I’m rebuilding history’s worst software bugs so you can break them yourself

On 4 June 1996, Europe’s new Ariane 5 rocket lifted off on its first flight. About 37 seconds later it was in pieces, along with roughly $370 million of rocket and satellites.

The cause wasn’t an engine or a weld. It was one line of code that converted a number into a box too small to hold it.

You can do it yourself, right now:

Fig. 01 Ariane 5 · 1996 · int16 registerDrag the slider

Horizontal bias · 64-bit float

float64

int16 limit32,767

Move the slider past 32,767 and you’ve recreated the failure. That’s the whole idea of this series.

What Post-Mortem Lab is

Every episode takes one real engineering failure and rebuilds it as a small, faithful simulation you can run in your browser. You get to:

1. Hear the story. What happened, who was affected, and why smart people missed it.
2. Trigger the bug. Not a diagram of it. The actual mechanism, running.
3. Fix it. And see why the fix works, with tests that prove it.
4. Keep the lesson. One idea you’ll recognise next time it shows up in your own code.

All the code is open source. Every episode is a tagged release, so you can check out exactly what you saw in the video.

Season 1

Ep	Failure	Year	The bug in one line
01	Ariane 5	1996	A 64-bit number squeezed into 16 bits
02	Mars Climate Orbiter	1999	One team used pounds, the other used newtons
03	Patriot missile, Dhahran	1991	A clock that drifted a third of a second
04	Therac-25	1985–87	A race condition in a radiation machine
05	Knight Capital	2012	An old feature flag, reused, cost $440M in 45 minutes
06	Cloudflare outage	2019	One regular expression took down a global network
07	CrowdStrike	2024	A config file that crashed 8.5 million computers
08	Your pick	—	Vote in the comments

How it’s built

Everything sits on a shared toolkit that grows with every episode: checked number conversions, fault injection, deterministic timing, unit-aware maths. By the end of the season it will be a small library for safely causing failures, useful well beyond these episodes.

I build it with AI as my pair programmer, and I’ll show that honestly, including the parts where it gets things wrong.