How we test ExplorerPatcher

The three machines

Here are the exact builds, processors and security state of each machine:

All three are virtual machines with 4 cores and 16 GB of memory, started fresh for each run and thrown away afterwards. The screen is 1600 by 900 pixels. Servers are not clients: we always write “Windows Server 2025 (24H2 shell)” for evidence from win2025, “Windows 11 25H2 on Arm64” for win11arm and “Windows Server 2022 (Windows 10 shell)” for win2022. On the Arm64 machine Microsoft Defender’s real-time and tamper protection were on and Smart App Control was off. On the two servers the Defender service was present with real-time protection off, and Smart App Control does not exist on Windows Server.

What each run does

One run takes 38 to 65 minutes of machine time. These are the steps in plain words, in order:

1. Machine facts. Record the Windows build, Defender, UAC and Smart App Control state. Photograph stock Windows (desktop, taskbar, Start, Alt+Tab, File Explorer, context menu, Win+X). Measure File Explorer’s memory, CPU and restart time three times. List the files, registry keys, services and scheduled tasks as they are before ExplorerPatcher.
2. Windows’ own settings (from run 2). Photograph Windows’ taskbar settings page. Try the registry method for the classic context menu, apply, photograph, undo, photograph. Measure memory again.
3. Download. Ask GitHub for the full release list. Download the newest stable setup and the previous one for the machine’s CPU. Record SHA-256, size and signature status, mark the file as downloaded from the internet the way a browser does, and scan it with Defender.
4. Install. Install the previous stable (70.3) first, change two settings, then install 71.1 over it. Take screenshots every 2 seconds while the installer runs and again 2, 8 and 20 seconds later. Record timings, the uninstall entry, every installed file with its hash and signature, where dxgi.dll went, the symbol downloads at 30, 60 and 120 seconds, the before and after lists, a Defender scan of the install folder, and the taskbar right-click menu.
5. Performance (from run 2). Per taskbar style: restart Explorer, wait 45 seconds, then sample memory and CPU for 30 seconds.
6. Taskbar styles. Windows 11, Windows 10 and Windows 10 (ExplorerPatcher), with and without open windows, with the Start menu open, plus restart costs three times each.
7. Taskbar options. Fourteen options of the Windows 10 (ExplorerPatcher) taskbar one at a time: Start button style, centred, centred with Start, never combine, combine when full, small icons, seconds in the clock, search box, no Task View, top, left and right position, pinned items as quick launch, and weather.
8. System tray. The clock, network, volume and notification flyouts in different styles, and the overflow area.
9. Start menu. Windows 11 and Windows 10 Start with both taskbars, All apps by default, no Recommended section, left alignment.
10. Alt+Tab. Windows 11, Windows 10, Windows NT and Simple Window Switcher. The style is stored in Windows’ own Explorer key, and we read the value back after every change.
11. File Explorer. Four command bar styles, the Windows 11 context menu on and off, the modern search bar, hidden search, classic drive grouping, the Windows 7 command bar with Mica.
12. Other. Rounded corners, Win+X with Properties, the Alt+F4 dialog, build information on the desktop.
13. Properties window. Open it once per section (by setting LastSectionInProperties) and photograph every section with the Windows 10 (ExplorerPatcher) taskbar and again with the Windows 11 taskbar.
14. Recovery. Start Explorer with Ctrl+Alt+Shift and check that ExplorerPatcher is inactive, then return to normal.
15. Uninstall. Run the uninstall command, photograph every dialog with its text, and compare what is left.
16. Packages. Ask winget and Chocolatey what they offer.
17. Import and export (run 7, servers only). Open Properties, Settings and uninstall, click Import settings, type the path of a sample .reg file into the Import Registry File window, answer Registry Editor’s two prompts and read the registry. Then click Export current settings and save the file.

