The research system

Adaptive scoring engine

The evolving scoring dimensions that combine into one final score — and how new formulas earn their place.

The scoring engine turns the agents’ evidence into numbers. It continuously evolves: the AI can propose new dimensions and new formulas, but only validated scoring models ever become active.

Current dimensions

Dimension	What it captures
Signal score	Quantitative evidence from the data.
AI score	The agents’ interpretation of the market.
Risk score	Risk evaluation — contract, liquidity, drawdown.
Memory score	Similarity with historical situations the lab has seen.
Strategy score	Confidence in the currently active strategy.
Execution score	Expected execution quality (slippage, fees, liquidity).
Final score	The weighted combination that drives the decision.

How a score is combined

signal_score_v1

Weighted, decomposable, penalized

S i g na l S cor e = i \sum w_{i} C o m p o n e n t_{i} - w_{r i s k} R i s k P e na l t y

A normalized, versioned blend of momentum, volume, breakout, liquidity and regime components, minus a risk penalty. Both the total and every component are stored, so any score can be explained.

A high score never overrides a hard rule

Scores rank candidates; they do not authorize trades. A signal with a strong score is still rejected if any deterministic safety rule fails — see the Risk engine and Deterministic scripts.

Only validated formulas go live

The AI can propose a new scoring dimension or reweight the combination — but a proposed formula must be validated on historical and shadow data before it replaces the active one. Scoring changes are a governance decision: the human approves them.

Why scoring is adaptive

Markets and narratives change, so a frozen scoring formula decays. By treating scoring as a hypothesis that can be challenged and improved — and gating every change behind validation and human governance — the engine keeps pace without ever silently rewriting how decisions are made.