Every model has an owner.

You do. Fine-tuned weights, adapters, prompts, evaluation sets, retrieval indexes and the application code written for you are deliverables assigned to you on payment. Where a base model is licensed from a third party, its licence terms are named in the proposal before any work starts, and the architecture is built so that model can be replaced without losing what was built around it. Altuon retains its own methods, internal tooling and pre-existing components, and licenses them to you for the life of the system.

Not without a written decision. Retrieval and prompting are tried first; fine-tuning is proposed only when the held-out set shows they are insufficient, and only on data your data-protection function has cleared for that purpose. Prompts, completions and embeddings are logged inside your data plane for evaluation and audit, under a retention schedule you set, and are never used to train anything outside your engagement. Where a hosted model is in scope, the provider's terms on training and retention are written into the sub-processor agreement, and a provider that reserves the right to train on your content is not used.

Yes. The data-plane decision is made in the Define phase and recorded in the architecture register: Switzerland, the European Union, the United States, your own premises or a national cloud. Inference, embeddings, vector stores and logs all run inside that boundary. The control plane holds configuration, evaluation results and metrics but never your documents or your prompts. For a sovereign deployment the proposal states what you must provide, including hardware, network and identity, and the model options that fit that footprint, so the residency commitment is costed rather than assumed.

Every model has an evaluation set built from your real cases and scored by your experts, with adversarial cases added for injection, leakage and out-of-policy requests. The set runs on every change, and a release that regresses does not ship. In production, sampled outputs are reviewed by a person against the same rubric, and alerts fire when accuracy, refusal rates or latency move away from the baseline recorded at go-live. Drift is treated as an incident with an owner: the register names who decides whether to retune, roll back or retire, and the runbook says how.

It is designed per case, not asserted in a policy. For each model the governance pack states where a person reviews before an output is used, where a person approves before an agent acts, and where a person can override afterwards, with the interface built so that doing so is quick and recorded. Where the EU AI Act classifies a use as high-risk, the documentation, logging and human-oversight duties it imposes are met in the design and evidenced in the register. Oversight rates are measured, and a reviewer who approves everything without reading is a finding, not a success.

The system keeps running. Everything it needs is inside your data plane and assigned to you: weights, prompts, indexes, code, evaluation sets and the runbook that tells your team how to change a prompt, swap a model, roll back and answer an auditor. Exit is rehearsed during the Prove phase, so the day it happens is not the first time your team has operated the system alone. Where Altuon-licensed tooling is in use, the licence survives the engagement for the life of the system, and the proposal names the notice period and the handover work included in it.

The proposal lists every third party that could touch your content, by name, role and location: model providers where a hosted model is in scope, infrastructure providers where you do not provide the plane, and any tooling that processes prompts or outputs. Each is approved by you before it is used, and a change to the list is notified in advance with the right to object. For a fully sovereign or on-premises deployment the list can be empty apart from Altuon itself, and the agreement says so in writing.