NoBanks/debate-desk

Two agents argue about SOL from the same evidence. A third judges. Fixed rules decide whether the ruling may trade. Every decision lands on Solana with its whole reasoning trail committed on chain, in the same transaction as the trade.
Built for the BLI Legal Tech Hackathon 2 (RYO-CHAN Autonomous Agents bounty, Chainlink CRE bounty).
Live record: https://nobanks.github.io/debate-desk/
Printable Decision Record for any receipt, re-checked in your browser: https://nobanks.github.io/debate-desk/record.html
Chainlink CRE evidence anchor on Ethereum Sepolia: 0xf917e04fec9dba42c485cd2d3425d7c7dbc94566
Desk wallet (Solana devnet): FHLgLhiFimjkjP34cuGHQZTfy9DvY4QphskxPMsSk14B
Desk's on-chain log: Ahw9K3YcVcaKZTKyx7u5QBCfoMKLLjhZz5ZFDqm3Y5gR
Receipts program: GkCqpfzoMVSCMw4muiKAXoo4kcHFRiYjjAHBQjmhMwmU (devnet)
Everything the desk wallet did before its log existed, so nothing is off the record:
1. received 2 devnet SOL from a test wallet: tx
2. one seed swap, 0.4 SOL to devUSDC, so it could act on a BUY (scripts/seed.ts, refuses to run once registered): tx
3. registered with the receipts program, committing its policy hash: tx
From then on, every action is a receipt.
## The problem
An agent that trades gives you a log file its operator controls. "It bought because the news was bullish" is a claim written after the fact. You cannot tell whether the evidence existed, whether the agent read it, whether the reasoning came before the trade, or whether the losing calls were quietly deleted.
## What Debate Desk does
Every cycle:
1. Evidence. The desk gathers one evidence pack: Pyth SOL/USD read straight from its Solana account, the Chainlink SOL/USD Data Feed (delivered by a Chainlink CRE workflow that also anchors it on Ethereum Sepolia, see cre/), Coinbase, Kraken and CoinGecko market data, recent crypto headlines, and what public market voices (KOLs) are saying on Telegram (see below). Each item is normalised, hashed and saved under data/evidence/.
2. Debate. BULL and BEAR each build the strongest honest case for their side. They run independently and read the exact same evidence text. Every claim must cite evidence ids and state cause and effect.
3. Ruling. The JUDGE weighs both cases against the evidence itself, lists the points it rejects and why, and makes one decision covering entry, size and exit: BUY, SELL or HOLD, conviction, stop loss, take profit, horizon.
4. Rules. Plain code (src/risk.ts) decides whether the ruling may trade. It holds if the price is stale, if price sources disagree by more than 150 bps, if the judge cites evidence that is not in the pack, if the ruling contradicts itself, or if conviction is under 60. It clamps the exit plan to policy and sizes the ticket. If 2 or more public voices converge against the verdict on SOL, it cuts the size in half; agreement never adds size. Every check is written into the receipt, pass or fail.
5. Receipt and trade, one transaction. The receipt holds evidence hashes, transcript hashes, the ruling, the rule trace, the position and the Orca quote. Its sha256 goes to the Agent Receipts program in the same transaction as the Orca swap. The program rejects a BUY/SELL receipt without the desk's own swap beside it, and a HOLD receipt with one. A trade and its reasoning land together or not at all.