szl-khipu

Knot the run. Hash the proof. Fail closed.

Governed-AI khipu: an advisory Λ-gate, original YARQA canal attention, a TinyKhipu NAVIGATE/ABSTAIN silhouette, and hash-chained receipts. CPU numpy is LIVE. CUDA is UNAVAILABLE. energy UNAVAILABLE. proven_trust is always false.

Canonical GitHub source: szl-holdings/szl-khipu
Hub card: SZLHOLDINGS/szl-khipu
Kernels: SZLHOLDINGS/szl-khipu-kernels

What it is

• A Python package (szl_khipu) that knots a run, hashes a receipt, and fail-closes on a zero axis or a broken chain.
• Λ: weighted geometric mean over 13 Yuyay axes, advisory, never proven trust.
• YARQA: contiguous canals; attend only inside the compartment; leak is the bound.
• TinyKhipu-Nano: NAVIGATE / ABSTAIN silhouette. A few thousand floats. Hard ID filter.
• ReceiptAgent-Nano: 4-way gate (HARD_DENY / DENY_DEFAULT / LAMBDA_VETO / ALLOW). The kernel is truth. The agent does not override it.
• Receipts: sha256 of weights, seed, steps, loss. Honesty REPORTED. Joules null.
• Organ integrity: five-organ fail-closed kernel of szl-holdings/anatomy. HEART/YUYAY, YAWAR, YACHAY, OTel, Khipu skeleton. Not a Three.js rehost. The 3D atlas is SLSA L1 static viz.

What it is NOT

• Not Qwen. Not a Qwen fine-tune, not a Qwen rehost, not a Qwen eval.
• Not 1.5B. Nothing in this tree trains or publishes SZL-Khipu-1.5B, KHIPU-R2, or ReceiptAgent 1.5B. Those Hub cards stay RESEARCH and are not produced here.
• Not a FlashAttention rehost. Not Dao .cu, not hopper/, not cute/, not Sage csrc, not vLLM paged .cu, not flex_attention.py.
• Λ uniqueness is OPEN. Conjecture 1 is not a theorem. Unconditional uniqueness under kernel A1–A5 is machine-checked FALSE.
• Energy UNAVAILABLE. Never a fabricated joule. MEASURED-NVML or honest null.
• Not trained weights at 1–2B. Not GGUF. Not a CUDA speedup. Not proven trust.

Install

pip install -e .
pip install -e ".[gradio]"   # optional — Gradio 4 space
pip install -e ".[torch]"    # optional — torch path; CUDA still UNAVAILABLE here

Requires Python ≥ 3.11. Runtime dep: numpy>=1.26. Apache-2.0.

Quickstart

Lambda (advisory)

from szl_khipu import evaluate_lambda, YUYAY_FLOORS

ev = evaluate_lambda(list(YUYAY_FLOORS))
print(ev.value, ev.blocked, ev.reason)
# advisory pass — uniqueness remains Conjecture 1 OPEN

zeroed = list(YUYAY_FLOORS)
zeroed[0] = 0.0
print(evaluate_lambda(zeroed).blocked)  # True — fail closed

CLI:

szl-khipu demo-lambda
szl-khipu demo-lambda --zero 0

YARQA (canal leak)

import numpy as np
from szl_khipu import yarqa_attn

q = k = v = np.random.default_rng(7).standard_normal((12, 4))
out = yarqa_attn(q, k, v, n_canals=3)
print(out.leaked)  # bound: leaked ≤ 1e-9

szl-khipu demo-yarqa --n-canals 3

Organ integrity (five organs)

Fail-closed substrate. Any DOWN organ or a WILLAY veto blocks the body. Λ stays Conjecture 1 OPEN. Energy UNAVAILABLE. Locked-proven stays 8. Not a Three.js rehost.

from szl_khipu import evaluate_anatomy

ev = evaluate_anatomy(seed=11)
print(ev.live_count, ev.blocked, ev.energy)  # 5 False UNAVAILABLE

print(evaluate_anatomy(zero_heart=True).blocked)        # True
print(evaluate_anatomy(fabricate_joule=True).energy_j)  # None — never a fabricated joule

szl-khipu demo-anatomy
szl-khipu demo-anatomy --tamper-chain
szl-khipu demo-anatomy --fabricate-joule

Canonical anatomy repo: szl-holdings/anatomy · 3D atlas: SZLHOLDINGS/anatomy

Train TinyKhipu

szl-khipu train --seed 20260721 --steps 280 --out artifacts
# writes artifacts/tiny_khipu.npz + artifacts/training_receipt.json
szl-khipu verify artifacts/training_receipt.json

from szl_khipu import train_tiny_khipu

result = train_tiny_khipu(seed=20260721, steps=280)
print(result.plan_valid, result.abstain, result.hallucinated)

