Pulse is the desk's resident quantitative laboratory: eighteen years of exchange-true tape, a growing library of engineered market features, and a standing court where every candidate model is tried before it may say a word on the desk. Its proudest results are the refusals.
The engine's effort goes where the evidence says it pays: into engineering informative representations of the market — price structure, volatility regime, microstructure, time-of-day, trend state — validated feature by feature for incremental worth, then watched for decay as regimes move. Algorithms are interchangeable employees; the feature library and the validation court are the institution.
Every candidate faces walk-forward validation on years it has never seen, permutation floors that ask "would noise have scored this well?", and statistical gates that separate genuine skill from lucky correlation. The court's record speaks for its honesty: when the engine tested whether short-horizon direction could be predicted from price features — across two feature generations and every horizon from three to forty minutes — the answer was no, and that program was closed in writing rather than shipped in hope. What survived is different in kind: knowing when the market is about to move, without pretending to know which way.
Trained on eighteen years and validated on held-out months it never saw, the desk's first serving model estimates the probability that a meaningful index move lands within the next quarter hour.
On the Terminal it is one honest tile: a probability, its level, and the claim it is being held to — never a bare number, never a stale one dressed as live. It advises; it cannot place, size, or block an order.
Every reading the model publishes is recorded at the moment it speaks. Each evening the engine grades that tape against what the market actually did — realized precision versus the claims made at validation, in a scorecard filed to the desk's reports. A model may only be promoted past observe-only on the strength of those forward scorecards, and the promotion itself is refused by the platform unless the evidence is fresh, sufficient, and on file. Models that decay are retired by the same paperwork that hired them.
The standing rule: staged rollout constrains authority, never quality. The best validated model serves — watched, graded, and outrankable by the risk committee on every order, like everyone else at this desk.
The architecture is built for many small specialist models — each answering one narrow question exceptionally well, each individually validated, monitored, and disposable — rather than one grand predictor. The same registry, court, and grading machinery that governs the first resident is designed to govern hundreds, across instruments and timeframes, as the desk's research program widens.