megarexus

One brain. One repo. Operate here.

MegaMind — Internal dreams · External judges · Central LoRA commits · operator garden pad.

Import	from brain import MegaMind
Primary docs	docs/PRIMARY.md
Pad status	synthesis_lab/orchestration/rdl/brain/STATUS.md
Lab (not primary)	docs/LAB.md
Mathematics	docs/MATHEMATICS_OF_WORLD_CAPTURE.md
Docs index	docs/README.md

Operate

# 1) Smoke — every planted idea on the pad
CUDA_VISIBLE_DEVICES= python -B synthesis_lab/orchestration/rdl/brain/smoke_all_ideas.py

# 2) Train flywheel — 24 games, level groups, keep-if-better
python -u -B synthesis_lab/orchestration/rdl/brain/train_flywheel.py \
  --root synthesis_lab/experiments/megamind_flywheel \
  --device cuda --max-cycles 0

# stop
touch synthesis_lab/experiments/megamind_flywheel/STOP

from brain import MegaMind

mm = MegaMind(device="cpu")  # or "cuda"
state = mm.observe(frame)   # HxW color-index board

Layout

brain/                    # public API
docs/PRIMARY.md           # single source of truth
docs/assets/              # hero art
synthesis_lab/orchestration/rdl/brain/   # MegaMind implementation
  operators/              # Loom · Rexus · pattern growth · Weco RSI · …
  train_flywheel.py
  smoke_all_ideas.py
legal_search/algorithms/  # operator backends (source only)
codexlab/contra_head/     # CONTRA view backend
synthesis_lab/experiments/other_games/environment_files/  # 24-game suite
notes/                    # doctrine sweeps (+ archive/)

Doctrine (short)

1. One brain — MegaMind only.
2. Plant useful channels; training raises what works.
3. Multi-game level groups — robust patterns, not one-game grind.
4. Keep only if better on private multi-game metrics.
5. Env grades only — no spill / win-bank / human-route train labels.

Honesty

Specialist scorecard policies and legacy CNN experiments are lab, not the primary. See docs/LAB.md.