The Verdict Engine retrains every week. That cadence is the product — markets drift, and a model scored against last quarter's regime decays quietly. Here is how we ship a new model every Friday without white-knuckling the deploy.
The tape is the label
We don't label headlines by hand. Every scored event is joined against realized price moves across multiple horizons, and those outcomes become next week's training labels. The ground truth is what the market actually did, not what an annotator guessed it would do.
Evaluation before promotion
A candidate model replays the full week of events shadow-mode and is diffed against the incumbent on calibration, ranking quality, and per-sector drift. Only a candidate that wins on the replay gets promoted — and the raw sub-scores we persist with every verdict mean we can re-weight historical scores without ever re-running a model.