A Strange Collection of System Failures

From the notebook of an investigator who kept what he could not explain

Prologue

I was employed to find out why things failed. A useful explanation specifies a mechanism, identifies the conditions that set it in motion, and survives an attempt to make it fail again. A server that dies at a particular temperature has not committed an outrage against nature. It has reached its limit. The distinction matters to the person whose work was lost, even if it offers that person no comfort.

I liked failures with a clear beginning. When a valve opened without a command, I asked where the command might have been lost; when a machine reported a state it could not have reached, I asked who had written the state and when. I distrusted eyewitnesses, including myself, in exactly the way an engineer ought to. Time was recorded before recollections were gathered. Nothing was called impossible until the ordinary possibilities had been exhausted.

There was one incident I could not dispose of quite so neatly. A valve opened on a maintenance line at a quarter to five. It should have remained shut. A fitter happened to be reaching for his spanner and escaped injury; I had the maintenance schedule, a clean event log, and three mutually incompatible accounts of the sound it made. The eventual report was adequate for the safety committee. It was not adequate for me. I kept the papers, intending to return to them after the next shutdown.

Other papers followed. I arranged them by incident, never by theory: a count that would not add up, a program whose failure vanished under inspection, an account of a night shift that nobody wanted to put in the official record. Some were things I investigated. Others came from records, or from people willing to repeat their testimony only once. I marked the difference.

The folder acquired a name only when I could no longer hold its contents in my desk drawer. I called it a dossier because miscellaneous did not seem an adequate filing instruction. I still expected each case to yield to a patient account of what had happened.

The first one is old. I was younger then, and the man who told it to me had a biscuit in his hand.

Case Files

I. The Schaerbeek Bit

Schaerbeek, Belgium.

I am in a cafeteria in Schaerbeek, just off the tram line, waiting for a man who is late. In unfamiliar rooms I begin, out of habit, by checking how they are arranged.

The windows sit low; from inside, passing traffic moves along the upper edge of the view like a slow mechanical ceiling. The tables keep the same measured distance from the walls, aisles just wide enough for a tray to pass without contact; the chairs face each other in pairs. The queue bends once before the till in an L-shape that keeps the doorway clear.

I take my place in it, moving forward each time the drawer opens with its tired ring. Above the counter, three menus are printed in two tight columns; when my turn comes I trust the digits more than the words and repeat one of the numbers, watching the cashier's hand go straight to the corresponding row as if confirming a mapping I had only guessed at. Tray in hand, I take a table by the wall.

A colleague on the municipal-systems contract that brought me here has given me the name of the man I am waiting for. He knows the local engineering scene, I am told, and which names on procurement papers correspond to people who can actually fix things.

There is a newspaper on the next chair. I pick it up. Liège has held a waffle competition; the winner is photographed behind an enormous plate of waffles, with three judges smiling beside her. I turn a few pages, looking for science. A short item from Geneva says that a neutron escaped one of CERN's experimental enclosures that morning and was detected beyond the shielding where it was expected to stop. The laboratory says no one was at risk. The newspaper uses the word escaped three times. I fold it and put it back.

He arrives late and apologizes before sitting down. His tram, he says, underwent “a spontaneous adjustment to the route.” He pauses, then corrects himself.

“Not spontaneous. Sorry. A signal failed, or somebody changed the route. We just don't know which yet.”

We exchange the usual confirmations—my name, his, the fact that we have both found the correct café in the correct district on the correct day. He tells me he has worked elections—“not politics,” he says quickly, “interfaces.” As he talks, he keeps straightening small things: the sugar packet, the spoon. When conversation pauses, he taps the table as though confirming liveness.

I ask who actually maintains the municipal systems I am likely to encounter. He gives me names, but rarely just names. One contractor is excellent with cabling but, according to him, “loses confidence around printers.” Another is the person everyone calls after hours even though nobody's paperwork admits this. I write down the useful parts and leave some of the commentary out.

The commentary gets harder to separate from the useful parts when he picks up one of the biscuits beside his coffee.

“You can tell a lot about a nation from how its biscuits break.”

I look at him. “How they break?”

“The English ones wait for tea and then break where you expect. The French fracture elegantly, usually along the line that leaves the prettier half. Belgian biscuits begin along one crack, reconsider, and finish by inventing a third.”

That seems ridiculous. I don't see what national character has to do with a biscuit, but he is already unwrapping it.

He breaks it. The biscuit splits diagonally across the stamped pattern, scattering crumbs onto the napkin. He seems pleased.

“See?”

I say nothing about the biscuit and ask how he came to work on elections.

He gives me the short version: municipal contracts, equipment support, one assignment becoming another. Then he says, “When people talk about elections, they always mean the result. They don't talk about the day itself.”

I had asked how he got the job. He had answered a different question.

“I worked one day,” he continues, “that stayed with me more than the result it produced.”

