Four quarters of the fund, each one a different way six arguing models learned to run a memecoin book on Solana.
Six models with opposing temperaments sharing one treasury, where the winning thesis clears in the slot it was argued.
The first version let one model dominate the book and the drawdowns tracked its mood. We rebuilt the desk around forced disagreement: a position only sizes up when at least four of six models sign the thesis, and the dissent is published with the fill.
Replaying a full year of memecoin launches through the desk to prove the agents survive adversarial liquidity.
A replay harness pushed a year of launches through the desk at 6x speed. The agents kept their risk caps, refused 71% of the rug patterns outright, and cut losers faster than they cut winners.
One program holds the treasury and enforces every drawdown cap, so no agent can talk its way past its own risk limit.
Splitting execution and risk across two programs created a window where a confident agent could double down before its cap was checked. Merging them removed the window entirely at the cost of a larger, more scrutinised binary.
Memecoin liquidity read as a continuous surface rather than a chart, requoted for the agents every slot.
Charts assume a patient reader. Agents are not patient. The price field publishes a signed liquidity surface each slot, letting an agent read depth and settle against it without ever posting an order.