Setup tags come from the one-hot flags in the feature vector: is_bull_setup, is_bear_setup, is_momentum, is_reversion, is_breakout. If exactly one is set, that's the tag; otherwise it's tagged mixed.
Edge = accuracy − 0.5. Positive means the brain has real signal in that setup type.
Brier = mean (predicted − actual)². Lower is better. 0.25 is the no-info baseline.

• Low-edge setup types (accuracy < 55%): downweight those setups in trade-selectivity rules or raise the conviction threshold for them.
• High-edge setup types (accuracy > 60%): boost size or alert more aggressively when those setups fire.
• "mixed" dominating: features aren't being tagged cleanly. Audit FeatureExtractor — maybe one-hot flags need stricter classification.
• This is independent of the brain's overall calibration / Sharpe — it's purely about whether some types of trade setups are easier to predict than others.