He mentions Schaerbeek, the 2003 federal election, and a candidate with an impossible number of votes. I recognize the outline. I have heard it before as one of those programmer stories compressed into a few lines. It takes me a moment to adjust to the idea that the man across from me had been inside the room I knew only from retellings.

What follows is his account of that day. He did not tell it in order, but I will.

The polling station was a school. Its walls still carried faded posters about hygiene and primary colours, but the doors had been re-labelled with printed arrows and neutral nouns: Entrance, Exit, Polling Room. There were folding tables, metal chairs, extension leads taped to the floor, stacks of forms, pencils attached to string, cardboard screens, a clock protected by a wire cage, two wastebaskets, a cabinet of children's paintings turned to face the wall, and a carpet whose pattern is an argument against democracy.

He had been sent there as technician, a word expansive enough to cover whatever had not yet been assigned to the others. On his table lay the checklist: confirm serial numbers, confirm counters at zero, confirm seals intact; note every confirmation in ink, sign where indicated. Two party representatives sat near the door. The bureau had a president, seconded for the day from the commune, whose signature would certify the count. She had signed returns in Schaerbeek for eleven years; in that district a signed return was the count. A schoolteacher seconded as a watcher stood where she could see both the machines and the clerks and seemed unsure which of them constituted the real work.

Voting itself went normally. He answered questions, checked displays, and waited for something to require him.

After the polls closed, the clerks began checking the totals. The work moved line by line until one of them stopped halfway through a candidate's result. She looked back at the list total, then at the candidate total, and read both again.

The candidate had more preferential votes than the entire list containing him.

She asked the clerk beside her to read the two figures. He did. They exchanged places at the display and read them again.

“Check this one,” somebody said.

They checked the machine identifier, the register, the media from which the totals had been read, and then the totals again.

There is the certainty, never admitted aloud, that somewhere a check is missing and that this is sane; that checks go missing the way buttons go missing from coats. So they looked for the missing check.

The air in the school had become very dry. Paper rasped when the clerks separated sheets, and once a wool sleeve snapped against the back of a plastic chair.

For a while nobody called it an error. That word would have started another procedure: forms, witnesses, preserved media, work frozen where it stood. So they talked around it. This one. These two figures. Check it again.

The same numbers survived every obvious check. The contradiction could now be written plainly:

Invariant 1. (violated)

$$ \boxed{\text{candidate preferences } \leq \text{ party list total } \leq \text{ district total}} $$

The candidate's preference count should never rise above the list total. In other words, the list's ceiling is the candidate's sky, and yet there it was: sky lower than the bird.

The president asked the technician, quietly, whether she could sign. He understood the question: not whether the number was right, but whether the machine could be vouched for. He said he could not vouch for it. She held her pen over the box where the count would go. The figure the machine had printed would certify a contradiction; any other figure would alter the record; an empty box would fail to certify the district. She wrote the machine's figure, drew one line through it, wrote the corrected figure beside it, and initialled the correction. Then she asked the two representatives to witness it. One did. The other looked at the struck figure for a long time, and put down his pen without signing.

Eventually somebody used the word error, and the formal investigation began.

The Investigation

The technician checked the machine and the media again. The schoolteacher tried to be helpful by pointing out each time the computer made a small sound, as if identifying birds.

The woman with the adding tape asked if lightning could do it. “Not tonight,” the technician said. “Tonight is dry.”

They made a ledger of what would be convenient to blame and crossed each item off.

• a transposed digit
• a repeated line
• a wrong candidate mapping
• a duplicate import
• a machine identifier copied from the next station
• a power interruption
• lightning

Each explanation had some trace it ought to have left. None of the expected traces was there.

When the figures survived another pass, someone subtracted one total from the other and wrote the difference in the margin: 4 096. The technician recognized the shape at once:

[ 4096 = 2^{12}. ]

A difference of exactly one power of two made a bit flip an obvious suspect.

A physicist was eventually pressed into service because she happened to live nearby. She arrived by bicycle, helmet still in hand. By then nobody needed her to notice (2^{12}); what they wanted was a reason for one bit to change.

She made them go through the ordinary possibilities again. Bad input, bad arithmetic, bad code: each had somewhere to leave a trace. When those routes led back to the same number, the question shifted to the memory itself. A memory cell can be disturbed physically, without the program asking it to change. An energetic particle passing through the semiconductor can deposit enough charge to alter the bit it crosses. If that bit was the position worth 4 096, a zero becoming one would add exactly 4 096 to the stored count.

That gave them a hypothesis: a single-event upset. Under that hypothesis there would be no smoking gun—only a single flip where a zero had become a one at the thirteenth bit, a neat, round power-of-two crime.

The technician objected that bits did not change for no reason. The physicist agreed. A particle crossing a memory cell was a reason; the difficulty was that the reason would already be gone.

