CARD honesty (SZL cookie): This Hub entry is a card / doctrine / kernel / witness surface. It does not claim production neural weights unless a MEASURED train receipt is linked below. States used:
NOT_TRAINED·CARD_ONLY·MEASURED·UNAVAILABLE·NOT_CLAIMED. Production promote requires StephenGO WEIGHT <model-id>. Soft Shield: Control before action. Evidence after.
TinyKhipu-Nano
A small synthetic fixture for proposing NAVIGATE or ABSTAIN from token embeddings and candidate handles.
Artifact: Bare NumPy weight archive · Stage: Synthetic test fixture
Explore in Command Lab · Build · Evidence
Before you use it
- The forward-pass implementation lives in the canonical package; this archive has no packaged Hub loader or config.json.
- Historical synthetic results and unsigned receipt claims do not establish generalization, a refusal guarantee, or production readiness.
- This fixture is separate from the Khipu 1.5B model. Verify the complete inference contract before use.
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.
TinyKhipu-Nano
Synthetic token embeddings and candidate handles produce a proposed NAVIGATE or ABSTAIN output. This fixture establishes no refusal guarantee.
Family. nano · Evidence. SYNTHETIC · Weights. numpy · Architecture. 24-token, 12-dimensional embedding with two-class and handle-scoring heads · Not 1.5B.
Hub: SZLHOLDINGS/TinyKhipu-Nano
Reference fixture
A synthetic token-and-handle reference fixture for the NAVIGATE/ABSTAIN schema. It mean-pools token embeddings and scores handle notes. This card makes no ecosystem-wide novelty claim and establishes no predecessor relationship or quality proxy for a 1.5B checkpoint.
Intended use
Unit-test the NAVIGATE|ABSTAIN schema before GPU spend.
Reported synthetic fixture evidence
TRAINING_RECEIPT.json seed 20260721 · steps 280 · honesty REPORTED
| Metric | Value |
|---|---|
| plan_valid | 1.00 |
| abstain | 1.00 |
| hallucinated | 0 |
| weights | tiny_khipu.npz receipt-reported sha256 cc8d0385b2c75079669df809d7e4823f1ad8d9d535aec511446347490b11dff9 |
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
- The shared unsigned receipt reports
plan_valid=1.0,abstain=1.0, andhallucinated=0.0on its synthetic fixture. Named sample count, split identity, and generalization are not independently established by that receipt. - These reported values are not a reliability guarantee or a sibling model evaluation.
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
The previous card reports tiny_khipu.npz (3,568 bytes). Receipt-reported SHA-256 (not rehashed in this review):
cc8d0385b2c75079669df809d7e4823f1ad8d9d535aec511446347490b11dff9
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.
| Array | Shape | Data type |
|---|---|---|
E |
[24, 12] |
float64 |
W |
[2, 12] |
float64 |
b |
[2] |
float64 |
Wc |
[12] |
float64 |
The retained receipt labels these reported synthetic fixture results REPORTED.
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 d890010f874ecafd3f2f1ef4066d84b810ba775e.
Reviewed publisher source: hf/TinyKhipu-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["tiny_khipu.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.
Card authoring scope
This mirror's card is authored separately in Forge at tinykhipu-nano/card/README.md.
Its reviewed canonical card reference is szl-khipu@57fab7f8c3a8514d597b055cf0dfbfd3592caea5.
A card/banner correction does not publish the canonical package, independently
verify archive bytes, replay training, recover a historical class contract, or
qualify a hosted runtime. The banner is a decorative schematic; its node counts
do not specify the arrays reported above.