AI-use policy / Bounded

Models may draft and inspect. Evidence and policy decide release.

External content is treated as untrusted data. Generators receive a bounded dossier, cannot access deployment credentials, and cannot edit protected workflows or security policy.

Independent review required

Required provenance before generated content ships

Source set and hashes
Prompt-template version
Provider and model identity
Sampling configuration
Output disposition
Resulting content diff and receipt

Private reasoning

Kedabek does not publish model chain-of-thought. It publishes enough inputs, evidence, outputs, rubrics, and tests to evaluate the released decision without exposing private reasoning traces.

No silent external routing

Local-only is the default. API and hybrid modes require explicit customer configuration, data classification, credentials, provider, budget, and agreement.