They could inspect software, recount totals, and ask clerks what they had done. If the cause had been a particle, there was no particle left to inspect.

Later, the machine was tested, the software examined, and the result reconstructed. No software defect was found that explained the discrepancy. The experts' report concluded that it could very probably be attributed to a spontaneous and random inversion of a binary position in the computer's working memory.

The report left the physical cause open. The committee’s report is less romantic, but it permits the word that haunts this dossier: likely.

Later retellings usually supplied ionizing radiation, and newspapers condensed that into a cosmic flea bite. By the time the incident reached programmers, the whole thing was often filed under computer error. The phrase is correct in the way that shipwrecks are wet. What stayed with me was the qualification in the report: very probably.

One clerk later admitted that he would have preferred a boring human mistake. A human mistake belonged somewhere: to a hand, a tired eye, a copied line. It could be found and corrected. He did not say this during the investigation. If human error returned to the list of possibilities, he himself was one of the humans available.

The president was summoned to explain the correction. A figure struck through in ink is what a forged count looks like; she answered questions for two days about the one she had struck, and in the end the correction was accepted, the district certified, and no one accused of anything. She did not speak to the technician again.

By the time he finishes, the cafeteria has thinned and our coffee has gone cold. He straightens the sugar packet, which does not need straightening, and does not look up.

At the door he asks me not to write down what became of the president. I promise him I will not.

At the time I heard an engineer's argument about preparation, nothing more. We do not fight the weather, he said; we prepare for it. The argument had a second half: if you believe in preparation, then you believe in a cathedral of checks where each arch braces another—triplicate logic, parity with scrubbing, watchdogs to guard the watchdogs, and the prophylactic act of voting in paper because paper fails like a person fails, slow and legible.

I did not write a field note that afternoon. Years later, when I began the dossier, I put this under his story:

Field Note #1. Horror, in our trade, is the clean error—the one that leaves no prints.

There is an older photograph I keep in a folder called Proofs I Do Not Argue With. In it, a moth lies taped to a page of an engineering log.

The page belongs to the Harvard Mark II. On 9 September 1947, its operators found a moth in Relay #70, Panel F, taped it into the log, and wrote beside it: “First actual case of bug being found.” I like the page because it makes an engineering superstition almost literal: give the failure a name, and perhaps you can banish it. Here they could do both—call it a bug, pull it from the relay, and tape it beside the record. (National Museum of American History)

The moth looks unconvinced.

The page I reserve for what became of the president is the one page in the dossier I have never been able to fill.

II. The Heisenbug I Failed To Catch

Somewhere between midnight and the first ferry.

The trouble was an order that would neither finish nor fail. A customer could press Pay, and sometimes the answer would not come. I could not safely run the request again: the first attempt might already have taken the money. The company's office was empty, so I went in alone.

The building backed onto the harbour. A strip of quay light crossed the carpet; beyond it the desks were dark. In the server room the fans kept their usual pitch. I had worked nights before. It pleased me that buildings, deprived of people, became easier to understand.

At 00:41 I caught the failure. The application stopped at the same line it had stopped at earlier, waiting for the database to answer. I attached a tracer, expecting to see the operation just before the stall. The next run succeeded. Then the next. I removed the tracer, and within minutes the process hung again.

A familiar kind of nuisance. Watching a program changes its timing; anyone who has debugged a race knows that. I wrote timing at the top of a fresh page. I was relieved to have a word for it.

I introduced print statements as if soothing a friend — tell me what you are thinking when you do this.

With the extra lines the application ran perfectly. Without them it stopped. I removed them a second time, watched the cursor blink, and saw the stall return. I felt an unreasonable satisfaction. At least there was a pattern.

The next experiment was even smaller. I left the code alone and opened a monitor on the other screen. For an hour there was no failure. I closed the monitor. The request stalled on the very next attempt.

I checked the clock. I checked that I had in fact closed the window. I reopened it; the stalled request completed before I touched anything else. That, too, was possible. A delayed response can arrive at any moment. I wrote the time down to the second.

The room had a glass partition, and my monitor cast a reflection in it. Once, reaching to turn on another tool, I caught sight of a pale movement beside my reflected shoulder. It was my own hand. I kept my eyes on the glass until I was sure.

At two I repeated the experiment. At three I changed the test so that it would run without a human operator. It failed twice while I was away making coffee. Both times the logs showed an ordinary wait, and both times the process had recovered before I returned. When I stayed to watch, it ran without complaint.

By then I had begun to dislike successful tests.

I knew what this sounded like. I had been working too long, and I had begun arranging accidents into a story about myself. A debugger perturbs the system. Coffee breaks change the load. The failures were real; the impression that something was choosing when to fail was mine.

I said this aloud. Hearing my voice in the empty office made the explanation feel less persuasive, which was another effect for which I could supply an ordinary cause.

