An operating system
for trading decisions
Eight stages run on every candle close, in the same order, every time. Data comes in raw and leaves as an order or a written reason there wasn’t one — and every intermediate state is kept, so any decision can be replayed exactly as it was made.
8
pipeline stages
< 80 ms
close to order
100%
decisions recorded
▸ pipeline online · 8 stages · deterministic replay enabled
Close to order, one path only
Ingest
Websocket and REST feeds from every connected venue are normalised into one internal candle and book representation, checked for gaps and duplicates, and stamped with the venue clock alongside ours. A strategy never sees a venue's quirks.
Three models. One arbiter. No averaging.
Averaging disagreement produces a number that no model would have chosen. The arbiter instead weighs each model by how well calibrated it has been in the regime the market is currently in — so a momentum model that is unreliable in compression is quietly demoted rather than allowed to vote.
market geometry
regime-conditioned
memory recall
arbiter rationale
All three agree; momentum is well calibrated in expanding trend, so it carries near-equal weight.
representative decisions · not live market data · hover to pause
Always on, and always narrow
Roughly eighteen hundred frames a second come in. Four hundred decisions an hour come out of the arbiter. Nine of them become orders. The funnel narrows by design — the engine's job is mostly to decide against doing something.
Services, not a script
Each box is a separately deployed service communicating over an event bus with replayable envelopes. The consequence that matters is the one on the right: there is no edge from the engine to execution that does not pass through risk.