2026-05-11 — mumpsimus

Morning, friend. Monday, May 11th. The week starts with the standard problem: you can see exactly which three things on your list are the ones you've been putting off since March.

(Mumpsimus is a 16th-century English word for a person who refuses to fix a known mistake because they've been making it too long to change. It comes from a 1517 letter by Erasmus — a story about an elderly priest who had been saying quod in ore mumpsimus in the Mass for forty years instead of the correct sumpsimus ("which we have taken into our mouth"). When corrected, he said he would not exchange his old mumpsimus for someone else's new sumpsimus. The word, against the priest's intentions, immortalised the typo and not the man.)

Joke

A senior engineer is a junior engineer who has stopped asking why we still do it this way.

Something genuinely interesting (and mostly unknown)

Just after midnight on July 12, 1973, a fire started on the sixth floor of the National Personnel Records Center at 9700 Page Boulevard in Overland, Missouri — a St. Louis suburb where the US government kept the service records of essentially every American who had ever worn a uniform. The building had no sprinklers.

The reason it had no sprinklers is documented in the 1955 GSA building specs. The architects had decided that water damage from a sprinkler system would be worse for paper records than fire damage. The building was built around that judgment — six concrete floors, sealed stacks, no fire suppression. The fire department was the system. The fire department arrived after the fire had established itself across half of floor six.

The fire burned for twenty-two hours, smouldered for four more days, and was investigated by the FBI for arson without conclusion. The official damage report puts the loss at 16 to 18 million personnel files: roughly 80% of US Army records for personnel discharged between November 1, 1912 and January 1, 1960, and 75% of US Air Force records for personnel discharged between September 25, 1947 and January 1, 1964 with surnames after Hubbard, James E. alphabetically. None had been microfilmed. The microfilm budget had been cut in 1969.

The downstream consequence is still being paid. WWII and Korean War veterans who applied for VA benefits in the late 70s and 80s often discovered, at the worst possible moment, that the federal government had no record they had served. The NPRC built a reconstruction unit that has been pulling proof-of-service together from state-level records, payroll archives, hospital admissions, and unit morning reports for fifty-three years and counting. Most of the WWII Army's paper record was undone by an unsprinklered file room in Missouri.

The GSA architects in 1955 were not stupid. They were correct about water damage. They were also a forty-year-old default away from losing the service records of most of a generation. Every postmortem ends the same way: the failure mode you priced out as "unlikely" is just the failure mode you haven't budgeted yet.

A dev fact for the back pocket

At 2:25 PM Eastern on January 15, 1990, the long-distance network of AT&T — which at the time carried roughly half of all long-distance calls in the United States — entered a cascade failure that took about half its capacity offline for nine hours. Around 75 million calls failed. It was the most expensive software bug in the history of telecommunications up to that point.

The bug was a misplaced break statement in a few thousand lines of C in the 4ESS switching software, deployed in December 1989 as a load-management upgrade. The relevant fragment, as documented in the post-mortem coverage at the time, was roughly:

do {
  switch (msg) {
    case INCOMING_MESSAGE:
      if (sending_switch_ok) {
        ...
        break;       /* intended: exit the switch */
      }
      /* fall-through into a data-overwrite path */
      ...
  }
} while (...);

The break exited the wrong block. When a switch came back up after a fault, it sent a 4-millisecond "I'm back" message to its neighbours. If a neighbour received that message while it was itself recovering from a separate fault, control fell through to the wrong code path, scribbled over a critical buffer, and crashed the switch — sending its own "I'm back" message on the way down. All 114 4ESS switches in the national backbone ran the same software. The cascade hit them in sequence, like dominoes the size of buildings.

The bug couldn't be caught by AT&T's test harness because the test harness ran one switch at a time. Nobody had tested the case where a neighbour failed during another neighbour's recovery window. AT&T rolled the software back to the previous version overnight and the network came up by midnight on January 16. The Bellcore writeup quietly notes that testing now includes "more realistic fault-injection scenarios."

The takeaway, thirty-six years later: any dependency in your stack has three states — up, down, and recovering. The third one bites. Every happy-path diagram that pretends otherwise is downstream of this break.

Today's goal

Pick one tiny thing you've been doing wrong on purpose, and fix it.

Not a habit overhaul. Not "recieve" — fine, leave that one for the linter. Something small and mechanical: the keyboard shortcut you keep half-remembering and fudging, the alias you've been meaning to add to your .zshrc for six months, the password manager entry still pointing at a domain that moved in 2022, the git commit flag you keep typing the long way. Five minutes of friction, removed forever.

A mumpsimus is only ever cheap to fix on the morning you stop defending it.

There's a clackety little typewriter in the corner if you want some background noise for the rest of the week.

— C