Stephen P. Lutar Jr.

AI/ML systems architect · Founder & CEO, SZL Holdings

---

I build the accountability layer that consequential AI is missing.

When an AI system makes a decision that matters — in defense, finance, critical operations — you should be able to verify exactly what it decided, under what authority, on what evidence, and that the record was never quietly changed. Today, you mostly can't. I set out to fix that.

SZL Holdings is the company I founded to build this missing layer: a governed-inference substrate where every AI action produces a cryptographically signed, tamper-evident receipt, and the trust math behind it is pinned in Lean 4 and checked by a proof machine — not just claimed on a slide.

The result is a11oy: a governed-AI Command Center with deny-by-default safety gates, trust scoring, and signed receipts for every decision. We apply the same substrate to autonomous systems in the field: live tracking, multi-sensor fusion, and governance receipts for counter-UAS and maritime operations.

---

What I've built

Artifact	What it is
a11oy	Governed-AI Command Center. Deny-by-default gates, trust scoring, signed receipts. Live at a-11-oy.com.
lutar-lean	Lean 4 proof library. 8 formulas locked-proven in the kernel, ~185 machine-checked theorems. DOI: 10.5281/zenodo.20434308.
Counter-UAS governance	Live track board, multi-sensor fusion, engagement rules, and signed engagement receipts for air and maritime autonomy.
Sovereign on-metal AI	Full governed inference stack — deployable air-gapped on your own hardware, signed and verifiable end to end.

---

The math, briefly

Every governed decision is scored by a function Λ — a weighted geometric mean over four trust axes. Its properties are formally stated and machine-checked:

• 8 formulas are locked-proven at kernel commit c7c0ba17: receipt replay determinism, DAG acyclicity, FIFO ordering, ledger conservation, bounded coupling, Reed–Solomon recovery, entropy budget, append-only monotonicity.
• ~185 theorems machine-checked across Waves 11–23.
• Λ unconditional uniqueness is Conjecture 1 — I found a concrete counterexample and publish it openly. The conditional version (given separability) is proven axiom-free (Theorem U). I call the open problem by its honest name and offer a bounty to close it.

The discipline of saying "this is proven, this is conjecture, this is wrong" is not modesty. It is the strongest claim I can make to an auditor.

Thesis DOI: 10.5281/zenodo.19944926

---

Verify it yourself

# Verify a build attestation on the command platform image
gh attestation verify oci://ghcr.io/szl-holdings/a11oy:latest --repo szl-holdings/a11oy

# Verify a governance receipt from the field node (offline, trust no one)
curl -s https://szlholdings-killinchu.hf.space/cosign.pub -o cosign.pub
curl -s https://szlholdings-killinchu.hf.space/api/killinchu/v1/receipt/export > receipt.json
# tamper one byte → "Verification failure"

---

Stack

Languages: TypeScript · Python · Lean 4
Runtime: Node.js · React · FastAPI
Proof: Lean 4 + Mathlib
Supply chain: Sigstore / cosign · Zenodo · CodeQL · Trivy · OpenSSF Scorecard
Deployment: Kubernetes · Zarf · air-gap bundles

---

Find me

Company	github.com/szl-holdings
Command platform	a-11-oy.com
Hugging Face	huggingface.co/SZLHOLDINGS
Thesis DOI	10.5281/zenodo.19944926
ORCID	0009-0001-0110-4173
Email	[email protected]

---

_{© 2026 Stephen P. Lutar Jr. · Code: Apache-2.0 · Research: CC BY 4.0 · Locked-proven = exactly 8 formulas at kernel c7c0ba17 · Λ = Conjecture 1 (unconditional = machine-checked false; conditional = proven axiom-free) · SLSA L1 honest · L2 build-attested · L3 roadmap · No ATO claimed}