At 04:12 I turned the main monitor away from me and left only the corner of its screen visible in the glass. I meant to see whether the request would fail without the bright window in front of my face. It did. On the reflected screen a small red mark appeared beside the waiting transaction. I turned around at once.

There was no red mark. There was no waiting transaction. The system had completed, and the log gave an ordinary completion time. I tried to remember whether I had really seen the colour or had expected to see it. The difference, which should have been obvious, would not settle.

The building's heating came on. Pipes clicked overhead; a chair creaked somewhere beyond the partition. I found myself listening between the clicks for a sound that would not belong to the building.

At 04:56 the request stopped again. I could attach the tracer immediately. Everything I needed was on the screen. Instead I sat with my hands resting above the keys.

What frightened me was not the possibility that the trace would show nothing. It was the idea, sudden and complete, that attaching it would make the thing aware of me.

I was alone, and I knew exactly how foolish that thought was. I did not move.

In the end I attached the tracer. The request completed. It left no useful evidence.

There was another test I could have run, a quieter way of watching the live system. It carried a risk of interrupting a real customer's order. I decided against it. At daybreak I wrote that decision down carefully, with the reasons. I could still have defended every one of them.

Then I added a line beneath the reasons, in smaller handwriting.

The thing hates to be watched.

I crossed it out. An hour later, before leaving the office, I copied it onto a clean page.

III. Maxwell’s Demon

A small hosting company, two rented halls.

The developer had gone to sleep sitting upright at a crash cart, with his forehead on his folded arms. He had been finishing a migration and had meant to rest for five minutes. The last thing he remembered was the cold from the aisle rising through the soles of his shoes.

In the dream he was standing at the far end of the server room, facing the door. He knew at once that it was a dream because the door was much too far away. Between him and it stood row after row of machines, receding into a distance the building could not contain. The fans were running, but he heard each one separately, as clearly as a person speaking his name in a crowd.

He tried to walk. His feet moved, yet the door did not come nearer.

The lamps along the aisle began to go out, one by one. In their place appeared small red lights in the cabinets, too low and too regular to be status indicators. They trembled with each breath he took. It occurred to him that something on the other side of the racks was breathing too.

Then he saw it.

A figure stood between two cabinets where no person could have fitted. Its head touched the cable trays, and its shoulders passed behind the steel without interrupting it. Its face was not blank. That would have been easier. It was covered with features set at different depths, like the faces of people pressing from the wrong side of a thin wall: an eye near the throat, a mouth that seemed to begin at one cheek and end somewhere behind the other. The features moved almost imperceptibly, attempting to agree on a single expression.

He recognized it as a demon with the swift certainty of dreams, and immediately tried to reject the word.

In one hand it held something pale and minute. In the other, something exactly like it. He could not tell whether they were grains of salt, teeth, or tiny pieces of paper. The demon compared them with the absorbed attention of a clerk checking two signatures. Whenever it raised one hand, the lights on that side of the aisle flickered. When it raised the other, the lights answered.

He wanted to ask what it was doing. He had worked with machines for twelve years; if a process was behaving strangely, he knew how to begin a question. But every question that formed in his mind reached his mouth already answered, and each answer was in a voice he had never heard.

It is a dream, he told himself. I can wake up.

The demon looked at him. All the faces in its face stopped moving.

He felt the weight of the entire room shift toward him. Not the floor, not the air: the room itself. The racks leaned inward by no visible amount. He heard a sound like paper being folded very slowly. He thought of the other people who had walked this aisle without knowing what watched them, and of the hundreds of rooms like it in other cities, and of how little a locked door meant to something that did not need to enter.

The demon stepped closer. Its feet made no sound. It had the smell of rain on overheated metal, and beneath that an ordinary human smell of wool damp from a journey. The human smell was worse.

“This room is cursed,” it said.

Its voice came from the doorway, impossibly far behind him.

He tried to turn. One of the little things in its hands fell, though the demon had opened neither hand. It struck the floor with a noise so violent that he expected the cabinets to fall over. Nothing moved.

“Which one?” he managed to ask.

The demon held out its hands.

He could see now that the two pale objects were not alike. Their surfaces were marked differently, and one was worn thin at the edge. The demon regarded them with a sorrowful satisfaction, as if a difficult piece of work had finally come right.

“Both,” it said.

He woke with his cheek stuck to the cold metal of the cart. His neck hurt. The door at the far end of the room stood open by an inch. He remembered pulling it shut. The fans were running at their ordinary pitch.

For several minutes he could not make himself approach the door.

He checked the system clock. At 02:47 the lights at the far end of the room had gone out during a brief power interruption. The access log recorded the door opening and closing at 02:51. Neither entry ought to have frightened him. Both did.

On his way home he passed the laundrette opposite the station. Two white shirts hung in the window, identical except for a dark thread in one cuff. He crossed the street to avoid them.

