Each signal run collects a live market snapshot covering the top ~100 assets by market cap. The snapshot is assembled from multiple providers for resilience:

• Price data: Binance and Coinbase spot markets (primary), CoinGecko (tertiary fallback).
• Order book depth: Binance Level 2 order book — best bid/ask, spread, and cumulative bid/ask depth.
• Volatility: 15-minute and 1-hour ATR-normalized volatility from live candle data.
• Funding rates: Binance perpetual futures premium index (when available; marked DATA_UNAVAILABLE otherwise).
• Smart-money signals: Tracked wallet activity from the Wallet Watcher engine (24-hour lookback).
• Chart analyses: Optional uploaded chart screenshots processed by vision analysis.

Every data point is timestamped with its source and collection time. Assets with stale or missing data are rejected before reaching the AI model.

1. Collect snapshot — Gather live prices, order books, volatility, funding rates, and smart-money data for the tracked universe.
2. Trim for model context — Remove raw candle arrays; keep only computed features to stay within the model's token limit.
3. Reconcile open signals — Check existing live signals against current prices. Update status (target hit, stopped, expired) and compute outcome grades.
4. Call AI model — Send the snapshot to the BWC model (Nemotron-3-Super 120B) with a PEARL-structured prompt asking for ranked trade setups.
5. Validate model output — Server-side schema validation gates each candidate: conviction, validator score, entry zone bounds, confidence range, evidence score range, failure scenario count, and reasoning chain length.
6. Override TP/SL — Use the model's PEARL-reasoned take-profit and stop-loss levels when they pass validation (RR ≥ 1.0); fall back to volatility-scaled levels otherwise.
7. Smart-money boost — Increase conviction by 5 points on signals where recent tracked-wallet activity aligns with the direction.
8. Persist and alert — Store signals to the database, log scan events, and send Discord notifications for high-conviction setups.

Each signal includes a confidenceRaw score (0–1) generated by the AI model along with four evidence dimensions:

• Technical (1–10) — Price structure, trend persistence, volatility profile, and indicator alignment.
• Derivatives (1–10) — Funding rates, open interest direction, and basis.
• Narrative (1–10) — Social sentiment, developer activity, and news flow.
• Liquidity (1–10) — Order book depth, spread tightness, and volume profile.

After a signal resolves (target hit, stopped, or expired), the outcome is graded and a Brier score component is computed: (prediction - outcome)². These accumulate across all resolved signals to produce an overall Brier score — a proper scoring rule where 0 is perfect and 1 is worst.

Platt scaling fits a logistic regression to the empirical confidence-outcome pairs, producing a confidenceCalibrated value that better reflects the true probability of a winning signal. Calibration requires at least 5 resolved samples; below that threshold, raw confidence is used directly.

Calibration dashboard: /performance/ai-accuracy

The engine applies multiple gates before a signal is published:

• Conviction gate: Minimum conviction score of 70/100.
• Validator score gate: Minimum validator score of 7.0/10.
• Price deviation gate: Current price must sit within the entry zone bounds.
• Spread gate: Bid-ask spread must be below 0.5% (0.8% for low-cap).
• Data freshness gate: All data sources must be within a configurable max-age window.
• Heuristic skip: If the AI model fails to produce valid output or the fallback validator disagrees, the candidate is rejected.

The engine prefers no trade over a weak trade. If all candidates fail validation, no signals are generated for that run — the risk summary explains which gates rejected each candidate and why.

When primary data sources or the AI model are unavailable:

• Price data: Falls back from Binance → Coinbase → CoinGecko in sequence.
• Funding rates: Market as DATA_UNAVAILABLE rather than substituting synthetic values.
• AI model: Falls back to a heuristic validator that uses the same technical gates but without AI-generated rationale or multi-dimensional evidence scores.
• Snapshot hash: Every run records a hash of the (trimmed) market snapshot so signals can be traced back to their input conditions.

Every signal that enters the live state is tracked through its lifecycle:

• Target hit: Price reaches TP1 or TP2. Classified as a win (actual outcome = 1).
• Stopped: Price hits the stop loss. Classified as a loss (actual outcome = 0).
• Expired: Time expires without target or stop being hit. Classified as breakeven (actual outcome = -1).
• R-multiple: Distance moved divided by the initial risk (entry-to-stop distance). Reported as resultR.
• Failure mode: Records which specific level triggered the close (tp1_hit, tp2_hit, stop_loss, expired).

Outcome data is used for the calibration dashboard and for ongoing model evaluation. Historical signal data is never overwritten — immutable outcome events are appended to the signal record.