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

chakana

This repository contains a small NumPy reference artifact: evidence-class crossing silhouette: MEASURED/SIGNED/MODELED kept separate. The broader organ remains roadmap work. The synthetic archive is present; earlier statements that this repository had no weights are superseded.

Canonical source

Canonical source: https://github.com/szl-holdings/szl-forge (kit directory chakana/README.md). The card text itself is mirrored from szl-khipu/atelier/hf/chakana.md. GitHub is canonical; Hugging Face is the mirror. Re-synced 2026-09-25.

Weights on this Hub ID

• OBSERVED — chakana.npz (3,219 bytes) is on the Hub tree, together with TRAINING_RECEIPT.json, LICENSE, provenance.json, status.json and bom/model-bom.cdx.json (listed 2026-09-25 through the Hub file API; revision SHA not captured by this pass). It is a synthetic NumPy silhouette: a small two-layer MLP pack from a CPU/NumPy run, described by its own receipt as "synthetic silhouettes only — honest placeholders with real weights".
• UNAVAILABLE — production weights, any checkpoint of the intended organ, any loader or config.json. Intended base per the szl-forge kit: Qwen/Qwen3-Embedding-0.6B (no fine-tune of it is on this Hub ID, so base_model is deliberately not declared in the frontmatter).
• UNAVAILABLE — TRAINING_RECEIPT.batch.json, the "full 6-organ receipt" the receipt refers to, is not on this Hub ID.

What the szl-forge kit says: chakana/README.md and chakana/HF_MODEL_CARD.md describe the intended organ as bridge embeddings (sentence-transformers, base_model: Qwen/Qwen3-Embedding-0.6B, optional Matryoshka 256/512/1024), status ROADMAP, Jobs UNKNOWN, nDCG@10 UNKNOWN, publication_eligible: false, with a launcher that refuses to fire. Both files still say the Hub repository is a stub "(README + .gitattributes only). No weights." That sentence describes an earlier state: the synthetic archive below is now on the tree. It remains true that no encoder weights and no Qwen3-Embedding fine-tune are on this Hub ID.

Disambiguation (from the kit). a11oy CHAKANA wiring / tinkuy (a11oy/organs/amaru/**/chakana_wiring.py) is Andean-cross topology wiring, not this organ; MiniEmbed in szl-kernels is a different, smaller artifact.

Intended use

Inspect synthetic artifacts and exercise software fixtures. The archive does not include a packaged loader or config.json in this repository. from_pretrained compatibility and hosted inference are unverified.

Artifact evidence

chakana.npz is present (3,219 bytes). SHA-256:

3849772ee86fe09e8752e98250573564de66e814231b50c36e7d29b5b6c3dbd5

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

Re-sync note (2026-09-25): the file was re-listed at the stated path and size; the SHA-256 match and the array inspection above are as stated by the card's original check and were not recomputed in this pass (the Hub file API exposes size only, and the archive was not downloaded). The receipt's stated digest is 3849772ee86fe09e8752e98250573564de66e814231b50c36e7d29b5b6c3dbd5; the receipt is unsigned.

Array	Shape	Data type
w1	[12, 16]	float64
b1	[16]	float64
w2	[16, 3]	float64
b2	[3]	float64
holdoutAcc	[]	float64
seed	[]	int64

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.

Reported training

The repository receipt reports seed 20260721, 2000 steps, synthetic accuracy 0.588333, and loss 0.820603. These values were read from the receipt and were not independently rerun. They do not establish field performance or authority to make operational decisions.

Evidence boundaries

• Artifact bytes and receipt hash: checked at the pinned repository revision.
• Training and evaluation: reported synthetic evidence only.
• Production deployment, runtime correctness, and energy: unverified.
• Intended organ (see the szl-forge kit): ROADMAP. Jobs: UNKNOWN. No job id, no held-out evaluation.
• Proven trust: false in the published training receipt.
• Lambda uniqueness remains Conjecture 1 OPEN, not a theorem.

Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr. · ORCID 0009-0001-0110-4173.