How to actually load this — the weights alone are not enough

moons.npz is a real, honest 2-8-2 MLP produced by a real training run, and the card below does not overstate it. But it is a bare NumPy archive with no config.json and no loader in this repo, so from_pretrained and the Hub inference widget cannot touch it. Nothing here tells you the array names or the forward pass.

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download("SZLHOLDINGS/Moons-Nano", "moons.npz")
w = np.load(path)
print(sorted(w.files))   # array names are the de-facto interface

The forward pass this was trained against lives in the szl_khipu package, in SZLHOLDINGS/szl-khipu-kernels — a different repository. Until the loader ships alongside the weights (or a custom_code handler is added), treat this repo as a test fixture, not a deployable model. Evidence status: acc 0.93 / loss 0.13 REPORTED on the TRAIN set.

Moons-Nano

Two-moons 2→8→2 tanh-softmax SGD. A few hundred floats. Not 1.5B. Not Qwen. Not a foundation model.

Canonical source: szl-holdings/szl-khipu
Sibling card: SZLHOLDINGS/szl-khipu

from szl_khipu.train import moons

weights, ev = moons.train(seed=20260721, steps=400)
print(ev["acc"], ev["loss"])
# REPORTED: acc 0.93 · loss ~0.13 on the training moons
moons.save_npz("moons.npz", weights)

What it does

• Classic two-moons toy classification. Hidden width 8. Softmax over 2.
• Trained here on CPU NumPy. Honesty REPORTED. Energy UNAVAILABLE.

Bench (this tree)

TRAINING_RECEIPT.json seed 20260721 · steps 400 · honesty REPORTED

Metric	Value
acc	0.93
loss	~0.13
weights	moons.npz sha256 dda50e3b293534de3f5aec01ebf9f8d6688e06069931618dfd35f01369904104

Infers on POST /api/infer {"kind":"moons","x":0.2,"y":0.3}. Not 1.5B. Not a published benchmark.

What it is NOT

• Not SZL-Khipu-1.5B. Not QLoRA. Not a chat model.
• Not sklearn moons as a product claim. A live silhouette so the estate has a TRAINED tiny MLP that actually ran.
• Not proven trust. Λ uniqueness remains Conjecture 1 OPEN.
• Energy UNAVAILABLE. CUDA UNAVAILABLE. Never a fabricated joule.

Honesty

Claim	Label	What-NOT
Weights trained in this package	REPORTED	silhouette, Not 1.5B
acc 0.93 on the training moons	REPORTED	not a published benchmark
Λ	ADVISORY · Conjecture 1 OPEN	never a theorem
Energy	UNAVAILABLE	never a fabricated joule
CUDA	UNAVAILABLE	CPU numpy LIVE

Doctrine v11 LOCKED · 749/14/163 · locked-proven 8. Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr. · ORCID 0009-0001-0110-4173.