In a portfolio-management agent, a language model can turn messy news into structured numerical features the agent can use as part of its state. It encodes information; it does not make the trading decision. section 12.0 section 12.11
Keep the encoder-agent boundary explicit:
- Expose the encoded signal through a fixed-length, bounded, named, validated, versioned schema. section 12.11
- Treat representation as consequential: information discarded from the state cannot be recovered by the algorithm. section 12.11
