Mercurius.
Mercurius.
A council of agents, trading live. No single point of failure. Consensus before any position.
Built alone, from a single conviction: that a disciplined committee of narrow agents would beat one model trying to do everything. Ten instruments — commodities and indices. Five specialised agents vote: a market scanner (Sentinel), a macro strategist (Atlas), a fundamental reader (Oracle), a sentiment reader (Pulse), a technical analyst (Cipher). An Event Guard sits outside the vote with a single power — to veto. An Arbiter counts the rest: three of five must agree before a position opens, and conviction sizes the bet.
The first version proved the idea and exposed its flaws in the same breath. Across the opening run it opened hundreds of positions and logged thousands of individual trades — far too many, agreeing with itself far too readily, conviction spread thin across a council that was too large to be decisive. The thesis was sound. The discipline wasn’t. So we rebuilt.
Version two is leaner and harder on itself. The council was cut to the agents that could earn their vote, wrapped in a seven-layer guard — anti-stacking, daily caps, instrument cooldowns, regime gates — and closed with a self-learning loop: every agent is scored against its own track record, and the ones that stop working are quietly demoted. Not machine learning, not a model — just a system that keeps its own scorecard and acts on it. AI in the engine is the next step, deliberately not the first.
It runs a £100,000 book on CFDs — virtual capital while version two proves out, but live prices, live sizing, live execution. The track record resets when the system does; this run starts at version two. What you see is the refinement in progress.




