Evidence-first stories about systems that should not behave this way, but do.
A lead is not a fact. Nothing moves from candidate to the catalog until its evidence, its uncertainty and its reproduction status are written down.
The falsehoods, ordered by wtf plus denial: how hard the behaviour is to believe, plus how confidently an engineer will tell you it cannot happen. Damage is deliberately left out, because ranking by damage gives the list of famous disasters everybody already has. These numbers are editorial judgements, not measurements.
A belief that is false, stated atomically, with the rule it yields. This is the half meant to be useful at three in the morning.
The belief that an open circuit is electrically inert, so a component wired to nothing can be ignored. Drafted here as a falsehood to bind evidence to; nothing is established yet.
Treat an unexplained physical change as a real input until measured, because not connected is a claim about a schematic and not about a circuit.
Shown by: The MIT AI Lab switch labelled magic and more magic
The belief that propagation delay is negligible next to processing time, so distance never has to enter a latency budget. In fibre, light carries about 200 km per millisecond, so a round trip costs roughly one millisecond per hundred kilometres of distance. Any timeout tight enough turns that into a hard geographic boundary: near peers succeed, far peers fail, and the application reports an ordinary failure rather than a map.
Put the round-trip light-speed floor in every latency budget before you tune anything else: about 1 ms per 100 km in fibre. A budget under that floor is not a timeout, it is a distance limit.
Shown by: The 500-mile email
The belief that content sniffing is a property of the content: feed the same bytes in and you get the same type out, so a sniffed type can be trusted to route behaviour. A signature is only as specific as the rule that encodes it, and a rule that silently matches on a prefix will claim files it has no business claiming. The caller sees a confident answer, never a near miss.
Never let a sniffed type decide behaviour when a declared type is available, and when you must sniff, treat a match as a hypothesis: log what matched, and fail closed on an ambiguous one rather than acting on the first rule that fires.
Shown by: OpenOffice could not print on Tuesdays
The belief that memory is a faithful medium, so a value that moved implies code that moved it. Drafted here as a falsehood to bind evidence to; nothing is established yet.
On any total that matters, carry a checksum or a redundant count, and treat a discrepancy that is a clean power of two as a hardware hypothesis rather than a logic one.
Shown by: The Belgian 2003 election count said to have gained exactly 4096 votes
The belief that a drive's service time depends on the queue in front of it and nothing physical outside the chassis. Drafted here as a falsehood to bind evidence to; nothing is established yet.
When latency moves and the queue did not, look outside the chassis before you look at the code.
Shown by: Shouting at a disk array said to raise its latency
The belief that throughput is bounded by available work rather than by whatever happens to wake the process. Drafted here as a falsehood to bind evidence to; nothing is established yet.
Before concluding a loop is starved of work, check what wakes it, because a scheduler that only runs you on an interrupt makes an unrelated interrupt source part of your performance envelope.
Shown by: A program said to run faster only while the mouse is moved
The belief that inserting a leap second is an ordinary tick that software need not model. Drafted here as a falsehood to bind evidence to; nothing is established yet and no source has been reviewed for it.
Treat a clock that can step, repeat or stall as a hostile input, and never let a wait be bounded by a wall clock you do not control.
Shown by: The 2012 leap second said to have taken down unrelated services at once
What actually happened, with its sources, rating the observation separately from the explanation.
A statistics department at the University of North Carolina at Chapel Hill could reportedly send email no farther than roughly 500 miles. A vendor OS patch had downgraded Sendmail 8 to Sendmail 5 while leaving the Sendmail 8 configuration in place; the older binary ignored the long-form option names it did not understand and had no compiled-in defaults, so the SMTP connect timeout became zero and aborted after about three milliseconds. On a low-latency fibre network, round-trip time tracked distance closely enough to produce a delivery radius. The account is first-hand and its author stands by the behaviour, but he states in his own FAQ that the narrative was altered for effect, that the three-millisecond figure was not the connect timeout, and that no logs, notes or configuration survive.
Falsehood: The speed of light is always fast enough
On Ubuntu in 2008, printing from OpenOffice failed on Tuesdays and worked on every other day. The cause was one line in the `file` command's magic database. A rule meant to match the twenty-nine byte string "Tue Jan 22 14:32:44 MET 1991" at offset 4 had its spaces left unescaped, so the comparison stopped after three bytes: any file with "Tue" at offset 4 was reported as an Erlang JAM file. A print job carrying the day name near the start of the file was therefore misidentified one day in seven, and the printing path acted on the wrong type.
Falsehood: A format signature is specific enough to identify the format
Leads under triage. None of this is established. They are published so that someone who knows better can say so before any of it becomes a claim.
An account circulates that an electronic voting machine in the 2003 Belgian federal election credited one candidate with 4096 extra votes, a power of two, and that the discrepancy was attributed to a single-event upset from cosmic radiation. No claim is made here about the count, the attribution, or whether any investigation reached that conclusion.
Would show: A stored number changes only when software changes it
An account circulates that the leap second inserted on 30 June 2012 caused simultaneous failures across unrelated operators, through a kernel path that spun when the clock stepped. No claim is made here about which operators were affected, which kernel versions carried the defect, or whether the failures shared one cause.
Would show: A leap second is just another second
An account circulates of a toggle switch wired into a PDP-10 at the MIT AI Lab, hand-labelled magic and more magic, with one terminal apparently connected to nothing, which nonetheless crashed the machine when flipped. No claim is made here about whether it happened, how it was wired, or why flipping it had any effect.
Would show: Hardware that is not connected to anything cannot affect a running system
An account circulates, with video, that shouting into a rack of spinning disks produced a measurable rise in I/O latency, attributed to vibration disturbing head positioning. No claim is made here about the magnitude, the method, or whether the effect reproduces outside the original demonstration.
Would show: Storage latency is a function of load, not of the room the disks are in
An account circulates of a program whose throughput rose while the user moved the mouse and fell when the mouse was still, attributed to the interrupt stream waking a loop that was otherwise idling. No claim is made here about which program, which operating system, or whether the proposed mechanism is the right one.
Would show: A program that is making no progress is a program with nothing to do