AI output verifier
Paste what ChatGPT, Copilot or Claude gave you. Real validators check it and produce a Trust report with green, amber or red items, why each matters, and a fix.
Try an example, each one is a mistake AI assistants really make:
Regex test strings (optional, one per line)
What the verifier does differently
Language models write code that looks right. This page runs real checks on it: 219 written rules across twelve kinds of output, plus the cron engine, a time-boxed regex worker and the Unravel Java analyzer. Every finding explains why it matters and gives a fix with a link to the relevant tool or concept page.
- Cron. Detects the dialect, counts fields, warns about day-of-week numbering (Quartz counts Sunday as 1), the day-of-month OR trap, uneven steps and daylight-saving windows, and lists the next runs.
- Regex. Compiles it, flags Java, JavaScript and PCRE differences, and measures catastrophic backtracking in a worker that is stopped after 0.7 seconds.
- SQL. UPDATE or DELETE without WHERE, SELECT *, comma joins, string concatenation,
= NULLand dialect mixing. - Config and infrastructure. Dockerfile, Compose, Kubernetes and GitHub Actions rules for roots, latest tags, missing limits, unpinned actions and script injection. See the config doctors for the full rule lists.
- Java. Swallowed exceptions, SQL concatenation, resource leaks and N+1 queries, using the same analyzer as Unravel.
How to read the Trust report
Green means no known trap was found. Amber means it probably works but carries a risk worth a decision. Red means a high severity problem, such as a statement that changes every row or a pattern that can freeze a server. Use the copy button to paste a short summary into a pull request or chat.
Frequently asked questions
Can an AI answer be trusted if every check passes?
A clean report means none of the known traps were found, which is a good sign but not proof. The checks are deterministic rules, so they never hallucinate, but they only know the mistakes written into them. Run anything important in a test environment first.
What kinds of output can I verify?
Cron expressions and crontabs, regular expressions, SQL, JSON and YAML, Dockerfiles, Docker Compose, Kubernetes manifests, GitHub Actions workflows, Spring Boot configuration, Maven pom.xml files and Java snippets.
How is this different from asking another AI to review it?
A second model can repeat the first model's mistake with confidence. These checks use real parsers and written rules: the cron engine computes the actual next runs, the regex is executed in a time-boxed worker, and the JSON is parsed by the same JSON.parse your code uses.
Is what I paste uploaded anywhere?
No. The checks run in your browser. The Java check loads the same in-browser parser as Unravel and does not send your code anywhere.
How does the regex backtracking check work?
It runs your pattern against a hostile test string inside a throw-away Web Worker with a 0.7 second limit. If the worker has to be stopped, the report shows a measured catastrophic backtracking result.