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

tiny_khipu.npz is a real, honest 4-6-2 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/TinyKhipu-Nano", "tiny_khipu.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: plan_valid 1.00 / abstain 1.00 (SYNTHETIC).

TinyKhipu-Nano

Four features in. NAVIGATE or ABSTAIN out. Abstain is the default class, not a post-hoc filter.

Family. nano · Evidence. SYNTHETIC · Weights. numpy · Params. 4-6-2 · Not 1.5B.

Hub: SZLHOLDINGS/TinyKhipu-Nano

The cut

Leaders train models to answer. We train a silhouette to shut up when overlap is thin or the lure is adversarial. This nano is the 4-6-2 silhouette of that cut. Not 1.5B.

A navigator that has never seen document text — only handles — and still knows when to refuse the walk.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	Refusal as a typed output, not a polite paragraph.
NVIDIA	Guardrail inside the head, not a sidecar.
Unsloth	The 1.5B QLoRA is the grown form of this MLP.

Nobody else ships this combination. That is the point of a one-of-one.

Intended use

Unit-test the NAVIGATE|ABSTAIN schema before GPU spend.

Bench (this tree)

TRAINING_RECEIPT.json seed 20260721 · steps 280 · honesty REPORTED

Metric	Value
plan_valid	1.00
abstain	1.00
hallucinated	0
weights	tiny_khipu.npz sha256 cc8d0385b2c75079669df809d7e4823f1ad8d9d535aec511446347490b11dff9

Infers on POST /api/infer {"kind":"tiny_khipu"}. Hard ID filter. Not Qwen. Not 1.5B.

Limitations

• Synthetic features. Perfect holdout is a design fact, not a field claim.
• The 1.5B abstain rate is 2/6 — this nano does not wash that.

Honesty

Claim	Label
This card's numbers	SYNTHETIC
Energy / joules	UNAVAILABLE unless a signed meter says MEASURED
Λ uniqueness	Conjecture 1 OPEN — not a theorem
GGUF as the signed object	FALSE

Doctrine v11 LOCKED · 749 declarations · 14 axioms · 163 sorries · locked-proven 8.

Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr. · ORCID 0009-0001-0110-4173.