CaseBrief

A model call is not a decision. An AI system can return a response without producing the artifact a workflow requires, or produce an artifact that fails the contract another system depends on. When the workflow is consequential, those distinctions matter before a human considers release. CaseBrief is an experimental, Apache 2.0-licensed project being developed as evidence-control infrastructure between AI output and consequential human decisions. Its Five-Gate Architecture separates execution, artifact, contract, evidence and human authority. A human may stop or review at any gate. The implementation includes bounded model adapters, artifact handling, a frozen synthetic fixture, deterministic contract validation, offline safety tests and a read-only Reviewer View. Workflow-specific evidence assessment is human-led or proposed; CaseBrief does not autonomously authorise a decision. Three bounded experiments show why the gates are separate. Run 001 received a genuine Meta Muse Spark Open response but no required final answer artifact. Run 002 produced visible content that the unchanged validator rejected as delivered because it began with a Markdown fence. Run 003 stopped at an authentication-related HTTP 401 before a model result; it did not test structured-output performance. Each attempt preserved its outcome without retry or repair. The public repository includes the architecture, synthetic fixture, public-safe receipts, offline judge quickstart and contribution guidance. The 67 reported passing tests are offline engineering checks, not model-quality or production-readiness results. The build contract is disclosed as prior work; event-period implementation is identified separately. Evidence before confidence. Human authority stays human.