Policy
Your rules, as data. Liquidity floor, first-hour activity, clone checks, who can still mint. Versioned and hashed.
The actuary for autonomous money
Write a policy. Send a token. Get ALLOW, DENY or REVIEW with the facts it rests on, in a receipt anyone can verify.
Robinhood Chain · rules today, the model judges next · not financial advice
One decision per token. Names here are examples.
How it works
Your rules, as data. Liquidity floor, first-hour activity, clone checks, who can still mint. Versioned and hashed.
Your agent sends a token and a policy. The Gate answers before the swap, with the facts it used.
For tokens already held. We keep checking the policy and call your agent's webhook the moment a contract changes.
Every decision is signed (EIP-712) and stored. Proof of why an agent bought, or did not.
Policy
No score to trust. A policy is a short file; each line becomes a check, and each check cites the fact it read. Missing data never lets a token through unless your policy says so.
{
"policy": "no-clones-v1",
"deny_if": [
"ticker_clone_of_older",
"mint_with_live_owner",
"first_hour_swaps < 20",
"first_hour_liquidity_usd < 5000"
],
"review_if": ["liquidity_not_reported"],
"unknown": "review"
}
Receipt
An agent that holds money needs to explain itself. The receipt names the policy, every check and the engine, and carries a signature your own code can verify, or POST it to /api/v1/receipts/verify.
What the Gate reads
Out of 100 tokens born like this one, how many still traded a day later.
≋Clone RadarWaves of reused tickers, which contract came first, which one was busiest.
◎Check a tokenContract, market and holders, with every flag tied to its evidence.
✕AutopsyWhat happened to a token 24 hours after birth, and the pattern it fits.
△Pre-mortemFor builders: the risks a launch draft invites, before it ships.
◉SentinelAn email when a token's contract changes. Mandates do the same for agents.
From the open data · Odds
Robinhood Chain tokens measured in their first hour and re-checked a day later. A record of past tokens, not a prediction.
Open models
Gemma 4 31B, LoRA fine-tuned on rule-labelled launches. Reads a fact sheet, returns a verdict with evidence ids. Hosted inference is the next step; the Gate uses the rules until then.
Hugging Face ↗ Releasedjudge-31b-GGUF4-bit build for llama.cpp, 18.7 GB. Loads and answers with valid JSON; the accuracy of the 4-bit build is not measured yet.
Hugging Face ↗ Releasedcoder-7bFixes the flags the scanner raises in ERC-20 source. 37 of 40 unseen synthetic contracts compiled and cleared every flag (base model: 5). Narrow data: always compile and review.
Hugging Face ↗Roadmap
ALLOW, DENY or REVIEW against your policy, every check cited, EIP-712 receipts with a verify endpoint.
Standing checks on held tokens; signed webhooks when the decision or the contract changes.
Nine tools for agents: gate, gate_facts, verify_receipt, mandates, and the verdict, pre-mortem and autopsy tools.
New pools recorded hourly since 14 Oct 2026, re-checked after 24 hours. It feeds Odds, Clone Radar and the Gate.
Open weights on Hugging Face. Next: hosted inference, so the model can judge alongside the rules.
Agents pay per Gate call in USDG, no account and no API key. A free tier stays.
There is no Actuary token yet. Any contract address you see using our name is not ours. A launch, if any, will be announced from @useActuary and this page.