File size: 3,213 Bytes
4c14585
 
 
 
ce0e089
 
 
 
 
 
 
4c14585
 
ce0e089
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4c14585
 
24b2a6e
4c14585
24b2a6e
4c14585
24b2a6e
4c14585
20d8cce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3da416e
 
 
 
 
 
 
 
 
 
 
20d8cce
 
 
 
 
4c14585
 
24b2a6e
 
 
 
 
 
 
 
4c14585
24b2a6e
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
---
license: apache-2.0
library_name: numpy
tags:
- governed-ai
- szl-holdings
- doctrine-v11
- nano
- synthetic
- needs-loader
- test-fixture
---

> ### 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.
>
> ```python
> 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](https://huggingface.co/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](https://huggingface.co/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](https://orcid.org/0009-0001-0110-4173).