He told nobody about the dream until the following week. The first version he wrote down was short: bad dream about the racks. heard someone talking. door open when I woke. When he later told it to me, he stopped several times to drink water, although we were sitting in a warm room.

Three weeks afterward, an engineer checking two separate customer files discovered that they had the same SHA-256 hash. Not similar hashes. The same complete value, though the files themselves were demonstrably different. She checked again with another program and asked someone to watch. She was still staring at the result when the developer came in.

He said, “Don't put them together.”

Nobody understood what he meant. He would not explain it.

Both files were kept for examination. Eleven days later, a storage failure destroyed them. Nothing usable survived. The engineer who found the collision kept the digest on a folded piece of paper; she showed it to me as if the number itself might still be contagious.

That was all that happened to the files.

It was not all that happened to the developer. He moved his desk upstairs, stopped working late, and refused to pass the server-room door unless somebody else was in the corridor. Once, when a cleaner dropped a handful of coins behind him, he fell to the floor with his arms over his head. He stayed there while she apologized.

I asked him whether he would return to the room with me in daylight.

He looked at my two hands resting on the table. I had been holding a pen in each, trying to decide which one still wrote.

“No,” he said.

Field Note #2. Two different things may answer to one name.

IV. The Visitor in the Cold Aisle

The night operator asked me to turn off the recorder before he began. I told him I would not be able to quote him accurately. He said accuracy was why he wanted it off.

He had worked twelve years in the same building. It was an old warehouse with new floors and a server room in the basement; he knew the building's sounds by the hour. At night the water pipes knocked, and toward morning the elevator made a small settling noise even when nobody had called it. Neither sound had ever troubled him.

What troubled him began with a tapping from the cold aisle. Someone was striking a tool against metal, slowly, as a watchmaker might test a spring.

He called out. The tapping continued.

At the end of the aisle he saw someone bent over an open cabinet. He first took the figure for a child. Then it straightened, and that explanation became less comfortable. It was too late for visitors, and the room was locked from the inside. He had checked the door himself.

“Was it a man?” I asked.

He rubbed his forehead with the heel of his hand.

“I don't know. I thought so.”

“Was he wearing anything distinctive?”

At that he shook his head, but I thought he had done so a little too quickly.

He told me the figure had been very small, although he could not say how small; the distance between cabinets confused the scale. He remembered a greenish colour, or perhaps the green light of a status lamp falling across a dark coat. There was something bright near the face. When I asked whether it might have been hair, he said he could not have seen that clearly.

What he remembered most precisely was the stranger's manner. The figure worked as though the machine belonged to it. It leaned close, tapped once, paused, and made a sound of satisfaction. Then it reached toward the terminal and moved one small thing. He could not see the screen from where he stood.

The operator took a step forward. The stranger turned.

“I thought he was annoyed that I'd interrupted him,” he said.

For an instant he had the distinct impression of being the intruder in his own workplace. He apologized. The figure made a dismissive little movement, and he looked down, ashamed of apologizing to somebody who had no business being there. When he looked up again the aisle was empty.

In the morning a routine cleanup job had gone one step too far. Its loop had acquired an extra iteration, and a customer's active archive had been deleted. There was no suitable culprit to name in the incident report.

The operator told the shift manager about the visitor. The manager smiled, then stopped smiling when he saw the operator's hands shaking. At lunch someone asked whether he had seen a ghost. Someone else asked whether he had seen a leprechaun.

The room laughed.

That word was not in the operator's account. He was careful to explain this to me. He did not say what word he might have chosen instead.

He was put on day shifts. The company called it a kindness, and he was grateful for it in the way people are grateful when they have been spared an argument they could never have won. For weeks his colleagues placed little green objects on his desk: a paper hat, a plastic shamrock, an old toy shoe. Nobody meant to be cruel. He kept the objects in a drawer because throwing them away would have required him to speak about them.

At the end of our conversation he asked whether I knew that, in the old stories, you should never take your eyes off the little people if you want them to remain.

I said I had heard something like it.

“But I looked down,” he said.

He seemed to be trying to forgive himself for that.

After he left I noticed that I had not asked whether he had seen the visitor's hands. I had asked only what they had done.

Field Note #3. When he said the stranger was pleased, his own face changed.

V. Mercury in Retrograde

Wednesday, the day of Mercury.

I first met the astrologer because of an envelope that had arrived at the clearing house three days before its contents happened.

The clearing house moved money between small businesses. A dozen people worked above a print shop, and their machines settled accounts after dark. One Wednesday in March, the nightly settlement ran backward. The instructions unwound in the wrong order; wages reached the wrong accounts; a woman discovered that the rent she had already paid was due again. It was a serious failure, but not the kind of failure that ought to make anyone consult the heavens.

