The Intelligence Loop

Scheduled refreshers pull on-chain, market, and social data continuously: radar convictions, early signals, catalysts, narratives, market sentiment, and the live intelligence feed. This is the raw material — volume acceleration, liquidity buildup, social surfacing, narrative momentum, capital rotation — before any judgment is applied.

Each engine's LLM takes the raw signal and forms a directional conviction: a 0–100 score plus a direction (bullish / bearish / neutral), grounded in authoritative live metrics. Critically, every prediction is logged at the moment it's made — symbol, conviction, direction, price-at-prediction, the market regime, and the evaluation horizon — so nothing is retrospective. What you saw is exactly what was scored.

When a prediction's horizon elapses, the resolver fetches the real price move and marks it hit or miss. A bullish call that rose over its horizon hits; one that fell misses. This is the ground truth: thousands of resolved outcomes, each a labeled data point tying a conviction level to a realized result. The wallet-health engine is scored the same way — its held basket is frozen at report time and repriced seven days later, so a portfolio call resolves without deposits or withdrawals distorting it. The token-authenticity audit engine resolves its verdicts against whether a token survived, got rugged, or was delisted.

The resolved outcomes fit a logistic model per engine mapping raw conviction to realized hit probability — conditioned on market regime and on discrete evidence features. The system also evaluates multiple horizons to find where each engine is actually predictive, and a drift detector watches for when recent performance diverges from the model, triggering an immediate refit so the model catches up to the new regime.

The learned weights flow back upstream: the feature calibration shows which evidence historically carries predictive weight, and that knowledge is credited into the curation prompts so the LLM rewards reliable evidence and discounts evidence that underperforms. Meanwhile, resolved outcomes are mined for patterns — empirical regularities stored as intelligence objects that self-improve as evidence accrues.

The result is a loop, not a feed: today's predictions sharpen tomorrow's probabilities.