szl-receipt-attn

KANCHAY · Doctrine v11 · Lean 749/14/163 · Λ = Conjecture 1 (advisory) · a-11-oy.com

Kernel sources are on this repo. CPU get_kernel import-LIVE is MEASURED. GPU UNAVAILABLE (no cubin + timed run). Not a model. Not listed next to Chaski or Qantu.

Status

The cut

Guardrails after generation are late. Masking before softmax is the thing nobody ships because it hurts scores. We want the hurt.

A model that cannot attend to what it is not allowed to see.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	Constitution applied at the attention head.
NVIDIA	Fused kernel, NVIDIA-shaped, SZL-cut.
Unsloth	No.

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

Intended use

Fuse into governed decode.

Limitations

• Kernel, not weights.

Canonical GitHub: szl-holdings/szl-khipu

STATUS: import-LIVE on CPU Kernel Hub get_kernel (kernels 0.16.1). GPU UNAVAILABLE.

Thing	Label	Method / N / date / what-NOT
Kernel Hub get_kernel	import-LIVE	MEASURED 2026-08-28 2:29pm ET on kernels 0.16.1. HEAD 42d9a31 (42d9a31c5ca1749fca16017d73880cbc4c5050fc). Legal name szl-receipt-attn (Python module szl_receipt_attn). Variants: build/torch-universal (default get_kernel) and build/torch-cpu (backend="cpu"). Working calls: get_kernel("SZLHOLDINGS/szl-receipt-attn", revision="main", trust_remote_code=True) and the same with backend="cpu". selfcheck ok (max_abs_vs_sdpa=0.0, path=torch_reference, chain_ok=true). What-NOT: no tokens/s; no joules.
GPU / Triton	UNAVAILABLE	No cubin + timed GPU run this session. Not claimed LIVE. No tokens/s, no joules.

Canonical source: https://github.com/szl-holdings/szl-receipt-attn

This Hub repo is the publish mirror. ATELIER owns cards. No CUDA benches. Λ = Conjecture 1. Apache-2.0.

import torch
from szl_receipt_attn import receipt_attn, ReceiptChain, selfcheck

q = k = v = torch.randn(1, 2, 16, 32)
chain = ReceiptChain()
y = receipt_attn(q, k, v, causal=True, chain=chain)
print(chain.verify(), selfcheck())