# A signal engine with no training, no tokens, no neural net URL

> Source: <https://github.com/JJardine919/voodoo-aoi>
> Published: 2026-07-08 13:57:25+00:00

A topological "collapse" engine that reads structure in signals with **no training,
no tokens, and no neural network.** It runs offline on a laptop — if you don't have a
quantum sampler it falls back to local simulated annealing automatically, no account
or key required. Clone it, run it, get the numbers below.

Built over five months by one person and an AI, and given to the world under copyleft. Institutions optimize for what they can measure; a lot of value falls through the cracks they can't see. This is a small tool for leveling that field.

These demos are **training-free topological analyses** (persistent homology). No model
is fit to any data. The same mathematical tool is applied across three unrelated
domains. **We show you where it fails, too — that's the point.**

As MRI k-space is undersampled, image loop-topology (Betti-1) degrades monotonically — no reference image, no training. Replicated across 6 brains (OpenNeuro ds000102):

| acceleration | 1× | 2× | 2.9× | 4× | 6× | 8× |
|---|---|---|---|---|---|---|
| Betti-1 (mean) | 112 | 94 | 71 | 56 | 34 | 23 |

`python demos/mri_qc.py`

On the two most look-alike glitch classes, Blip vs Koi Fish, where peak amplitude
barely separates them (AUC 0.24), topology does (**AUC 0.84**).
**Honest limit:** on the easy, already-separable pairs a trivial spectral-bandwidth
feature beats topology. This is a *narrow* result, not a universal one.

`python demos/ligo_glitches.py`

Topology tracks capacity fade strongly (r ≈ −0.7 to −0.9 across cells).
**Honest limit:** it does **not** forecast *future* capacity beyond what you'd get from
current capacity + recent trend (partial r ≈ 0.04). Descriptive positive, predictive null.

`python demos/battery_predict.py`

`engine/`

is the full 96D octonion-collapse organism: entropy gating, Fano/octonion
projection, Jordan-Shadow decomposition, 33 transposable-element families,
Monster-moonshine grading, and chain-complex homology. Token-free, no LLM, runs offline.

``` python
import numpy as np
from aoi_collapse_96d_dwave import aoi_collapse_96d_dwave

out = aoi_collapse_96d_dwave(np.random.normal(size=96))
print(out["betti"], out["intent"], out["chaos"], out["backend"])
# -> [..] 0.xx 0.xx SimulatedAnnealing
pip install -r requirements.txt
bash fetch_data.sh   # downloads the public datasets into ./data/ (MRI, LIGO, battery)
```

Then run any demo from the repo root, e.g. `python demos/mri_qc.py`

.

**AGPLv3** (see `LICENSE`

). You may use, modify, and redistribute freely; if you run a
modified version as a network service, you must release your source. Every copy carries
attribution (see `NOTICE`

).
Commercial licensing (to use it without AGPL obligations): ** james@lattice24.com**.

James Jardine, *AOI Shell v1.1: Token-Free Domain-Agnostic Signal Organism via 96D
Octonion Collapse.* DOI [10.5281/zenodo.20200607](https://doi.org/10.5281/zenodo.20200607)

**James Jardine** — Lattice24 / VoodooAOI · [james@lattice24.com](mailto:james@lattice24.com)
Built in collaboration with Claude (Anthropic).