Status: SOFTWARE / REFERENCE / TEST FIXTURE. Not a production model.

This Hub repository contains a bare NumPy archive. The loading and forward-pass implementation lives in the canonical szl_khipu package; no packaged Hub loader or config.json is shipped alongside these weights. Treat this as a software fixture until its complete inference contract is independently verified.

ReceiptAgent-Nano

ALLOW · WARN · BLOCKED · ESCALATE. Escalation is a class, not a retry loop.

Family. nano · Evidence. SYNTHETIC · Weights. numpy · Architecture. 24-16-8-4 MLP

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 24-dimensional 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.

Artifact evidence

receipt_agent.npz is present (6,014 bytes). SHA-256:

8aca4d24c90d6159cbb2bb885c7a94822d715899437f58abe69fb5c9664a1381

The archive hash matches TRAINING_RECEIPT.json. Its arrays were inspected with numpy.load(..., allow_pickle=False); numeric values were finite.

Array	Shape	Data type
W1	[16, 24]	float64
b1	[16]	float64
W2	[8, 16]	float64
b2	[8]	float64
W3	[4, 8]	float64
b3	[4]	float64

A matching unsigned receipt establishes local artifact consistency. Training metrics remain reported synthetic results; this check does not independently reproduce training or establish deployment, general intelligence, or production readiness.