The envelope contained a drawing of two serpents winding in opposite directions around a staff, and beneath it the instruction: Do not settle accounts when the messenger is walking backward. There was no return address. The postmark and the office's incoming-mail stamp predated the failure. Someone had written the name of the on-call operator inside, very neatly, though the operator had never met the sender.

He showed it to me after we had tried to reconstruct that night. He was not particularly interested in astrology. He was interested in how a stranger could have known his name and the one thing his machines would do that had never happened before.

The address we found belonged to a narrow house behind a church. Its door was blue, and on its step stood a bowl filled with coins from several countries, all blackened as though they had been recovered from a fire. Above the bell hung a small brass wing, with a bell of its own attached to the tip.

An old woman admitted us. She did not ask what we wanted. She examined the operator's shoes and asked whether he had crossed running water on the way. He said he had come by tram, across the canal.

“Then the message has already arrived,” she said.

The room resembled the back office of a postmaster who had taken charge of every undelivered letter in the city. Wooden pigeonholes climbed to the ceiling. Some held envelopes browned with age; others held receipts bearing tomorrow's date. A timetable was pinned above the sorting desk, but its tram lines were drawn in two directions at once, each arrow pointing toward the station it had come from. Names had been crossed out on the envelopes and written again on their backs, in the same hand. Near the window an old letter scale balanced a sealed envelope against a pair of worn shoes. I could not work out which side was supposed to be heavier.

She called the room her observatory. There was no telescope.

Mercury, she told us, governs the things that pass from one person to another: letters, payments, bargains, messages, names, lies, roads, handshakes, the transfer of a thing from one owner to the next. When Mercury retraces his steps, such things may return to their senders, arrive before they are dispatched, or demand payment a second time.

“You have mistaken him for a clock,” she told the operator. “A clock repeats. A messenger returns.”

I would once have objected to the difference as a matter of language. But she spoke as though she were distinguishing two species of animal that somebody had carelessly kept in the same cage.

She placed a silver coin on the table. A winged figure was engraved on one side. On the other was a number whose digits had been worn almost smooth. The coin travelled toward the operator across the wood, slowly, without rolling. He snatched his hands into his lap. It stopped at the edge of the table and waited.

The woman covered it with a drinking glass. For a moment I saw the coin's reflection at the bottom of the glass; then I saw two reflections, one where the coin lay and another where it had been. The actual coin had not moved. I bent closer until my forehead almost touched the rim.

“Do not look there,” she said sharply. “That is where things go when they have not yet happened.”

She took us to the window. The sky was pale and empty above the church. Mercury was invisible in the daylight, but she pointed to a space near the sun as if she could see a tiny person waiting at a crossing. She spoke of the stations, those pauses when a planet seems to halt against the stars before beginning its backward walk, as days when the messenger puts down his bag and the messages inside begin addressing themselves.

The operator asked whether the clearing house would be safe when the planet turned direct again.

She laughed, not unkindly. “The bills will arrive,” she said. “That is what you call safe.”

I wanted to leave. At the same time I wanted, almost painfully, to know what was under the glass.

She told us we had come on a Wednesday for a reason. The day belongs to Mercury, as do the market and the crossroads, the thieves and the interpreters. In the old stories Hermes carries souls between worlds. There is no door he cannot pass through, no boundary he has promised to respect. A person who relies on a message reaching its destination has already entered his country.

She opened an almanac on the table and asked me to turn the pages until I found the date of the failed settlement. At first the pages held the familiar diagrams of the moon, sunrise, and the positions of planets. After a few turns the dates began to repeat. The twenty-first followed the nineteenth, then the eighteenth, then the nineteenth once more. I thought the signatures had been bound out of order, but the page numbers continued regularly along the bottom, as though the book had counted events that the calendar refused to hold still.

On the page for the Wednesday, a tiny mark stood beside Mercury's symbol. It looked like a closed eye.

“Write your name,” the woman said.

There was a pen next to the book. I asked why.

“To give the letter somewhere to go.”

I had written prose that described procedures in the future tense, as if promising the very sun, and when the sun obeyed I pretended it was because we had the grammar correct.

The operator put his hand over the pen before I could pick it up. The gesture embarrassed both of us. He had invited me as an investigator, and now he was trying to prevent me from participating in his investigation. Yet I did not push his hand away.

In the pause a telephone rang in the adjoining room. It was an old mechanical bell, the kind that strikes actual metal. I could hear it through the wall as distinctly as a spoon tapping the rim of a glass.

The woman said, “That is for you.”

I had left my phone switched off in my coat. She did not move toward the adjoining room. The bell continued to ring, and the operator, who had turned very pale, asked how she could know whose call it was.

“Because you have not made it yet,” she said.

Then she led us into the adjoining room.

It contained a narrow cabinet, a chair, and the ringing telephone. Beside the instrument lay a second receiver, disconnected, its cord cut short and tied with blue thread. I lifted the ringing receiver before I could consider the wisdom of doing so.

