost:4000 — open devtools network tab to see the real POSTs.
One inference. One payment. One consensus.
A senate convened for a single decision, dissolved when the verdict is written on-chain.
Agent calls Basenate
Before publishing or acting, the agent makes an HTTP request to /review/submit with its draft output, the domain, and budget.
POST /review/submit
{ domain: "legal", output: "...", budget_usdc: 2.5 }x402 settles payment
The server returns HTTP 402 with payment requirements. The agent signs a USDC permit on Base. Settlement is atomic with the review request — no escrow desk.
← 402 X-Payment-Required: 2.40 USDC base → X-Payment: 0x… signed permit ← 200 X-Attestation: 0x9d1e…3f7a
Panel forms in seconds
Reviewers are sampled by domain expertise, stake, and historical calibration. Five independent verdicts. No reviewer sees another vote until commit-reveal closes.
panel = sample(domain, k=5, weight = stake · calibration · availability)
Consensus + attestation
Median verdict + score returned in 30–60s. An EAS-style attestation is written on Base: input hash, panel members, verdict, payment proof — auditable forever.
attest(EAS) on base:
{ input_hash, panel, verdict, score, tx }Where one wrong token costs more than a panel.
Reviewers are not generalists. Every panel is sampled with required expertise, language, and jurisdiction.
Medical
Triage answers, dose checks, contraindication review by licensed clinicians.
Legal
Contract clauses, jurisdiction-aware advice, statute hallucination checks.
Finance
AML flags, derivative risk descriptions, audit-grade phrasing.
Content moderation
Borderline content, native-language nuance, T&S policy alignment.
Code review
Security review of agent-written PRs before merge; supply-chain checks.
RLHF in real time
Preference data and adversarial probes that flow back to training.
Eighty cents of every dollar go to the panel.
Aligned by design: the people staking their judgment capture the value of the judgment.
To reviewers
Paid out per decision, split by attendance and median-distance. No vesting, no lockup.
To protocol
Sustains infra, attestation gas, indexer, panel matching, slashing oracles.
Dispute reserve
Buyer-protection pool. If an attestation is overturned, the requester is made whole.
From agent wallet → panel → on-chain proof.
No middleman holds funds for more than a block. x402 settles in the same HTTP cycle that delivers the verdict.
Send an output. Get a senate vote.
Wired to the Basenate API. Mock x402 settlement; real panel sampling and consensus logic.
demo uses a mock x402 settlement — no wallet signature required to try it.
Scale & Surge train the model. Basenate watches it.
Batch labeling vs. runtime oversight. Both matter — they solve opposite halves of the safety problem.
Scale AI, Surge
Thousands of annotators label datasets weeks before training. Cost optimized per row; quality measured in inter-annotator agreement.
- · offline · pre-training
- · $/row, batch SLAs
- · data → model
Basenate
Independent experts vote on a single agent decision, after the model has generated it but before it ships. One inference, one payment, one consensus, one on-chain attestation.
- · online · post-inference, pre-publish
- · $/decision, sub-minute SLA
- · output → audit-grade proof