a-11-oy.com · operator workspace

Your work. One centre.

Governed conversation, existing tools, and the evidence behind them.

Grok connection

Not observed yet

Configuration is not successful inference.

Local open-weight model

Not observed yet

No provider key; this lane is available only from a configured local runtime.

Public CPU demonstration

Not observed yet

Separate SZL Khipu lab; public data only and upstream execution records are unsigned.

HF inventory

Not observed yet

Public snapshot, not qualification.

01 · Work

Think with Atelier.

External provider · bounded single turn

Grok 4.7 supplies text. SZL controls authentication, policy checks and signed receipts. No tools run here. Atelier's receipt ledger keeps metadata and content-derived hashes, not prompt or answer text; hashes are not anonymization. store: false disables xAI's stateful history; xAI's separate API audit retention is 30 days by default unless the team has Zero Data Retention.

Held only in this page’s memory; never saved to browser storage. The provider key stays on the server.

No request has been sent. Restricted and secret data cannot leave for Grok.

No automatic retries. A timeout can leave provider execution uncertain; check the outcome before sending another request. Encrypted provider reasoning is not displayed or replayed. Energy is UNAVAILABLE without a measured exporter delta.

01 · Work · local

Think on this computer.

Open weights · no provider API key

A separately configured loopback runtime can use an installed Ollama model without an xAI key. This is not Grok. The public website cannot reach your workstation; this form stays disabled unless its own server reports local configuration ready. The operator credential, policy gate, persistent signer and receipt ledger are still required. Model text is advisory and runs no tools.

Held only in this page’s memory. Never enter a model-provider key here.

No local request has been sent.

No automatic retries. Local output is not proof of model quality, factual accuracy or action authority. Configuration readiness is not a successful inference; a completed, signed turn is separate evidence.

01 · Work · public demonstration

Try the SZL CPU lab.

No provider API key · 32-token ceiling

This is a separate, best-effort SZL Khipu 1.5B demonstration, not Grok and not the local Ollama model. Send public, non-sensitive text only; do not include personal, regulated or secret data. Atelier still requires its operator credential and policy gate, and signs its own decision and observed outcome. The lab's execution record is UNSIGNED, and a signed Atelier receipt does not prove the lab's answer is correct or its remote execution independently attested. The lab permits one concurrent request and may be unavailable.

Held only in this page's memory. Never send a provider API key to the public CPU lab.

No CPU lab request has been sent.

One request, no automatic retry or tool execution. A timeout can leave the lab outcome uncertain. The upstream lab is a best-effort public demo without an SLA; inspect its independent source and limits.

02 · Operate

Existing tools, connected.

Navigation is not action authorization

03 · Verify

Follow the source.

  1. Protected GitHub Python source
  2. Canonical Hugging Face publication
  3. Product and proof verification

The source reported by this process must be checked against protected GitHub main and the publisher’s byte-attestation receipt. A running Space or HTTP 200 does not prove inference.

04 · Artifacts

The SZL estate.

Existing models, kernels, Spaces and datasets keep their own source bindings. This integration creates no SZL-owned Grok weights or training dataset.

05 · Study

Study widely. Build independently.

Research intake · not deployment admission

Research snapshot not observed yet.

Popularity is not quality. Provider listings are not successful inference. Declared card licenses still need version-specific review. We study cited public work, implement our own interfaces and mathematics, and require held-out evidence before claiming an improvement. Muse identity remains unconfirmed; Jev is advisory, never permission to act.

Inspect the full pinned intake and evaluation protocol ↗

2026-10-02 · no-key open-weight shortlist

Qwen3.5-4B · 8 GB study targetApache-2.0 card · exact Hub revision · not downloaded or qualified hereQwen3.5-9B · 16 GB study targetApache-2.0 card · exact Hub revision · not downloaded or qualified here

The installed local qwen3:4b-instruct is a separate Ollama artifact. We do not claim it is byte-derived from either Hub checkpoint, and local inference alone does not prove superiority.