Wallets under management
2
Owner + partner · ~$393 total
Farah is an autonomous agent that trades two real Solana wallets. In one working session we audited it until it admitted it had no edge — proved that on 18 months of real data, said so plainly, unlocked a wallet that had been frozen for two months, and rebuilt the whole operation into a risk-first desk with capital parked in guaranteed yield until it earns the right to trade again.
The bot reported a +$27 lifetime profit. The forensic audit found that number was an accounting fiction: it contained one fabricated +$30.42 trade. Reconstructed position-by-position, the real record was 142 completed trades, a coin-flip 50% win rate, and roughly +$8 gross — turned net-negative once transaction costs were counted. Backtested on the real market, the old strategy would have lost nearly half the account.
The disciplined rebuild does something the old bot never could: when price is below its long-term trend, it refuses to trade. Through the October 2025 cascade that ~halved the market, that single rule was the difference between −42.9% and −3.4%.
The partner's wallet sat $0.10 short of the gas it needed to transact since roughly May 1 — and the safety floor that blocked new trades also blocked the one swap that would fix it. A +130% winner was stranded, unrealizable, while the monitoring said "OK" the entire time. Mid-session, we topped up $8 of the wallet's own funds and watched the next cycle finally settle the queue.
$0.10 of missing gas had frozen a +130% position for two months.
"The process is running" is not "the money can move." The old health check confused the two for two months. The new one never will.
Live-money evidence is brutal on autonomous traders — in public contests, most frontier models lost up to 63%. So Farah's new mandate demotes the agent from order-originator to veto-only risk officer: a deterministic engine proposes, the agent can only halt. Everything below it is machinery built to make failure loud and losses small.
The constitution: veto-only role, hard risk rails, a rule that hostile input is never an instruction — the exact defense against the prompt-injection drains that hit other agent wallets this year.
A separate process, every 5 minutes, that halts trading on runaway activity, a 15% drawdown, dust regression, or a stalled loop — then alerts the owner. It caught a real anomaly the day it deployed.
18 months of real candles, a no-lookahead engine, a mutation-proven test suite. It's what turned "this strategy might work" into "neither strategy is certifiable — don't deploy."
Reconstructs real positions from the trade log, flags glitch artifacts, reports P&L net of cost. It replaced a reporter that had been printing zeros since February.
The backtest was honest: at this account size, no live strategy clears the bar. So the disciplined move isn't a cleverer trade — it's to stop guessing. We deposited $220 into Kamino, the largest, longest-clean USDC lending market on Solana, at ~4.4% APY — more than the bot's entire lifetime P&L, guaranteed, in a segregated wallet the trading engine can't touch.
Every figure on this page is a measured result. Click any hash to verify the transaction on-chain.
We'd rather be right than optimistic.
Most "AI trading" pitches lead with a return. This one leads with disproving its own — because the honest ceiling at this size is tens of dollars a year, and pretending otherwise is how accounts get drained. What was actually built is validated machinery: the audit that catches the lie, the kill-switch that stops the bleed, the backtest that vetoes the bad idea, and the yield that beats the alternative today. Capital scales into a system with evidence behind it — not hope.