Demonblade Labs · Issue 005
The Button That Ignored A Fast Thumb
A tap on "Yes, prepare the request" about 400 milliseconds after the previous step did nothing, one time in three. The cause was a 700 millisecond double-tap guard. The proof was a slowed clock, because more runs at normal speed passed on the broken build by luck.
If you build, buy or sell anything people tap on a phone, this one is for you. It sells nothing and it costs nothing to read. This follows Issue 004: the same unsolicited prototype, built from a building materials maker's public website, so the company is still not named. They have the prototype now, and the company stays unnamed. Nothing of theirs was touched.
What you will get out of it
- One bug, its one-line root cause, and the fix.
- Why a timing bug needs a slowed clock, not more runs.
- Proof on the live link, not only locally.
- Why a download that works is not a download that sells.
- The misses still open.
Chapter 0. The number is 700
Before the client saw the release candidate, an independent audit agent reviewed it. It flagged one item as "plausible, not attributed": tapping "Yes, prepare the request" about 400 milliseconds after the previous step did nothing, and a second tap worked. Reproduced on the live build, 1 dead tap in 3 at normal speed.
The consent chip and the confirm chip both send the same value, "yes". A double-tap guard dropped any repeat of the same value within 700 milliseconds, so a fast thumb on the second chip looked like an accidental double-tap on the first. The fix keys the guard on the chip that was tapped, value plus label. A genuine double-tap on one chip is still blocked.
Chapter 1. A slowed clock, not more runs
A timing bug that passes by luck will keep passing by luck. The first control run at normal speed passed on the broken build: 0 of 4 taps dead. A green result on a bug you know exists is a warning about the test, not good news about the build. So the page clock was slowed 10 times.
| Build and clock | Runs | Dead taps |
|---|---|---|
| Broken, normal speed | 4 | 0 of 4 (luck) |
| Broken, clock slowed 10x | 3 | 3 of 3 failed |
| Fixed, clock slowed 10x | 3 | 3 of 3 passed |
| Fixed, live URL behind its password | 3 | 0 of 3 dead |
The slowed-clock test is now a permanent regression test. The suites stayed green: sales browser 87 of 87, actions 58, cart 58, command selector 40, and 4,474 unit checks.
Chapter 2. A download that works is not a download that sells
The same morning, the staff dashboard's PDF reports downloaded fine, 10 of 10. But one table header was cut off with "..." and the pages were half empty. They were rebuilt as branded report dossiers with wrapped text and page numbers, then re-checked on the live link: 6 of 6 PDFs branded, 0 cut-offs.
One more detail. A live check seconds after the deploy read the old edge cache. The re-run a minute later read the new files. Wait for the edge before you call it.
The limits of all of this
- The assistant's first answer is still 839,235 bytes against a 600,000 byte target.
- The accessibility and performance budget misses disclosed in Issue 004 are still open.
- An auditor found this bug, not my own tests, which did not yet cover two different controls sending the same value. They do now.
- It covers one maker's public catalogue, and whether any of it changes a purchase is UNKNOWN until they have seen it.
Your three questions
Does your tap guard key on the control or only on the value? Do your timing tests run on a clock you can slow down? When a test passes on a build you know is broken, do you trust the test or the build?
What has to be true before I call a bug fixed: it fails on the broken build for a reason I can name, it passes on the fixed build, the test stays in the suite, and the fix is checked on the live link after the edge has caught up.
If you build for phones, I would genuinely like to hear about the worst timing bug you have chased. Reply on X with the one that took you longest. Would it be okay if I asked you for the one control in your app that people tap twice?
Receipts
- Every figure above is a test run result or a count from the project, measured on 1 and 2 November 2026.
- The catalogue is the maker's public web pages and data sheets. No private data, no login, no account.
Next issue: The exam the builder never saw.
Previous issues: I built a product assistant that is not allowed to guess, I built a gym that judges its own best strategies as luck, I freed 334 gigabytes and the disk got smaller, Why more AI agents don't automatically mean more output. All issues: Demonblade Labs.