Data & methods

The two data planes

• Live plane: Binance USDⓈ-M public REST and websocket endpoints, requested by your browser while the terminal is open. Prices, funding, open interest, order flow, klines — all public market data.
• Model plane: the authenticated forecast endpoint runs the qualified own joint hourly generator plus TRAIN-only direction expert against 6,144 completed five-minute candles per asset, with causal funding and disabled untrained OI. One in-flight result and one atomic joint cache serve all three markets at the same input close and model/code identity. Changes to the generator, expert, calibration, qualification evidence or inference code invalidate the cache. A labelled previous cycle can display while the next completes. The archived research artifact distinguishes one-hour reused historical confirmation from exploratory longer-horizon development. Strategy research remains independent of model forecast scores.

Fixed honesty rules

• Unavailable data renders an explicit state — never zero, never a stale number pretending to be fresh.
• Missing forecasts render unavailable when there is no retained result for the same market and model. Retained results keep their original input timestamp and show updating or retry status; they are never fabricated neutral probabilities or an old-model fallback.
• Token-model evaluation uses chronological train/development/holdout splits; normalisation and codebooks fit on training data only. The terminal refreshes inference for each new five-minute close. Training and holdout evaluation run offline.
• Fees are included in net figures wherever a net figure is claimed.
• Gain/loss color semantics never change, under any theme.

What this is not

PerpIntel is research tooling. Nothing on the desk is financial advice, a solicitation, or a guarantee of performance — markets in perpetual futures are brutally volatile and you can lose everything. The terminal's job is to make the state of the market legible; the decisions stay yours.