PowerMachine/HAKO-v1

HAKO — Hybrid Attention Kohonen Orchestrator

Orchestrated / routed / cooperative hybrid-attention system fusing GHSOM + CPNN + BKN + DASOM (Kohonen core), MoE cross-attention, Tree-GNN reasoning, chain recursion + beta fusion, a Diffusion game-loop generative layer, a cyclic thinking orchestrator, and Lyapunov-guarded auto-tuning (Robbins–Monro Θ hyper-gradient, EWC-QAC, CRB-gated growth, ABMO barrier, orthogonal ensemble, GP-PGO), built on frozen quantized sources: AMD Qwen2.5-0.5B int4 ONNX and AMD granite-4.0-1b AWQ grp32 ONNX (weights decomposed into int4_proto, int4_xi, expert keys K, and 4 source-grounded adapters per source, quantization preserved).

All mathematical statements (lemmas, theorems, propositions) and their proofs live in the module docstrings (hako/**.py); tests/test_math_properties.py verifies them numerically before any training or publication.

Layout

hako/
  config.py               hardware probe, budgets, storage caps
  memory_manager.py       aggressive RAM/disk watchdog (P-MEM)
  telemetry.py            5 JSONL streams (learning/routing/orchestration/diffusion/thinking)
  checkpoint.py           compressed state save/load
  tokenizer/byte_bpe.py   parallel byte-BPE, Lemma-1 exact sharded counts
  sources/loader.py       streaming HF fetch (token via env HF_TOKEN ONLY)
  sources/decompose.py    mandated geometry decomposition (P-DEC, T15)
  sources/runtime.py      frozen ONNX embedding runtime (T-EXH)
  core/kmeanspp.py        K-Means++ + Lloyd (T-KM)
  core/plateau.py         2D plateau density (T-PLATEAU, Morse)
  core/ghsom.py           growing hierarchical SOM (T-GHSOM)
  core/heads.py           CPNN/BKN/DASOM heads (T-HEADS)
  core/moe_attention.py   Mechanism I (T-ATTVAR corrected proof)
  core/tree_gnn.py        Mechanism II (T-DIRICHLET + Brouwer)
  core/chain.py           chain recursion + beta fusion (T-CHAIN)
  generative/diffusion.py DDPM + game curriculum (T-CURRICULUM)
  orchestrator/cyclic.py  4-step thinking loop (T-ORCH termination bound)
  orchestrator/router.py  REINFORCE router + Gumbel-Softmax (T-ROUTER)
  autotune/meta.py        Robbins–Monro Θ controller (T-RM)
  autotune/ewc_qac.py     Fisher clip/smooth + QAC cubic (T-QAC)
  autotune/growth.py      CRB growth gate + N-lift (T-GROWTH)
  autotune/abmo.py        attention meta-optimizer + log barrier (T-ABMO)
  autotune/ensemble.py    orthogonal ensemble + PID (T-ENS, T-PID)
  autotune/pgo.py         GP surrogate + EI + L-BFGS (T-GP)
  train/phase*.py         phase pipelines
  publish/push_hf.py      token-gated publication to PowerMachine
tests/test_math_properties.py   15 numerical proof checks
run_all.py               phase 0–4 runner (math suite gates everything)

Run

pip install -r requirements.txt
export HF_TOKEN=...        # used ONLY at runtime, never stored
python3 run_all.py         # full pipeline (~1h budget)
python3 run_all.py --smoke # short validation pass

Telemetry

telemetry/{learning,routing,orchestration,diffusion,thinking}.jsonl — append-only, crash-consistent JSONL (P-TEL).

Security

The HuggingFace token is read exclusively from HF_TOKEN at call time and is never persisted; publish aborts if the token string appears anywhere in the payload tree.