Verification Engine
AI code verification tools
Check AI-written changes before merge with independent verification and a reviewable evidence record. Evaluate production controls against your own workflow.
For engineering and QA teams deciding how to check AI-written code before merge. A generated change still needs independent review, tests and an explicit release decision.
What Zof does
An independent check
Zof is an independent verification and control layer between AI and production. Start by checking the change, recording what was checked, and keeping missing evidence visible to the reviewer.
Verification chosen from the change
Zof determines which verification is appropriate from the change, the System Graph, the risk and your policy, rather than running every possible test on every change. Existing test suites and CI results count as evidence.
Platform evaluation
The Verification Engine is the platform path for broader testing and control. Confirm the checks, connectors, evidence retention and approval enforcement available in your proposed deployment.
From a change to a review decision
- 01
Select a change
Choose a supported repository and an explicit diff or changed files. Keep existing code review, unit tests and CI checks in place.
- 02
Inspect the result
Read findings, changed-file scope and any incomplete execution. A clean result means only that the checks that ran found no matching issue.
- 03
Make the decision
Review the result alongside executed tests. Configure your CI rules explicitly if a result should block merge. A report alone does not enforce a production gate.
What evidence you get
Results with their artifacts
Structured results, artifacts, traces and logs are kept together in an evidence bundle attached to the decision. Confirm the formats your reviewers need during evaluation.
Execution status
An incomplete or missing check is not a pass. Evidence should show which checks ran, which were skipped and which failed to complete.
Existing results count
Playwright, Cypress, Selenium and unit framework results keep running in your pipeline and become part of the evidence.
Prerequisites and limits
- Checks can miss defects. A passing result does not prove functional correctness, security or compliance.
- Supported languages, frameworks and connectors depend on the deployment. Confirm them on a representative repository before relying on a result.
- Running your tests means running trusted code. Agree which commands may execute, where, and with which credentials.
- Hosted integration, evidence retention and approval enforcement must be validated in a scoped evaluation for your deployment.
In practice
Define what a convincing verification result looks like
A check is useful when its execution answers a concrete question about a change. Agree on the target, expected behavior and acceptable result before running it. That makes a failure actionable and a passing result interpretable.
01
Select a target and capability
Choose the application or repository and the checks relevant to the change. Establish the environment, credentials and permitted execution scope. Available capabilities and prerequisites should be confirmed in your evaluation.
02
Evaluate actual execution
Read the observations behind the result: assertions, findings, logs and artifacts where available. Distinguish a behavior failure from a run that could not complete because of authentication, connectivity or setup.
03
Compare with the release criteria
Check whether the result covers the revision and environment being reviewed. Review missing evidence separately. A result from an earlier revision may be useful context but does not verify the new change.
Illustrative scenario · not a customer result
An API response changes
A pricing endpoint adds a new discount field. Check the response contract, a valid discount, an invalid discount and a consumer that ignores the field. A successful HTTP response does not establish that totals or downstream behavior are correct. Retain the assertions and input cases behind the result.
Questions to resolve before you begin
Should every change run every test?
Use risk and dependency context to select relevant checks, while retaining baseline regression requirements. Broader verification remains useful when impact is uncertain or the release changes shared behavior.
What happens when a check cannot run?
Investigate the execution problem or obtain other required evidence. An unavailable target or missing prerequisite must not be interpreted as a successful behavioral check.
Next step
Start with one change
Bring a representative change, the tests you already run and your release rules to an evaluation on your own stack.