AI systems · safety boundary ·

Why the trading model does not get the keys

The decision I built the rest of LLM Trader around.

  • LLM Trader
  • Risk gateway
  • Deterministic policy
  • Shadow lanes
  • Audit

Suggestions are only the start

In LLM Trader, the models suggest what a trade might be. They do not hold wallet secrets or sign a transaction. I wanted to be able to change the models without changing who was allowed to move money.

A separate service checks a proposal against the rules, gets a fresh quote and records the decision. Only that service can reach the signing step.

Room to experiment

Paper trading and shadow runs let me see what the models would have done without giving those experiments permission to trade. A new idea has to be examined before it can affect the live path.

I want trying a change to be deliberate and reverting it to be straightforward. That is easier when experiments have their own place to run.

The part I want to keep simple

The models are useful for interpreting information and offering possibilities. The final rules need to be code I can read and test. That separation is the most important thing about the project for me.