ADR-0012: Keep providers optional, untrusted and advisory¶

• Status: Accepted — retrospective baseline.
• Recorded: 2026-09-29. This is the recording date, not an invented original approval date.
• Authority and implementation boundaries: see Evidence below and the register conventions.

Context¶

Models can help propose specifications, but provider SDKs, responses and source-derived metadata create trust and disclosure risks.

Decision¶

Use a provider-neutral typed exchange with immutable trusted instructions, explicitly untrusted metadata, schema and original-request fingerprints. Validate bounded proposals before application; proposals cannot approve, generate, mutate safety policy or access sources.

Send safe metadata only. Even allowed local categories and matching categorical predicates become field-scoped labels outside the local boundary. Provider request/response/token/time/retry budgets cover the full request. SDKs remain optional; normal tests use fictional transports.

Optional semantic-value providers are a distinct narrow hook: bounded deterministic replay and synthetic namespace/privacy checks still apply. They are not authority for model-authored datasets.

Alternatives and consequences¶

Provider-specific logic in the core couples safety to SDK behavior. Passing raw examples or trusting model-declared safety is rejected. Custom application-owned adapters can use the contract without gaining filesystem/database authority.

Evidence¶

Sources: advisor guide, advisor API, semantic provider guide, support policy. Executable anchors: advisor, OpenAI, GigaChat, semantic provider.

Revisit when¶

A provider transport or SDK update needs bounded/redacted contract evidence; new model capabilities do not confer product authority.