_{Part of the SZL Holdings governed estate — claims are designed to carry checkable receipts. Verification proves integrity & origin, never accuracy or performance.}

The cut

Leaders treat GGUF as the model. We print on the card: signed hash does not cover these files. CPU-honest, receipt-honest.

Edge navigation that still fails closed, with a card that refuses to launder a quant as a weight.

Silhouette → leave → SZL

Nobody else ships this combination. That is the point of a one-of-one.

Intended use

llama.cpp / Ollama / LM Studio. Still proposal-only.

Limitations

• Derived. Numerics drift vs BF16.
• Do not cite GGUF as the signed checkpoint.

Canonical GitHub: szl-holdings/szl-serve

| Base model | Qwen/Qwen2.5-1.5B-Instruct | | License | apache-2.0 | | Parameters | 1.5B | | Hardware | Runs CPU-only via GGUF Q4_K_M (~0.99 GB); GPU optional | | One command | ollama run hf.co/SZLHOLDINGS/SZL-Khipu-1.5B-GGUF:Q4_K_M |

GGUF quantizations of SZL-Khipu-1.5B — a QLoRA fine-tune of Qwen2.5-1.5B-Instruct for governed, grounded-only navigation of the SZL receipt lake. Small enough for a laptop, honest enough for an audit.

Published provenance: the base model ships with owner-signed training and eval receipts. They are Ed25519 signatures over canonical JSON and chain the evaluation receipt to the training receipt. This repo carries those receipts plus the repo-declared public key so you can verify repository-key continuity and receipt integrity before you load a tensor. They are not DSSE envelopes; the repository does not establish external key provenance, and an owner signature is not an independent evaluation.

Quants

Chat template (Qwen2.5 ChatML) is embedded in every file.

Run it

Ollama

ollama run hf.co/SZLHOLDINGS/SZL-Khipu-1.5B-GGUF:Q4_K_M

llama.cpp

llama-cli -hf SZLHOLDINGS/SZL-Khipu-1.5B-GGUF:Q4_K_M -p "Navigate: which receipt signed decision d-42?"

LM Studio — search SZLHOLDINGS/SZL-Khipu-1.5B-GGUF, pick Q4_K_M.

Prompt contract

The user turn is a single JSON object {query, candidates:[{nodeId, nodeKind, label, note}]} — handles only, never node content — and the model returns a single JSON plan (decision=NAVIGATE citing offered handles, or decision=ABSTAIN with an abstainReason) per khipu.schema.json. The full contract and expected output shape live on the BrainNavigator card.

Verify before you trust

# The owner-signed receipts travel with the weights:
#   training_receipt.signed.json; eval_receipt.signed.json; owner_pubkey.json
# They are Ed25519 signatures over canonical JSON, not DSSE envelopes.
# Verify them offline against the repo-declared public key before use.

Quantization: llama.cpp convert_hf_to_gguf.py -> llama-quantize (F16 -> Q4_K_M / Q5_K_M / Q8_0), 2026-07-15. Quantization changes numerics; the signed evaluation receipt covers the pre-quantized BrainNavigator evaluation artifact described by that receipt, not any GGUF. The LFS hashes above bind the exact uploaded GGUF bytes. No post-quantization quality evaluation or independent benchmark is claimed.

---

---