Tune the engine.
Every threshold here changes when calls fire, which signals contribute, and where they get dispatched. Paper mode is locked on until you accept the disclaimer below.
Model profile
A saved configuration — signal mix, confidence bar, cost band and damp — applied as one unit. While a profile is active it wins over the manual settings below. Bankroll, caps and the kill switch always stay manual.
Manual (current settings)
The gates and signals configured on this page.
Live-trading gate
Paper mode is hard-locked on. Unlocking the live toggle requires an explicit acknowledgment. Even after unlock, paper mode stays the default until you flip it.
Disclaimer
15-minute BTC direction is close to a coin flip. After 5 bps taker fees and 2 bps slippage, anything below ~54% hit-rate is net negative. Run at least two weeks in paper mode and inspect the per-signal hit-rate before flipping live. Tape is not your broker.
Trading caps
Hard limits on real-money orders. Even when live trading is on, the scanner still respects paper mode and will only place orders when paper mode is off.
Thresholds & gates
How strict the engine has to be before it commits to a call.
Auto-tune the confidence gate
Periodically recalibrates the confidence threshold from the last 7 days of resolved calls using Brier score and hit rate. It is biased toward no trade: it only loosens the gate when the sample is large enough (n≥20), the hit rate clears your target, and calibration beats a coin flip. Thin samples, poor calibration, or a sub-54% hit rate tighten it instead — and tightening moves faster than loosening.
tightening to 0.58 — Brier 0.354 no better than a coin flip (n=155)
Entry-cost band
Refuses any entry priced outside 35c–70c per contract. Across every settled, contract-bound trade since the logic epoch, entries inside the band returned +$344 on 46 trades, while entries above 70c lost −$106 and entries under 30c lost −$48. Rich contracts still win often — they just pay too little to survive a normal loss rate.
Regime-conditional signal weights
Fits a separate weight multiplier per signal for trend and chop tapes from the last 30 days of graded, contract-bound calls. A bucket is only adopted when replaying it through the score gate improves both win rate and average P&L — otherwise your live weights stand. Multipliers are clamped to 0.5×–1.5×, so no regime can switch a signal off.
no regime bucket improved both win rate and P&L — live weights kept · trend: candidate fires only 9 of 113 (need 12) | chop: candidate fires only 7 of 107 (need 12)
Throughput tuner
Keeps a minimum number of graded windows per day so calibration has something to learn from. When the trailing 24h falls short, it temporarily loosens three gates one notch at a time — confidence threshold, counter-trend damp, and the fair-value tradeable band — and reverts to your defaults the moment the target is met or the relax window expires. Direction logic is untouched; only the gates move, and only temporarily.
Signal mix
Disable a signal when its 7-day per-signal hit rate sits below 48% — it is actively hurting the composite. Weights multiply each signal's z-score; set to zero to mute without disabling.
Webhook dispatch
Fired calls are POSTed to your endpoint with an HMAC-SHA256 signature in the X-Signal-Signature header. The shared secret lives server-side as INGEST_HMAC_SECRET.
When 7-day cumulative P&L falls below this floor, webhook dispatch is automatically disabled and a banner is shown on the dashboard until you re-arm.