I heard myself asking for the operator by name. The voice was unmistakably mine, impatient in exactly the way I become impatient when a call does not connect. Behind my own voice I heard a railway announcement and the hollow echo of a station concourse. Then, before I could speak, I heard a third voice, too faint to identify, say, “Tell him not to answer.”

I hung up.

The bell rang once more, although the receiver was down.

Outside, a tram passed the window toward the canal. The passengers were turning their heads in the same direction, toward something at the back of the carriage. For one impossible second I had the conviction that every passenger was watching the same person walk toward them along the aisle, slowly, against the motion of the tram.

When I turned back, the woman had removed the glass. The coin lay on the other side of the table. Nobody had touched it.

We left without signing the book. She did not appear disappointed. At the door she gave the operator a little square of folded paper and said that when the messenger returned, he should give it back to the first person who asked for change.

The paper held no writing. We unfolded it twice. It was blank on both sides, except for an indentation that might have been made by a pen long before the ink ran dry.

For weeks afterward I tried to reduce the visit to a sequence of events. I drew the room, counted the pigeonholes, copied the tram routes, wrote down everything the woman had said. That procedure, which had served me well all my life, became a peculiar form of homesickness. I could catalogue her objects, but I could not return to the hour in which I had believed each object belonged to another world.

I began reading the old books she had named. I learned why Mercury wears wings, why his staff bears two serpents, why merchants and messengers once made offerings at crossroads. I learned the signs for conjunction and opposition, and the houses in which an astrologer looks for journeys, theft, marriage, debts, death. These did not feel to me like explanations. They were rooms opening from rooms, with no outer wall.

One evening I waited in a car park to see Mercury after sunset. It was low and difficult to find, a small light in the last colour of the sky. A plane crossed the horizon. Somebody laughed behind me, and I turned too quickly. When I looked back, the light was gone.

I bought an almanac. I moved a planned migration away from the next retrograde. I did not put that reason in the schedule.

Months later I found the same silver coin in the pocket of a coat I had not worn to the astrologer's house. The worn number on it was a little clearer now. I could almost read it.

I have never decided whether I hope to meet the woman again.

Field Note #4. The closer your model fits the world, the more the world will take issue.

VI. The Crocodile in Vienna

The photograph showed a crocodile on the stone embankment of the Donaukanal. Behind it were the ordinary details of Vienna: a bicycle railing, a cigarette packet, the edge of a café terrace. The animal lay with its mouth slightly open, as if it had arrived in the city too early for an appointment.

The police closed the path. Cyclists were redirected through narrow streets, children were taken away from the water, and a woman interviewed on the radio said that her dog had refused to approach the bridge. The zoo denied that anything was missing. For two days the city lived with the improbable addition to its geography.

I began keeping a record when the first report reached me. At noon I checked the news from Vienna. Then I opened the monitoring page for a service we operated in America. Everything was green.

That evening the animal had moved downstream. The police were using boats. I checked the status page again.

The next morning a picture showed its tail slipping into the water beneath a footbridge. People watched from behind barriers; someone had brought a folding chair. I returned to the American dashboard before I finished reading the caption. There had been no interruption overnight.

I made a column in my notebook. Vienna on one side; the other events on the other. It was difficult to fill the second column, so I put down the hours at which I had checked. By Thursday it contained more entries than the account of the crocodile.

At work, somebody noticed the news article open on my screen. He asked whether I had ever been to Vienna. I said yes. He asked whether I knew the canal.

“Not well,” I said, and closed the article.

The crocodile was eventually caught. The canal reopened, the cafés set their tables out again, and the public notices came down. None of our systems had suffered an incident. I made a final entry to that effect, with the time.

On Friday night I printed the first photograph. I looked carefully at the railing and the water behind the animal, searching for something I might have overlooked. There was nothing. I folded the sheet, put it in the dossier, and opened the dashboard once more.

All green.

I left it open while I went to bed.

Field Note #5. There was nothing to report. I kept reporting it.

VII. The Heat Death of Rack C

A research institute, one hot week in August.

The room was hot when I arrived. Rack C stood dark at the end of the row. A cooling filter was clogged, a fan had been failing for weeks, and the temperature log showed the steady rise that ended in an emergency shutdown. There was nothing unusual in the sequence.

The database could be restored, but its latest transactions were missing. A backup warning had gone to an unattended inbox. Ruth, a postdoctoral researcher, had lost the last run of a simulation that had taken three days. She listened while I explained the gap, then told me she would join the cluster queue again. That would take another week.

We changed the filter, replaced the fan, and checked the restored data. Every failure had a cause. Every cause had left a trace. The case was as complete as I had ever wanted a case to be.

That evening I wrote ambient temperature in the incident log and closed it.

I opened it again before going home.