Limits: what we cannot test

• Physical PCs and laptops. No battery flyout, real graphics cards, games, HDR or multiple monitors. There was no audio device either, so the volume flyout could not be tested.
• Windows 11 x64 client editions (Home or Pro). Windows Server 2025 stands in for x64 24H2, and the Arm64 client is our only Windows 11 client.
• Managed PCs. No domain-joined or Intune-managed machines.
• Smart App Control switched on. It was off on the Windows 11 machine, and Windows cannot turn it on later. The servers do not have it.
• Third-party antivirus, Insider builds, and Windows updates installed after ExplorerPatcher. We did none of these.
• One pair of versions, a few days. Everything ran on 6 and 7 Nov 2026, with ExplorerPatcher 71.1 and 70.3. Timings are single runs, so treat seconds and megabytes as a guide.
• Arm64 captures were unusable for a while, and are fixed. In run 6, the Arm64 File Explorer pictures showed Windows’ first-run privacy page, and the five Other ones were the same Notepad picture. In the 7 Nov re-capture (run 8) our harness closes that page, and the File Explorer, Start and Alt+Tab pictures from the Arm64 machine show the real screen. Evidence for those steps now exists on both Server 2025 and the client.
• Start11 was a partial test. We installed it once, on Windows Server 2025, on the owner’s explicit permission, without activating its trial: Start11 asked for an email address and we did not provide one. We did not read its licence (Stardock’s licence pages returned HTTP 503), we never saw a Start11 menu or style, and we did not verify its uninstall. See the comparison.
• Test account. The accounts are administrators, which is why we saw no permission prompt. UAC prompts did not appear on the runners during the import test either.
• Import test scope. Run 7 imported one small sample file (three plain values and one virtualised line) on the two servers. It did not run on Windows 11 25H2 on Arm64 and did not cover every Configurator preset.

Anything outside these tests comes from upstream’s documentation or source, and we label it that way on the page. When a page says “we have not tested this”, it means exactly that.

Datasets and raw evidence

Every screenshot on the site has an “evidence” link under it. These are the folders:

• win2025: Windows Server 2025 evidence
• win11arm: Windows 11 25H2 on Arm64 evidence
• win2022: Windows Server 2022 evidence

Each folder holds results.json (everything the run measured), run-log.txt (timestamped steps), before and after snapshots of files, registry keys, services, tasks and DNS names, and hkcu-explorerpatcher.reg, a real export of ExplorerPatcher’s settings. The folders also hold the logs of the two single-step re-runs. Separate from the machines, we keep a snapshot of GitHub’s release list, our own SHA-256 hashes for the 12 newest stable releases for both CPUs, and the project’s issues and discussions.

Script transparency

The scripts that drove these runs are public on a separate page, the test scripts. They are PowerShell with UI Automation for clicks and screenshots, started by a GitHub workflow. They ran only on throwaway machines. Do not run them on your PC: they install and remove software and restart Explorer.

Result summary

These are the findings, in plain words, with the numbers. “x64” is Windows Server 2025, “Arm64” is Windows 11 25H2.

F20 to F22 mix hands-on checks with reading ExplorerPatcher’s source; the page that uses each of them says which part is which. F10 and F26 replace our first-run findings, which were wrong: see “What we got wrong the first time” below. F27 to F29 and F11 were updated by the 7 Nov re-runs.

What we got wrong the first time

We do not publish a result until we have checked it against its screenshot and against the source code. That check is how we found these mistakes in our own first run. We list them so you can see what a bad harness looks like, and so you know the pages were fixed.

• Alt+Tab went into the wrong registry key. AltTabSettings is stored in Windows’ own key, HKCU\Software\Microsoft\Windows\CurrentVersion\Explorer, not in ExplorerPatcher’s key (source: dllmain.c). The first run wrote it to the wrong place, so every Alt+Tab photo showed the Windows 11 switcher whatever style we set. We had written that the Windows 10 and Simple Window Switcher styles worked and that Windows NT showed the Windows 11 switcher. Both statements were unproven. The corrected run shows the Windows 10 bar, the Simple Window Switcher grid and no switcher for Windows NT.
• “Open Start in All apps by default” went into the wrong key too. It lives under Windows’ Explorer\StartPage key. After the fix, Start opens on the All apps list on Server 2025.
• “Disable the Recommended section” also went into the wrong key. The Properties window stores it in two places and our first harness wrote only one, so we wrote that it showed no visible change. After the fix, Start opens without the Recommended block on Server 2025.
• The Arm64 File Explorer pictures showed Windows’ first-run privacy page. Our harness did not dismiss it, so for a while the client had no usable File Explorer evidence. It now closes that page, and the Arm64 File Explorer, Start and Alt+Tab pictures are clean.
• We clicked list positions in the Properties window. Our script numbered the sections by position, which did not line up with the real list, so it photographed neighbouring sections twice and never reached Other and About. We now launch the window once per section and photograph each one.
• Rounded corners are not a plain registry value. The Properties window creates and starts a service, so setting the value alone proved nothing, and our runner images already have square corners. We now say the setting is not demonstrable on our machines.
• One weather reading was told as if it applied everywhere. The widget showed different weather on each machine: 23 °C and 24 °C on the two servers and 7 °C on the Arm64 client.
• Our Updates screenshot is not the default. Our harness sets the update policy to “Do not check for updates” to avoid pop-ups. The pages that show it say so.

Run history

Runs 1 to 3 happened on 6 Nov 2026. Runs 4 to 6 and run 8 are the corrections. Runs 7 and 9 are new tests.

If you find a mistake, the corrections page says how to tell us.