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

receipt_agent.npz is a real, honest 4-10-4 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/ReceiptAgent-Nano", "receipt_agent.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: see TRAINING_RECEIPT.json / BENCH.*.json (SYNTHETIC).

ReceiptAgent-Nano

ALLOW · DENY · ABSTAIN · ESCALATE. Escalation is a class, not a retry loop.

Family. nano · Evidence. SYNTHETIC · Weights. numpy · Params. 4-10-4

Hub: SZLHOLDINGS/ReceiptAgent-Nano

The cut

Anthropic refuses. NVIDIA rails. Unsloth trains. We add a fourth way: hand the decision to a human with a receipt. Loop-tax lives here.

A policy head that cannot silently succeed. Every output is one of four named gates.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	Constitutional refuse → typed DENY/ABSTAIN.
NVIDIA	NeMo Guardrails flow → four-way head.
Unsloth	Grown form is SZL-Forge-1.5B-ReceiptAgent.

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

Intended use

Fail-closed unit tests for the 4-way gate.

Bench (this tree)

TRAINING_RECEIPT.json seed 20260721 · honesty REPORTED · kernel is truth

Metric	Value
held-out agree vs rule_check	0.905
weights	receipt_agent.npz sha256 8aca4d24c90d6159cbb2bb885c7a94822d715899437f58abe69fb5c9664a1381

Infers on POST /api/infer {"kind":"receipt_agent"}. Surrogate may disagree. Kernel wins. Not 1.5B.

Limitations

• Synthetic 4-D features. Not a substitute for the 1.5B agent.
• Hub kernel labels are ALLOW/WARN/BLOCKED/ESCALATE — kernel is truth. This atelier MLP is a 4-class silhouette, not rule_check.

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.