Export all three silhouettes (TinyKhipu, ReceiptAgent, moons) and a combined receipt:

python scripts/train_and_export.py
# artifacts/*.npz + artifacts/TRAINING_RECEIPT.json
# honesty REPORTED · energy_j null · proven_trust false · whatNot listed

Kernels vs field leaders (honest deltas)

SZL cut	Field leader	Honest delta
TileReceipt (szl-receipt-attn)	FlashAttention (Dao et al., NeurIPS 2022)	Numeric silhouette of tiled fused attention. SHA receipt of the tiles. No speedup claim. Not a rehost of Dao .cu / hopper / cute.
BlockWitness (szl-block-kv)	PagedAttention / vLLM (Kwon et al., SOSP 2023)	Paged-KV gather with a digest of the block table. No tokens/s. Not a vLLM rehost.
ScoreMod Fiber (szl-maskmod)	FlexAttention (PyTorch, arXiv:2412.05496)	Original score_mod + block-mask path that receipts the mask digest. No CUDA benches. A swapped mask cannot stay silent.
YARQA-ATTN	SageAttention (Zhang et al., ICLR 2025)	Not quantized attention. Contiguous canals (Quechua yarqa), attend inside the compartment, receipt the partition. GPU cubins UNAVAILABLE. CPU numpy path LIVE.
Governed RMSNorm	RMSNorm (Zhang & Sennrich, 2019) / LayerNorm	Matches a reference and optionally chains a digest of the rounded output. No speedup. Digest is integrity, not authorship.
Λ-gate	UN HDI 2010 WGM / OECD composite indicators	WGM over [0,1] axes, fail-closed on any zero or non-finite axis, A1–A4 runtime checks. Uniqueness is Conjecture 1 — OPEN. Advisory, never proven trust.
govsign + provctl	in-toto / SLSA / Sigstore / DSSE	Governance predicate (Λ, energy, decision) as an envelope. proven_trust locked false. UNSIGNED is a first-class fallback, never a fake signature. SLSA L1 honest, L2 roadmap.

This repo silhouettes Λ, YARQA, TinyKhipu, ReceiptAgent, and receipts in numpy. It does not replace those field leaders and does not inherit their benches.

Locked-8

Lean locked-proven formulas = exactly 8. Lab numerics CHECKED ≠ Lean PROVEN. The F-number mapping onto any larger registry is UNKNOWN — never fabricated.

ID	Name	Caveat
F1	Replay-Hash Determinism
F4	Khipu DAG acyclicity	edges dst < src
F7	Chaski FIFO	drain(enqueueAll([], msgs)) = msgs
F11	Ayni Reciprocity Conservation	Σ in = Σ out
F12	Kuramoto boundedness	additive fragment ONLY — not full nonlinear sync
F18	RS(10,6) recovery	recoverable iff ≥ 6 of 10 shards
F19	Bekenstein additive scaffolding	monotone fragment ONLY — not S ≤ 2πkRE/ℏc
F22	Khipu emit monotonicity	sequence numbers strictly increase

Conjecture 1

Any two aggregators satisfying A1–A4 agree on every input. OPEN (sorry). Unconditional uniqueness under kernel A1–A5 is machine-checked FALSE.

Λ is advisory. Trust ceiling 0.97. proven_trust is false in this artifact and every receipt it emits.

Axioms carried at runtime (checks, not a uniqueness proof):

ID	Name
A1	IsMonotone
A2	IsHomogeneous Λ(c·x)=c·Λ(x)
A3	IsEgyptianExact Λ(c,…,c)=c
A4	IsBounded Λ(x)≤max(x)
A5	IsPermutationInvariant

Honesty

Claim	Label
CPU numpy path	LIVE
CUDA / GPU cubins	UNAVAILABLE
Energy / joules	UNAVAILABLE (never fabricated)
Λ uniqueness	Conjecture 1 OPEN
proven_trust	false
TinyKhipu / ReceiptAgent / moons trained here	REPORTED silhouette, not 1.5B
SZL-Khipu-1.5B / ReceiptAgent 1.5B	RESEARCH, not trained in this tree
Weights at Hub 1–2B	not applicable

Doctrine v11 LOCKED · 749 declarations · 14 axioms · 163 tracked sorries · locked-proven 8.

Hub siblings

Artifact	What
hf/TinyKhipu-Nano	NAVIGATE/ABSTAIN silhouette
hf/ReceiptAgent-Nano	4-way gate; kernel is truth
hf/szl-khipu-kernels	Kernel Hub card; get_kernel
spaces/	Gradio 4 space

License

Apache-2.0. Copyright 2026 SZL Holdings.

Stephen P. Lutar Jr. / SZL Holdings · ORCID 0009-0001-0110-4173