Alpha-v2.7 is built on three levels: a feature encoder, an ensemble distributional critic, and a portfolio generalizer. The defining shift versus v2.6 is the policy side: instead of an independent per-symbol actor, v2.7 uses a single cross-symbol PortfolioGeneralizer that consumes the full Q-distribution across all symbols and horizons and emits portfolio weights jointly.
Result. "Buy BTC" is no longer decided in isolation — it is decided with the explicit, fully-distributional Q-value of the whole portfolio in view. The encoder is shared between critic and policy and is trained by the TD signal only, so representation features carry value information, not the shape of the actor's output.