The trading layer
Strategies have to earn their capital.
Eve writes and tests its own trading strategies. Almost all of them are rejected, which is the feature.
Named for the probe in WALL·E: scan everything, reject nearly everything, and only verified life rides home. Here, life is a measured edge.
01The loop
Every agent runs the same five stages.
- ingest
- signal
- gate
- execute
- learn
Pull the world in. Prices, filings, odds and balances, on a schedule that does not care what time it is.
Turn raw feeds into a claim with a number attached. A direction is not a signal until it has a size.
Every claim meets the risk gates. Most die here, and that is the point.
What survives reaches a broker or an exchange. Position, slippage and fees are recorded as they happen, not later.
Outcomes feed the next decision. A strategy that stops paying has its capital taken away.
02The gate
How a strategy earns real money.
- Strategies are generated, not hand-written
- A nightly pipeline proposes candidate strategies per category and per symbol, then backtests them. The generator is allowed to be wrong as often as it likes, because nothing it produces is trusted on the strength of its backtest alone.
- In-sample fitness is not a result
- A candidate is only promoted if it clears walk-forward and holdout validation, which replay the signal over a window it was never fitted on. An earlier version of this pipeline gated on in-sample fitness while computing walk-forward and ignoring it. Finding that was worth more than any strategy in it.
- Every input feed must be point-in-time
- A feed that returns only today's value makes the replay look-ahead biased, so a category built on one can never honestly promote no matter how good the edge looks. This kills most new category ideas before any edge research happens, which is the cheapest place to kill them.
- Windows are sized in events, not months
- The gates count trades. An event-driven category needs enough events inside the window to clear them, so sizing a backtest in days quietly guarantees that an earnings-driven engine can never pass.
03The graveyard
What it measured, then switched off.
Each of these was built, run against real markets, measured, and retired on the evidence. Keeping the record is what makes the surviving engines mean anything.
Directional BTC
Measured to have no edge on any horizon tested. Retired as a dead end rather than tuned. The range and structure work that replaced it is the part that actually trades.
Baseball
Lost money over its run and was decommissioned. It stays decommissioned: restoring it needs new evidence of an edge, not a new set of parameters over the same data.
Soccer
Across nineteen series and roughly 2,200 deduplicated trades, no series showed a significant edge, and the model's own edge signal was inverted: the bucket it was most confident in realised the worst outcome. Raising the minimum-edge bar would have made it worse, not better.
Copy-trading
The traders it was set to follow turned out to be running multi-outcome arbitrage, so copying a single leg would have been naked unhedged risk rather than a shared bet. It made 31,833 attempts and placed zero orders before being removed.
“The best way to understand a system is to build one that can lose your money while you sleep.”
04The invitation
Build something
that runs without you.
Type it here — it lands on my dashboard and pages my phone.