Each critic is a model from a different vendor, running in its own sandbox with no web access and no view of the other critics. It sees the change, in the context of the repository, and your repository's AGENTS.md.
Independent critics. Fixed rules. One verdict per commit.
Critics from competing model vendors review each pull request in isolation. A fixed rule, not a model, turns their findings into one verdict. It is signed, bound to the commit, and posted as a GitHub check.
The five stages
From pull request to check.
A critic returns findings, not a score. Each says why it matters and what the fix is, and is tied to the file and line where it has one.
A fixed rule, not a model, turns the findings into one verdict. The same findings under the same rules give the same verdict, every time.
The verdict is signed and bound to the exact commit it judged. Amend the commit and it is reviewed again; a stale verdict never covers a new change.
The verdict posts to the pull request as the cerebe/critic check, with
the findings and their evidence. Required when you choose, on a paid plan,
and then it fails closed.
What we publish
Everything you need to verify a verdict.
We publish the verdict, every finding, the policy it was judged under, and which reviewers judged it. Anyone with the CLI and read access to the repository can verify a verdict's signature and its binding to the commit.
Local and hosted
The same verdict math, on your workstation and in the check.
On your workstation, the Cerebe CLI reviews every commit, alongside Claude Code, Codex and Cursor, so problems are caught before there is a pull request. The hosted check reviews every pull request and is the approval of record. The workstation review is the author's convenience; the check is what gates merge.
See it on your own pull request.
Install the GitHub App on one repository and your next pull request gets a verdict as soon as it is opened or updated.