SZL Khipu
Inspect advisory NumPy kernels, small reference fixtures, and hash-chained run receipts.
Artifact: Python / NumPy software package · Stage: Software / reference
Explore in Command Lab · Build · Evidence
Before you use it
- Λ is advisory: Conjecture 1 OPEN and proven_trust=false.
- CUDA and energy measurement are UNAVAILABLE; software source binding does not establish a production deployment.
- The package and Nano fixtures do not qualify a 1.5B checkpoint. Verify each artifact and its separate inference contract.
Technical details and evidence
szl-khipu
Control before action. Evidence after.
Part of the szl-holdings estate · Product: a-11-oy.com · Proof: a11oy.net
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
The canonical publish-hf workflow mirrors szl_khipu/, README.md, LICENSE,
and pyproject.toml from immutable Git blobs in one Hub model commit. Its
MODEL_SOURCE_BINDING.json records the Git commit, tree, and each file digest;
publication succeeds only after immutable provider readback matches every byte.
Historical Hub files outside this software scope are preserved and are not
attested by this binding. Software parity makes no trained-checkpoint claim.
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: the reviewed
szl_khipu/train/receipt_agent.pydefines0=ALLOW,1=WARN,2=BLOCKED,3=ESCALATE. The deterministicrule_checkdecision remains authoritative; learned predictions do not override it. This software binding does not attest the sibling archive, its card, or its historical training/class-index contract. - 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 Sagecsrc, not vLLM paged.cu, notflex_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 6 demo; gradio>=6.27.0,<7
pip install -e ".[torch]" # optional — torch path; CUDA still UNAVAILABLE here
Requires Python ≥ 3.11. Runtime dep: numpy>=1.26. Apache-2.0.
Python frontend + backend
Stdlib HTTP. Same process serves the holographic UI and runs the kernels. No Gradio required.
szl-khipu serve --host 0.0.0.0 --port 7860
# GET /healthz /version /api/lambda /api/greenlight /api/anatomy
# POST /api/lambda /api/anatomy /api/greenlight /api/yarqa /api/tiledigest
Gold = OPEN. Proof teal = LIVE. Never green-as-proven. Energy UNAVAILABLE. Conjecture 1 stays OPEN.
Static JS silhouette (not the NumPy kernel): docs/index.html · szl-holdings/khipu-pages · htmlpreview.
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
For numerical output without dense probability diagnostics, use
yarqa_attn_output(q, k, v, n_canals=3). This NumPy path computes query tiles
inside each canal and avoids full sequence-square score/probability matrices.
The API contract and reproducible CPU comparison
document validation, memory limits, attribution, and measurement boundaries
at the reviewed source commit. This document is outside the model software
mirror scope and is not included in that Hub model snapshot.
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 |
| space/ | Hub docker hologram — stdlib HTTP, no Gradio |
| spaces/ | Gradio 6 demo (GitHub only; separate Docker publisher uses space/) |
License
Apache-2.0. Copyright 2026 SZL Holdings.
Stephen P. Lutar Jr. / SZL Holdings · ORCID 0009-0001-0110-4173
Work with SZL Holdings
szl-khipu participates in the SZL governed-AI estate. The flagship control plane targets governed agent change management, but production/runtime qualification is tracked separately through the GitHub → Hugging Face → product → proof evidence chain. This repository does not itself prove or authorize a production deployment.
- Design partners (6-month, paid): governed-action receipts in your environment → stephenlutar2@gmail.com
- Verify our claims offline: github.com/szl-holdings/szl-gov — signed estate receipt + public verifier
- Pricing + SKUs: not maintained in this repository.