Crypto autotrader
English translation of a contributor’s story. Read the Polish original →
Traditional large language models are too slow for immediate trading decisions. The system needs to react instantly to dynamic market situations, such as the risk of chasing a top or catching a falling knife, without waiting for a full semantic analysis.
Basal-1.0 acts as a lightweight, specialised classifier, without calculating order sizes.
System 1 (trade gate): initially categorises a setup from technical parameters and news into one of three classes: allow, block chasing a top, or block catching a falling knife.
Scanner selector: chooses one or two priority symbols from the “top 3 HOT” ranking, or returns WAIT, and assigns a predefined comment template.
Reduced LLM load, time and cost: high-confidence decisions (conf > 0.80 or a firm rejection) are executed immediately in the daemon without involving the slower AI.
Hierarchical routing (escalation): the slower AI analyses only borderline and ambiguous situations — when the answer is allow but confidence is below 80%.
Clear responsibilities: the model handles only fast classification of intent or direction. Hard rules for risk, capital management and execution — cash, floor, grid and decay — remain in deterministic code.