#4 — IronNet fixed-topology NEAT

Hypothesis — what it tests

Whether fixing the network topology (no topology mutations, no connection disabling) prevents NEAT's degenerate one-direction output collapse while preserving the apparent edge.

Description

Standard NEAT kept producing degenerate genomes with dead outputs — the V3 general genome had BUY stuck at a constant 0.9445 and went 100% LONG live (−791p in 10 days). IronNet fixed the architecture at 4 inputs → 4 hidden → 3 outputs, fully connected with skip connections (40 always-enabled connections), letting NEAT mutate only weights, biases and activations from {tanh, sin, cos, gauss} — dead outputs became impossible. A 3-phase pipeline (sine pretrain 30 gens → zigzag pretrain 50 gens → walk-forward P&L evolution 200 gens with a >=15% minority-direction hard gate) trained pilot pairs EUR_GBP and CAD_JPY. Evaluated at S5 cadence the M5-trained genomes reached EUR_GBP +3,421p and CAD_JPY +12,287p OOS (40%/74% more than at M5 eval), all passing Monte Carlo sign-shuffle at sign_p=0.0000, and were deployed head-to-head on accounts 001 (S5 fine-tuned) and 009 (M5). The topology fix itself was sound engineering, but the inputs came from the compute_asi_mc pipeline, so the RCA verdict applies: the backtest numbers were lookahead artifacts.

Key result

Indicators

Algorithms

Code