What the engine outputs: reading order, states, the grade scale, factor weights and grounding.
The verdict reads in decision order rather than sales order: invalidation first — the price at which the thesis is wrong — then the grade and its calibrated confidence, the confluence breakdown, the risk arithmetic, the thesis, scenarios, and coaching ranked by where the score has the most headroom.
A verdict is in one of three states — setup, no-setup or suspended — and carries one of five signals: strong-buy, buy, no-trade, sell or strong-sell. Signals are derived deterministically from the engine’s own numbers; the language model never writes one. Every no-trade names its single deciding reason.
A score between 0 and 100 maps to ten bands from A+ down to F. Thin data caps the band regardless of how clean the setup looks: too few bars of context, too few confirmed swings, or no readable higher timeframe each impose their own ceiling, and each names itself on the verdict.
Nine base factors, each scored between −1 and +1 and individually weighted: pattern, trend, multi-timeframe agreement, location, reward-to-risk extension, volume, momentum, structural confirmation and volatility regime. Crypto adds three more. Weights always sum to exactly one for the class being graded.
The confluence score measures how much of the checklist agrees. Confidence is a different question — what actually happened to comparable setups — and it frequently sits well below a high score. It is built by nested shrinkage: a replay curve first, then the platform’s record, then your own.
Two narrators share one contract: a deterministic composer that needs no API key, and a model rewrite in the house voice. Every generated sentence passes a grounding verifier — a sentence citing a number the engine never computed is rejected, and the deterministic bullet takes its place.