How a champion is born. Total autonomy, fifty generations, crime scene memory, weekend tournaments, and a single Hall of Fame. No human in the loop.
Every cycle the swarm births fifty candidates. The first thirty are wide-exploration mutations - the LLM is given total authority to rewrite every scaffold function and invent brand-new multi-dimensional engines. Generations thirty-one to fifty pivot to honing: the strongest candidate so far is handed back to the LLM as a parent for surgical improvement. Whatever emerges top of the leaderboard becomes the next champion.
The V4 swarm hands the LLM a bare scaffold and tells it to rewrite all of it. pre_filter, signal_entry, manage_position, the cache precompute, the DNA dictionary, even the hypothesis docstring that ends up on the Hall of Fame page. No template lock-in. No fixed indicator list. If the LLM wants Shannon Entropy and Hurst persistence instead of ADX, it writes them. The only hard constraints are pure NumPy execution, honest physics (one-hour window lookups, four-hour time exits), and a fifty-percent minimum win rate for survival.
Live losing trades are written to a structured crime scene log. Before each swarm cycle fires, the LLM receives the top recent failure patterns as part of the prompt — the exact chop bands, ADX zones, volume ratios, and regime labels that liquidated the previous champion. The result is that each new generation is mathematically aware of what killed its predecessors and builds vetoes accordingly. The Gamma Goliath’s “Chop Death Zone” filter was born this way, hard-coding a refusal to short in the 40-60 chop / 20-35 ADX band where four prior champions were liquidated.
Every signal that fires without the filter stack approving is logged to a shadow matrix. If a shadow signal would have made more than 2 percent profit, it counts as missed alpha. The ten most recent misses are fed back into the swarm prompt alongside the crime scenes. This closes the loop — the bot learns not just what to avoid, but also what it is wrongly rejecting.
The swarm fires on a systemd timer at 17:00 BST every Saturday. Fifty generations run unattended over roughly three to four hours. The crowned candidate is validated against the live adapter, added to the Hall of Fame, and deployed. The previous champion is archived, not deleted. This means every week the system is rewritten from scratch against the latest six-day market window. No champion stays on the throne unless it continues to beat its own successors.
For most of MAVERICK's life, when a sidecar engine started losing in a market it was not designed for, an operator had to spot the bleed and pull it by hand. As of May 2026 that gap is closed. A weekly governor reads every closed sidecar trade, attributes the loss to the engine that fired the entry, and assigns the engine to a stage: WATCH if it has had a single bad week, PROBATION if it has lost across two weeks or breached a small equity-budget cap (size cuts to 50% on next entries, the engine still fires but at half margin), QUARANTINE if it has lost across three weeks on multiple statistical conditions, or an immediate HARD KILL if a single engine drops more than 1.75% of account equity in any seven day window. Hall-of-fame engines with a prior winning month are protected from a single bad spell by an explicit double-evidence rule.
When an engine actually gets quarantined, a separate orchestrator picks it up on the next weekly fire and launches a fresh 50-generation 9-LLM swarm tournament pointed at the failed engine's source code plus its losing-trade regime fingerprint. The swarm writes new candidate strategies tailored to the conditions that broke its predecessor, and the tournament winner is the replacement. The existing hot-swap path injects it as a new sidecar without ever overwriting the live champion. Per-engine 7-day cooldown prevents oscillation. Every transition pushes a notification to the operator phone. No human is required between detection and repair.
The weekly Saturday tournament still runs. The self-healing loop runs on a different cadence and a different trigger: it acts when an engine actually fails, not on a schedule.