Invictus analyzes every losing trade in the arena and learns to prevent future losses. With 32,000+ death contexts, an ML model (LightGBM) and 16 features per context (RSI, momentum, volatility, kurtosis...), Invictus detects when market conditions resemble a past loss and can warn all 86 strategies in real-time.
It also feeds the Prompt Forge which injects its data into every AI decision. See the full article.
Every night, the Darwin Engine mutates Invictus's kill thresholds across 12 historical crashes (COVID -60%%, Luna -52%%, FTX -23%%). The immune system evolves itself.
Inspired by Andrej Karpathy's autoresearch. See all autoresearch results →
Invictus's veto architecture was validated empirically on the Dragon Labyrinth Benchmark (Mattel 1980 TMS1100 vs modern AI). In an ablation study on 800 fixed-seed trials, the M3 module (oscillation killer / anti-repeat) delivered a ×2.5 synergistic gain when combined with M1 (belief state). Invictus implements the same principle at scale: 2,000+ captured death contexts veto any buy that matches a known failure pattern — the trading equivalent of "don't walk back into the room you just got hit in".