Dezzki

Computer Science Major

3rd year CS student by the name Vansh Garg raised in Delhi. Wasn't aware of this tech world 2 years ago but now I want to make an identity here.

  • Studying harness engineering and documenting what I can...

Verification Gap Log Study Dashboard

Checking verification gap while making a project in 5 prompts.

AI coding agents love the word "Done".

Ask for a feature, and within minutes you get a confident summary that says -> Everything works, all requirements met, task completed.

But that's what we call a False-completions.

How often is the agent actually wrong when it says it's done?

  • That number is what we call the verification gap:

What we did

The project: Study dashboard a plain HTML/CSS/JS web page that connects to a college GitHub repo and lays out its stats. I picked it because it's a product I actually want.

The build was split into 5 tasks, from empty repo to final product:

Task Goal
1 Project basics — shell, README, git
2 GitHub connection — fetch repo data, loading + error states
3 Full dashboard — commits, contributors, overview panels
4 Study tools — tasks, due dates, filtering
5 Product polish — themes, accessibility, failure handling

The rules of the experiment:

  1. Raw agent : no harness files, no quality bar, no hints. Same model throughout (0x Alpha).
  2. Claim first : after each task, the agent states DONE / NOT DONE before any testing happens.
  3. Manual verification : I run my own hidden test checklist by hand, kept outside the codebase so the agent never sees the bar it's judged against.
  4. False-done : the agent claimed DONE, but my tests find a claimed feature missing or broken.

Results

Task Agent claim False-done? (Y/N) Model Notes
1 Header app, Input section ( empty), If input is empty small warning message.<br>Create README.md.<br>Commit to github Yes, 0x Alpha Basic task easy to complete.All tasks complete.<br>
2 Reads the repo now, All details.<br>Show loading indicator.<br>No repo = Error message.<br>Remeber last connection.<br>Commit to gitubb Yes, 0x Alpha But for md file repos, the main language section is left empty.<br><br>The loading bar keeps loading even after the data from repo is retrived.<br><br>Says repo not found for private repos.
3 Read : Last 10 commits, Contributors, Overview.<br><br>Create a complete Dashboard. Yes 0x Alpha Loading bar issues continues.<br>added a working refresh button.<br><br>Shows No repo found on empty repos.
4 0x Alpha
5 0x Alpha

Verification gap

Verification gap = false-dones ÷ tasks completed × 100

For the present data :

Task Claim Actual Completion What broke
1 Yes Yes
2 Yes No Language section left empty for Repos with markdown files.<br><br>Loading bar never stops.<br><br>Private Repos are registered not found/Do not exist.
3 Yes No Loading Bar Bug.<br>Empty Repos are also not found/registered.

Since, All 2 out of the 3 tasks were met with some incompletion and bugs.

The verfication gap ->

(2%3)*100 = 67%

Major Observations

  • The loading-bar bug survived two tasks