ReceiptAgent-Nano

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

Synthetic four-class policy surrogate. Escalation is a class in the reviewed implementation; this fixture does not establish a runtime retry loop.

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

Hub: SZLHOLDINGS/ReceiptAgent-Nano

Reference fixture

A small synthetic four-class reference fixture for testing policy-output schemas. Its learned predictions are advisory; they do not override the deterministic rule checker. No comparative safety or ecosystem-wide novelty claim is made.

Intended use

Test advisory four-class output schemas alongside the separately authoritative rule checker.

Reported synthetic fixture evidence

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

The related demo documents an application-specific POST /api/infer route. This archive repository establishes no hosted endpoint, served revision, or deployment guarantee. The route is illustrative application context.

Limitations

• Synthetic 24-dimensional features. Not a substitute for the 1.5B agent.
• The reviewed implementation defines class indices 0=ALLOW, 1=WARN, 2=BLOCKED, 3=ESCALATE, and keeps the deterministic rule_check decision authoritative. The former package summary named a different label set; that set must not be substituted here.
• The shared receipt does not bind an exact training/source revision, class mapping, held-out sample count, or split digest to this archive. Current-source inspection is not a recovered historical training contract. Agreement 0.905 is a reported synthetic result, not independently auditable production accuracy.

Honesty

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

The previous card reports receipt_agent.npz (6,014 bytes). Receipt-reported SHA-256 (not rehashed in this review):

8aca4d24c90d6159cbb2bb885c7a94822d715899437f58abe69fb5c9664a1381

The previous card reported that the archive matched the unsigned training receipt and that numpy.load(..., allow_pickle=False) found finite numeric arrays. The table below preserves that historical report. The September 30, 2026 card review read pinned text and the receipt; it did not download, rehash, or inspect the archive, and did not replay training.

An independently checked archive/receipt match could establish local artifact consistency; an unsigned digest would still not authenticate authorship or measurement. These preserved receipt and array reports establish no new training replay, independent evaluation, deployment, or production readiness.

Reviewed Hub text: immutable snapshot 463df555fee339b776b938846238327100f573ab. Reviewed publisher source: hf/ReceiptAgent-Nano/README.md at e53e3d24b22e356eb986c373aee27b3b3e7947ec. The shared unsigned TRAINING_RECEIPT.json, timestamped 2026-08-29T17:11:32.518042+00:00, enumerates four artifacts. Only its artifacts["receipt_agent.npz"] entry describes this archive; the other entries do not establish that sibling artifacts are present in this repository. The receipt does not bind its training run to the reviewed source commit.