SZL Holdings

AI that signs its work — and refuses to lie.

Every claim carries a cryptographic receipt you can verify yourself. Anything it can't prove is labeled UNAVAILABLE instead of a plausible-looking answer. No silent confidence. No fabricated "LIVE." Signed evidence, honest bounds.

See it in 30 seconds

Open a-11-oy.com — the living command fabric. Every governed action emits a hash-chained, DSSE-signed receipt. Pull one and verify it against the public contract.

Start here

1. The product — a-11-oy.com: the A11oy command substrate, live capability status, governed workflows.
2. The proof — a11oy.net: the proof registry: receipts, evaluations, and known bounds, including what we can't yet prove.
3. Install the stack — PyPI: 20 packages with OIDC trusted publishing and build provenance.
4. Browse the artifacts — Hugging Face / SZLHOLDINGS: models, kernels, datasets, and Spaces, each with evidence and stated limits.
5. Read the source — github.com/szl-holdings: canonical source, evaluation lanes, and per-repo controls.

The receipt stack on PyPI

Twenty packages that are the provenance tooling: receipt primitives, guardrail and energy witnesses, calibration and retrieval benchmarks, an OpenTelemetry evidence exporter. Each ships by OIDC trusted publishing (no stored tokens) with PEP 740 attestations and fails closed: MEASURED, BLOCKED, or UNAVAILABLE, never coerced.

pip install szl-receipts szl-guardrail-receipt vsp-otel szl-nemo

The doctrine

Source, running software, and model evaluations are separate claims, kept separate. A green test is not a deployment; a deployment is not a proof. Unknowns stay visible. Supply-chain security is converging on the same discipline — SLSA, in-toto, Sigstore — except here it isn't bolted on. It's the product.

Portfolio topology

One inference flagship: SZL Router is source-owned at szl-holdings/szl-router; its presentation target is SZLHOLDINGS/llm-router-live, product integration is https://a-11-oy.com/code, and proof originates at a11oy.net. Portfolio roles distinguish one inference flagship, three commercial flagships, five public domain bodies, and six internal engines. Those labels describe topology, not availability, operational readiness, or publication policy.

Current state (honest, dated)

Public Hugging Face estate observed 2026-09-30: 50 models, 35 datasets, 33 Spaces (repository counts). Twenty PyPI packages published. Dated observations from the live Hub API, not marketing numbers.

HISTORICAL: estate-alignment contract v1 recorded 16 portfolio Spaces, 1 inventory-only Space, 45 models, and 34 datasets — prior snapshot, superseded above. Its named topology: A11oy, Forge, Killinchu, Terra, PRISM Counsel, PURIQ Finance, LYTE.

Verification contracts: a11oy. Training and evaluation: szl-forge (kits, datasets, runbooks, measured limits). Trust boundary: TRUST.md. Killinchu effectors stay SIMULATED; SZLHOLDINGS/SZLHOLDINGS is HISTORICAL.

Governed AI. Evidence you can inspect. Work you can verify.