The agent that wrote the program should not be the only one who can explain it

Most code review asks a human to reconstruct intent from source. That gets harder when an AI agent can generate more code than a reviewer can read. Revia explores a different contract: an agent writes an executable program; a human reviews the checked semantic graph that describes what the program is allowed to do.

What is public now

The Revia source repository is public under MPL-2.0. The current native preview includes project manifests, module checks, deterministic build artifacts, translation, execution, and a bounded server slice. The checked-in project template now runs through project-check, project-test, project-build, project-translate, and project-run. The latest CI run passed on Linux x64, Windows x64, and source validation at commit 7c242e1.

That evidence is intentionally narrow. Revia is not Stable V1.0 or a production backend; TLS/auth, complete A2A/MCP interoperability, and the remaining platform targets are still outside the verified claim.

A falsifiable challenge

Clone the public repository. Can you make a project pass review while causing the translated or executed program to differ from the reviewed source or manifest? Try stale digests, an undeclared file, or a changed project handoff. Include the exact commit, platform, commands, exit codes, and smallest counterexample. A no-defect control matters too: show that the check rejects the wrong object without rejecting the intended template.

Open a focused PR or reproducible issue directly; you do not need to wait for an invitation. The question is not whether an agent can write code. It is whether another agent or human can verify what that code means before trusting it.