Running AI when a vendor can go dark

A model API can disappear from a company’s operating environment without the model becoming technically worse. Access policy, sanctions, payment rails, regional support, contracting, or a provider incident can interrupt service. If critical procedures live inside that vendor’s assistant, recovery becomes a rewrite under pressure.

A VPN is not continuity architecture. It does not solve lawful procurement, data processing terms, billing, credentials, service support, or predictable availability. The answer is to separate company capability from inference supply.

Keep the procedure in files

In Neuro OS, the role’s skills, project memory, connector definitions, permissions, tests, and output schemas live in a git repository. The model receives a scoped package for the task. It does not own the procedure or the durable memory.

Avoid provider-specific conversation stores as the system of record. Preserve source references and approved decisions in project memory. Wrap model calls behind a small interface and keep structured outputs portable.

Define three inference lanes

The primary lane uses the preferred approved provider for quality and cost. The sovereign lane uses a Russian provider, VPC endpoint, or self-hosted open-weight model for workloads with residency or access constraints. The emergency lane prioritizes continuity and may offer lower quality or narrower capabilities.

Map every role to allowed lanes. Sensitive data must not fail over to an unapproved endpoint. If no compliant model is available, queue or stop the task. Silent policy degradation is worse than downtime.

Prepare for a one-week migration

Day one inventories roles, models, data classes, and hard dependencies. Day two runs representative evaluations against the fallback. Day three fixes output adapters and model-specific instructions. Day four load-tests sandboxes and connector paths. Day five rehearses key rotation and traffic switching.

Days six and seven run shadow traffic, compare accepted outcomes, train reviewers on changed failure modes, and decide whether to cut over. This schedule is plausible only if skills are files, connectors are independent, and evaluation cases already exist. Otherwise the “migration” is a new implementation.

Test quality beyond benchmarks

Use real Russian-language tasks, company terminology, long documents, tool calls, and adversarial content. Measure acceptance, factual errors, schema validity, latency, reviewer minutes, and cost per accepted case. A fallback that scores well on public tests may fail on a 1C export or internal contract format.

Record model versions and serving configurations. Open-weight is not one stable product: quantization, context implementation, hardware, and host can change behavior.

Secure local operation

Self-hosting removes one external dependency and introduces operational ones. Plan GPU or CPU capacity, patching, model provenance, monitoring, backups, and on-call ownership. Isolate each run. Keep credentials outside sandboxes and expose only approved connector operations.

Allow/Ask/Block remains model-independent. Reads may be Allowed, consequential writes default to Ask, and prohibited access stays Blocked. A local model does not become trustworthy merely because it runs nearby.

Contract for exit

Procurement should require export of configuration and logs where applicable, documented retention, usable rate limits, and clear termination handling. Keep company API keys under company control. Avoid connector logic that only works inside one provider’s proprietary agent product.

Run a quarterly failover exercise. Disable the primary lane, process a representative queue, and measure recovery time and quality loss. Update the playbook from evidence.

Sovereignty is not a promise to use only local models. It is the ability to choose cloud, VPC, regional, or self-hosted inference without losing the company’s operating knowledge. When the model is a replaceable component, a vendor going dark becomes an incident to manage, not an existential rebuild.

This work runs on Neuro OS. To scope a first role, get started.

Explore Neuro OS →