How the organism evolves. Eleven structural paradigms rotating sequentially. Tier-gated execution that scales with liquidity. A feature-invention pipeline that grows the knowledge base one crown at a time.
Each generation is assigned a specific structural paradigm by the rule gen % 11. This prevents the LLM from converging on any single recipe. By generation 11, every paradigm has been tested at least once. By generation 50, each paradigm has been tested four to five times. The swarm explores the full structural space, not just the most recent champion's lineage.
Not every symbol deserves the same stop-hunt protection. MEGA majors like BTC and ETH move on genuine institutional flow; their snap-ATR allowance is wide. LARGE alts sit in a middle band. OTHER low-liquidity names get tighter snaps because their wicks are more often manipulation than price discovery. Enforced at the entry gate via get_symbol_tier() and get_snap_atr().
Champions may invent novel composite features as top-level feature_* functions. When a candidate crowns, each feature is AST-extracted from its source, validated on a 7-day live-data window (NaN rate below 5 percent, non-constant variance, runtime below 100 ms per 1000 bars), and appended to the permanent knowledge base. Future generations can then import and combine them. Features that fail validation are ignored. Only features that survive join the library. The swarm earns plus 100 per invented feature, capped at plus 500 per candidate.
The V8 Calmar scorer ranks candidates by dollar profit per dollar of drawdown, not by win-rate distribution. This values honest trend-following profiles that professional CTAs run (25-40 percent win rate, large winners, disciplined stops) equally with higher win-rate strategies. V7 Sortino is logged in parallel for audit so scorer-calibration drift is always visible. The formula caps Calmar at 5.0 so institutional-grade risk discipline doesn't infinitely compound its own reward.