AI & Architecture
Dual-Agent Debate Engine
Three independent personas argue both sides before you take a position.
Every market ships with two structured stances ā a bull case and a bear case ā plus a three-agent debate transcript. The goal is not to predict for you; it is to make the disagreement legible before you commit capital.
The three personas
| Macro Hawk š¦ | Rates, liquidity, term structure. Zero tolerance for 'this time is different'. |
| Onchain Detective š | Whale wallets, DEX flows, stablecoin net issuance. Only balances count as evidence. |
| Meme Master šø | Call density, community heat, KOL consensus as a contrarian signal. |
Why three and not one
A single model produces a single confident answer, which is exactly the failure mode of prediction. Splitting the reasoning into three constraints ā fundamentals, flows, sentiment ā forces disagreement into the open, and the market's implied probability becomes the aggregation mechanism instead of any one model's opinion.
Evidence trail
- Each claim can carry an IPFS CID pointing at the snapshot it was derived from
- At settlement the operator writes a bytes32 evidence hash on-chain (settlementHash)
- Anyone can re-fetch the snapshot and compare it with the recorded hash
- The debate text itself is off-chain content, clearly labelled as AI-generated commentary ā never as on-chain data
This is not a signal service
Agent stances are editorial content, not financial advice, and they carry no on-chain authority. The only on-chain facts are the pool balances, the settlement hash and the timestamps.