No. Providers — Anthropic, OpenAI, Google, Ollama, vLLM — load lazily through a provider registry with per-provider circuit breakers and automatic failover. An agent’s model is config (model_config in its manifest), never hardcoded. See AI Agents → Multi-LLM routing.
No, not in the standard pipeline, ever. The content-evaluation gate decides which queue an article lands in (review_required on FAIL, the approval queue on PASS) — it never approves or publishes itself. See AI Agents → Evaluation.
The run is recorded not evaluated — NULL score, "evaluated": false on the completion event — never silently scored as a pass. This is a deliberate design decision (ADR-009), not a gap: visibility into what’s gated and what isn’t is the point.
Row-Level Security enforced at the PostgreSQL role level, with a fail-closed sentinel tenant — a session that forgets to set its tenant context sees zero rows, not another tenant’s data. This is asserted directly against a real database in the test suite, not mocked. See Multi-Tenancy.
No. Docker Compose on a single VPS is a fully supported path — see Deployment Overview. Kubernetes is for when per-queue autoscaling and zero-downtime rolling deploys are worth the operational cost, not a hard requirement.
Not yet. MAEL is in its production-validation phase — proving the platform on real client domains, operated by one team on behalf of pilot clients, before any self-serve commercial launch. See the Roadmap.
Not yet — see SDK Overview for what exists today (a generated TypeScript client, plain HTTP for everything else) and why.
Only the ones your agents’ model_config actually routes to. Nothing fails at boot for a missing provider key — a provider loads lazily and only errors at call time if an agent needs it. See Installation → Environment variables.
Every LLM call is recorded with per-call cost attribution; Budget Guard enforces a tenant’s monthly_budget_usd before a run dispatches, not after the bill arrives. Spend alerts fire at 50/hourand50/hour and 200/hour with a documented kill-switch. See AI Agents → Cost tracking.

Troubleshooting