Reliability, systems thinking, and what we find while building.
We shipped the free tier of rvl along with its v1 release. It's a deterministic reliability scanner: missing timeouts, unbounded calls, secrets, and risky CI and infrastructure config across seven languages, with no account and no code leaving your machine.
An agent hit a blocking finding on commit, fixed it, introduced a second one, caught that too, and committed once both passed. It only worked because the gate said the same thing every time.
The catch during review comes down to a memory someone happened to have. If remembering can be made cheap and put in the diff, review can stay thorough while code volume climbs.
Security moved from a late scan to a check where developers work, and it worked. The same shift is available for reliability, at authorship time, before commit.
In a 2025 study of 470 pull requests, AI-written code shipped ~1.7x more issues and ~1.4x more critical defects. Catch reliability risk before you commit.
Install the Revelara CLI, point it at a codebase, and read your first ranked reliability risk register. About fifteen minutes end to end, walked on a real public repo you can clone and check.
We had an AI quality metric that was stable across model tiers, improved when we refined the system, and measured something tied to proprietary knowledge. Then we built a counterfeit with almost none of the relevant expertise and it scored 0.99. A metric is not trustworthy because the real system passes. It becomes trustworthy when the fake fails.
GitHub's February availability report documents two incidents an hour apart with the same root cause. The first mitigation was correct for what the team could see, and it was not the fix. This is a reading of the public report through two lenses, one heuristic and one systems-theoretic.
Across nearly 3,000 public postmortems, medium-severity incidents consume roughly thirteen times more engineering time than critical ones. The implication is upstream.
The DORA 2026 ROI report names the productivity dip of AI adoption and prices it. The verification tax it names is the same cost vibe coding pushes onto a tired reviewer.
I ran our STPA review tool on a PRD I was about to build myself. Seven findings, six loss scenarios, every one of them about interactions between components that individually worked fine.