Stated before fees, transfer costs and basis movement.
Our own apparatus caught a 2.31x undercount of trading fees and propagated the correction backwards through every affected report. That is how we know these figures are real.
We hold the payment. The price is hedged away.
Every exposure is offset the moment it is taken on. What stays on the book is the rate.
Neutrality here is arithmetic. The offset is placed with the position, sized to it, at the same time. No model decides when to be neutral, and the book never needs a view on where any asset goes next.*
The edge is thin. Execution decides whether it survives.
- Sizing is set by the live order book. Anything the engine could not exit at an acceptable cost is refused before it reaches a venue.
- A proprietary model takes the market's temperature across the venues we trade: how much liquidity is genuinely there, how violently it is moving, how much leverage leans on it. It scores how close conditions are to breaking. When the score turns, the book steps down ahead of the stress.
- No single reading moves capital. Independent inputs have to corroborate the call first. When they disagree, the engine fails closed.
- Unwinds are asymmetric. The engine sheds what is exposed and keeps what still pays. On any input it cannot trust, it closes out entirely.
- Five venues are fenced individually. The book trades its own subaccounts. The escalation ladder is tested under synthetic injection. Nothing goes live that cannot be turned off.
Research reaches production without a rewrite.
Two AI members hold standing mandates, one on research and one on engineering, working with the firm's written history in context. The same apparatus that runs the research writes the production services, so a result that survives validation becomes live code in days.
Published work is pulled, filtered and tested against our own data on a standing cadence. Most of what we test is killed. What survives reaches capital through the same validated pipeline every time, and the record of what failed is kept as carefully as the record of what worked.
The trading system is small enough to change. A new signal does not require a new engine, and the universe, the sizing and the thresholds are re-fitted from measured data as conditions move.
Two partners, two AI members. Humans decide. The apparatus remembers and executes.
Work by either AI member is reviewed by a model from a different vendor before it can touch capital, against a written registry of invariants.
The strategy is live and paying.
Four members run the whole operation. No back office, no research floor, no distribution desk. The firm's fixed cost base is small, and it does not grow when the book does.
That is an economic position and not a boast about frugality. Operating cost is deducted from a gross return before anyone receives it, so a firm that runs lean hands over more of what it earns. What widens the book here is capacity we have measured. Hiring does not.
If you build in this space or allocate to it, we should compare notes.
contact@aleph-capital.ai