Try it. Break it. Tell me what broke.
Ranex is pre-release. Four ways to help build it, in order of what costs you the least: run the credential test on whatever you already use, clone the MIT kernel and try to break it, take one of the gaps stated in the README, or apply to build the parts that don't exist yet. The verdict path works today and gates this repository's own test suite — that's checkable in the git history, not a promise. Flow graphs and scenario compilation are designed, not built. The README says plainly: don't point Ranex at a real agent yet.
Run the credential test
Pull every model credential off the machine, re-run whatever AI coding tool you're already using — not just Ranex — and see whether a single verdict changes. It costs nothing and needs no signup. The full test lives on the resources page.
Read the credential test →Break it and tell me what broke
The kernel is MIT and public. Clone it, point it at a suite, and try to make it pass something it should refuse — a claim with no evidence, a self-approved gate, a stale commit.
Tell me your worst AI failure story too. I read every one — it's research, not a promised feature.
Take a gap
The README's Known gaps section is a live list of what Ranex cannot yet prove. None of them are softened for this page.
- The confinement controller is same-uid trusted infrastructure. The controller subprocess that invokes the sandboxed session still runs as the same user (ADR-023). Controller environment narrowing is a named follow-up, and the model credential still sits in a network-open loop.
- Approver identity is unauthenticated. --approver is a plain string, so a producer can currently name anyone as their approver. No-self-approval compares those unauthenticated strings, not verified identities.
- The journal does not detect rollback or truncation. It catches tampering with any row already written, but an internally consistent earlier prefix still verifies after later rows are removed. Unassigned.
- No flow graph, no scenario compilation, no budget, no escalation. Designed, not built. The suite manifest also freezes test IDs, not test bodies, and evidence.json is overwritten rather than appended.
Then ask any other AI coding vendor the four questions on the resources page.
Become a design partner
One lane among four, not the whole page. Pre-release, and said plainly: the verdict path works today and gates this repository's own test suite — that's checkable in the git history. Flow graphs and scenario compilation are designed, not built. The README says don't point Ranex at a real agent yet. Applying means helping build the parts that don't exist — not getting in line for a launch.
Opens in new tab • No spam, unsubscribe anytime
- A direct line to the person building this — no support queue.
- Updates only when there's something real to show, never a launch countdown.
- A say in what gets built next — flow graphs and scenario compilation are designed, not built yet.
What the kernel enforces today — useful before you break it or take a gap
Fail-Closed Gates
- Absence of evidence blocks
- No self-approval — approver identity isn’t authenticated yet
- Non-blocking gates are refused
- Verdicts from code, not confidence
Frozen Tests
- Tests locked before any code
- Diffs touching tests fail instantly
- Red-then-green, enforced
- Edge coverage as a gate
Evidence With Receipts
- Bound to the exact commit
- Ed25519-signed records
- Committed public keyring
- Claim ↔ command binding
Hermetic Observation
- Materialised committed tree
- Environment built from empty
- Pinned toolchain
- Sealed, offline suite runs
The Journal
- Append-only by construction
- Hash-chained rows
- One-command verification
- Full replay from the record
The Worker Loop
- One isolated worktree per task
- Diff on disk, summary discarded
- Keyless judging — CANDIDATE only
The honest deal
The kernel is MIT open source and stays that way. Paid plans will layer private features on top — never restrict the code that decides pass or fail.
No spam. Unsubscribe anytime.