A systematic desk doesn't buy a feed on a demo. It buys after the backtest survives — and most news feeds fail there, because the vendor's "sentiment" column is a black box that changed silently three times over the sample period. Forecite was designed against that failure mode from the start.
Scores you can replay
Every verdict persists its raw sub-scores — novelty, materiality, surprise, specificity, directness for actionability; clarity, magnitude, consensus gap for conviction. The headline score is just a published weighting over them. That means you can re-weight history in a vectorized one-liner and test your own composite without ever re-running a model, and when we retrain, your research on the raw components still stands.
Labels come from the tape, not annotators: the Verdict Engine is trained on realized price impact across horizons. Your backtest and our training loop are looking at the same ground truth.
History that pages like a database
The REST API paginates on a (published_at, id) keyset, so a multi-year replay walks the corpus in stable, reproducible pages — no drifting offsets while the live feed appends. Filters compose the way your universe file does: symbols, exchanges, four tag dimensions, sentiment bounds, actionability flags.
Latency you can audit
Production is the same contract. The WebSocket feed delivers verdicts with per-event telemetry — ingest, parse, score, publish timings on every message — so you can measure our end-to-end latency on your own traffic and put the number in your slippage model instead of taking it from a sales deck. When the desk needs delivery guarantees rather than an open socket, signed webhooks carry the identical payload.
If your research stack speaks Python, the SDK gets you from key to first historical page in one call. The backtest is the pitch.