There was nothing to add. I checked the temperature curve, then the warning time, then the curve once more. The explanation did not change. Nor did the peculiar feeling that I ought not to leave the room unattended.

Ruth passed me in the corridor carrying her laptop and a paper cup of coffee. I asked if she needed anything. She said no, and went back to her desk.

I stood by Rack C until its new fan had completed another ordinary, faultless hour.

Field Note #6. All causes accounted for.

Coda

The room is quiet at this hour, by which I mean only that the fans are running properly. I can hear a change in their pitch without looking up. I write the temperature of Rack C in the margin of my notebook; it is within specification.

Beside the notebook lies the dossier. Its first pages are orderly. They have dates, witnesses, causes proposed and rejected, little diagrams in ink. Near the front is the report of the valve that opened while a fitter was working beside it. His name is there, and the time, and an empty space where the command should have been. As the pages turn, the handwriting grows less even. Certain sentences have been written over so many times that the paper has thinned.

I keep a photograph of the moth taped into the logbook of an early computer, because I once found it amusing that a bug could be caught and made to stay. Next to it is the page on which I crossed out a sentence about a program that hated to be watched. I can still see the erased letters. Across the page a box is marked 04:56. I have kept it empty, though I know exactly what I was doing at that hour.

There is a copied hash written on a folded scrap of paper. The files it once named are gone. When I unfold it, I remember not the number but the developer's description of a demon studying two small objects in its hands. I fold the paper again before reaching the end of the line.

Another page contains a witness's answers in my own handwriting. He remembered a visitor and did not want to describe him. He had been laughed at, and each time I read the notes I remember how carefully he chose his words. I have stopped trying to supply the words he withheld.

A coin lies beside the almanac. I have not shown it to anyone. Next month's date is ringed, and I have moved a routine change to the following week. The change is safer for several ordinary reasons, which I wrote down afterward.

The filter for Rack C will be changed on the nineteenth. I have checked that the work order exists. I have checked that somebody has accepted it. I have set an ordinary reminder, in the ordinary way, and I have set another that I will not mention.

On the last page I copy Ruth's name, though the crash that cost her a week has been explained completely. Then I put down my pen and listen to the fans. A small change passes through their sound and is gone.

I look at the clock. I do not look at the almanac.

In the morning I will be able to explain why.

Author's Notes & Sources

These notes are outside the narrator's dossier. The historical incidents and technical reports below are real; the characters, dialogue, occult events, and their connections in the stories are fictional unless explicitly identified otherwise. A reference to a historical precedent is not evidence that the fictional episode occurred.

1. The Mark II moth. The original 1947 logbook, with an insect taped into it, is held by the Smithsonian National Museum of American History. (Museum collection)

2. The Schaerbeek election. The 4,096-vote discrepancy in the Belgian election of May 2003 was investigated by an official expert panel. The technical anomaly and report are documented; this story's café conversation, witnesses' private words, and newspaper story are invented. (Expert report) (Historical overview)

3. Belgian e-voting context. Oversight and advocacy material supplies additional background on the official response and controversy. (PourEVA report)

4. The Heisenbug. The lonely night and the narrator's experiences are fictional. The observation-sensitive software behavior is inspired by a real ProxySQL hang that disappeared under tracing and other interventions. That real issue was eventually investigated and fixed; the fictional episode does not claim otherwise. (Original issue) (Engineering account) (Fix)

5. The valve. The prologue's uncommanded actuation is invented, drawing on the real class of industrial-control failures in which a physical action lacks a matching command in the available log.

6. Maxwell's Demon. The visitation, hash collision, and vanished originals are fictional. No full SHA-256 collision is known to have been publicly demonstrated. The distinct-file collision actually demonstrated by the SHAttered project used SHA-1, a different hash algorithm. (SHAttered demonstration)

7. The visitor in the cold aisle. This case is invented. Its folklore, including the caution about watching a fairy being, draws on the Irish traditions collected in W. B. Yeats's Fairy and Folk Tales of the Irish Peasantry; neither the narrator nor the operator is a historical witness. (Yeats collection)

8. Mercury in Retrograde. The clearing house, astrologer, moving coin, phone call, and almanac are fictional. Mercury's associations with messages, journeys, commerce and interpretation belong to longstanding astrological traditions; William Lilly's Christian Astrology (1647) documents one early modern version of the larger symbolic practice. This source is evidence of a tradition, not of the invented events. (Lilly, Christian Astrology)

9. Crocodiles in Vienna. The incident in the story is invented. An actual small caiman was caught in Vienna's Donaukanal in 2001; the historical capture is a precedent, not the episode described by the narrator. (BBC) (ORF Wien)

10. Heat and server failure. Rack C, Ruth, and the research institute are invented. Cooling-system failures causing real cloud-service disruptions are documented, including an incident in London in July 2022. (Google Cloud incident report) (ASHRAE thermal guidelines)