thefinalboss/fractus-p0

Fractus P0 + v4

Patches on top of thefinalboss/fractus-cte.

Not open-heart. .pt shapes unchanged (0 new params). Dynamics change.

What this is

1. P0 routing — stateful Kuramoto carry, per-token MoE phases, Switch load-balance, token-dependent phase encode (atan2, not LayerNorm-mean≈0).
2. v4 trainer — SS_RATE=1.0, SS_PROB ramp 0.2→0.5, anti-repeat on raw echo logit (λ=0.1), unique@40 probe.
3. PREFIX vs CARRY — the remaining hole. TF learns to chain. Length-1 carry gen locks on the first generated token. Language gate = PREFIX decode.

Mini-run (CPU, counting task, 80 steps)

decode	unique	echo	0 1 2 →
PREFIX (TF-matched)	15.7 GO	0.00	4 5 6 7 8 9 10…
CARRY (length-1)	1.7	0.00	4 4 4 4 4…

Do not crank REPEAT_COEF above 0.1 (CE explodes).

Tests

python -m pytest tests/test_p0_routing.py tests/test_ar_trainer.py -q
python scripts/probe_p0.py
python scripts/mini_v4_run.py

1B smoke (one GPU)

CUDA_VISIBLE_DEVICES=0 GPU_ID=0 START_TOKEN=<manifest> \
  BATCH=8 CE_CHUNK=2048 FRACTUS_ATTN_IMPL=chunked BLOCK_CKPT=1 \
  python -u scripts/fast4gpu_boost_v4.py

Read UNIQUE@40 PREFIX. That's the gate. CARRY is logged, not the blocker.

Push

See PUSH.md.

DiffusionBlocks (Fractus-native, 2026-08-27)

See DIFFUSIONBLOCKS_FRACTUS.md. One CTEBlock trained at a time, noise band on the residual, CE + local denoise, 0 new params. Mini:

python -m pytest tests/test_dblock.py -q
python scripts/mini_dblock_run.py