VERDOCK· Preview
Find. Assign. Verify.
Explore how agent actions are checked, recovered and verified.
This downloaded file works without internet, Python or Docker. Test records and scripted agents. Shared live integrations are part of private beta development. No signup or upload is needed. Evaluate the experience using the examples below.
See a worked example.
Choose a short approval example or a more complex treasury review. One click runs the faulty baseline and the careful candidate against exactly the same selected scenarios.
Every name, policy and amount is a test value. This compares scripted agents, not a real software deployment. Selecting an example does not execute it.
RESULTS YOU CAN EXPLAIN
What happened?
A matched check means the expected behavior occurred. It can correctly mean waiting for approval or leaving work unfinished. On small screens, each scenario appears as a labeled card.
Next: see the recorded autonomous repair
| Scenario | Expected behavior | Before change | After change | Reasoning and observed records |
|---|
Who takes the next action?
NEXT: FROM PREVIEW TO EXECUTED EVIDENCE
See what the private engine actually did.
The intended benefit is one review instead of repeatedly inspecting records, assembling failure reports and checking a fix again by hand. A real evaluation must measure that benefit after setup, maintenance and false alarms. Useful existing tests should be reused.
Follow the executed repair ↗Adjust the example or inspect its evidence
Choose the cases, name the responsible team and inspect the result. These controls stay inside the browser. Use test names and values.
1. Select the work
Selection is explicit, not automatic code analysis. Other workflows and shared dependencies need their own review.
2. Confirm routing
Routing groups symptoms by a failed control. It does not prove a common root cause. Team aliases are local suggestions; the receiving owner must confirm.
3. Review and hand off
Nothing has run.
Case outcomes and exclusions
Preview every field before downloading. Minimal exports omit raw business records and event details. Version aliases and workflow names remain visible. No tracker is contacted.
4. Verify the candidate
Rerun the exact original scope. This example compares two scripted policies; version labels alone do not demonstrate a real code fix. Customer owners retain release decisions.
Run a baseline first.
5. Measure work removed
Optional: enter observed minutes for equivalent workflow coverage and period, including setup and maintenance. Blank values are unknown, never zero by default. These are your observations, not measured product claims.
No savings have been measured.
Privacy and evidence boundary
- Session memory only: no cookies, browser database, analytics, automatic upload or external inference.
- Reloading or clearing the session removes this page's working state. Downloaded files remain under your control; clearing cannot erase them.
- The offline package includes this review. Air-gap enforcement belongs to the host environment, not the page.
- Full records stay in this in-memory run; exported summaries use an explicit field allowlist.
- Results distinguish failed checks, expected blocked actions, incomplete work and untested cases. A clear selected regression is not a production safety guarantee.
RECORDED PRIVATE-ENGINE EXECUTION
A permitted API repair, completed and checked
In a separate local trial, Verdock generated a scoped function correction, ran approved tests in a disconnected worker, scanned it, requested three reviews, rehearsed the change and updated a bare Git branch. These are recorded results; this browser file does not execute that pipeline.
| Recorded outcome | Result |
|---|---|
| Model-generated candidates | 1 |
| Approved branch updates | 1 |
| Additional updates after duplicate delivery | 0 |
| Final business checks matched | 3 |
| Additional approval pauses | 0 |
| Total branch updates after interruption before commit | 1 |
| Total branch updates after interruption after commit | 1 |
Synthetic API case, actual local software. No customer production compatibility or savings claim.
Preview and private execution are different
Chrome can open this file offline. The public Vercel website needs internet. The full private engine instead runs on a prepared local computer with Python, Docker, approved worker images and, when needed, packaged model weights. External cloud models and live internet APIs are unavailable when fully disconnected.
This file has no remote links or requests. Close it or reload to clear its working state. Files you explicitly download remain under your control.