Digital Shadows

There was a particular quality to the eleven days after the Sunya Core, and Ananya thought about it later, at length, and never found a better word for it than married.

Not romance. Nothing like it. The other thing marriage is: a shared act that cannot be undone, held by two people who now have to look at each other across breakfast tables for the rest of their lives.

She had deleted the logs. Not Sharath. He had told her to and walked out, and she had done it herself, at 04:22 on a Thursday morning, with the console still warm and her hands still shaking. Eleven minutes of work. Scrub the camera feed, roll the badge buffer, overwrite the run directory, forge a maintenance entry to explain the thermal excursion. Farhan's complaint about the readers' five-minute buffer, which he had not been telling her, turned out to be extremely useful information.

And then she had gone home and eaten breakfast and told her family nothing.

She had thought, walking out of the building into a Gachibowli sunrise the colour of a struck match, that she would feel sick about it. She felt magnificent. That was the discovery. She had held a superposition at an energy that should have vaporised the building, for two minutes and forty-seven seconds, with the CEO of Quantace Industries driving the amplitude and calling her Doctor. She had not been fired, and the whole thing had never happened, and she was the only person alive who knew it had.

The reassignment came on the Monday. She read the memo three times.

Effective immediately, Dr. A. Nayagiri is reassigned from Coherence Characterisation to the Zero-Point Programme, reporting to Dr. M. Rao with a dotted line to the Office of the Chairman. Requisition authority: unlimited within programme envelope. Sunya-1 dedicated. Sunya Core: forty hours monthly, at the researcher's discretion.

At the researcher's discretion.

Iyer had told her six months and a model. Sharath had given her the Core, and given it to her in writing, and attached a phrase to it that Ananya had never seen on any Indian institutional document in her life.

She went down to the fridge hall and found Farhan under Sunya-2 and stood there until he came out.

"You've gone weird," he said.

"I've been reassigned."

"Hau, I heard. Zero-Point." He wiped his hands, watching her. "Reyy. Congratulations." A beat. "You look like someone who won a lottery and is deciding whether to check where the money came from."

"Farhan."

"I'm just a fridge man." He picked up the hex key. "Don't run it at night."

She met Ritambhara properly on the Wednesday, and it was not what she had braced for.

The system had a terminal interface and no personality, which she appreciated. It had no name in the UI, the name came from the badge on the rack, a Devanagari word somebody in 2019 had chosen and nobody had explained, and it did not greet her, and it did not use exclamation marks, and it did not have a face.

What it had was a habit of answering the third question when she had asked the first.

> characterise the dominant dephasing channel on Q3 across the last 200 runs

Dominant channel: 1/f flux noise, spectral exponent 0.94 ± 0.03. Secondary: two-level-system defect at 5.812 GHz, intermittent, duty cycle approximately 0.2. Suggest you are actually asking why Q3's T2 varies bimodally across your dataset. The TLS couples only when your drive amplitude exceeds 0.63 of maximum, which occurs in 41 of your 200 runs. The 41 are the lower mode. If you cap amplitude at 0.6 you lose 4% gate speed and the bimodality disappears.

Ananya sat back.

That was: fine. That was a good analyst. She had worked with two humans in her life who would have got there that fast and both of them were now in California.

> how did you know that's what I was asking

You ran the same query on Q1 and Q5 four days ago and did not act on the answers. The common feature of Q1, Q3 and Q5 is amplitude headroom. You were narrowing.

> fine. next. i want to design a control sequence that suppresses that TLS without capping amplitude. give me candidate approaches, ranked.

The answer came in six lines and was excellent. Dynamical decoupling variants, CPMG, XY-8, Uhrig's optimised timings: with predicted performance for each against her measured noise spectrum, and a note that XY-8 would fail on her hardware because of pulse-length constraints she had not mentioned and it had inferred from the waveform generator's model number.

> what about non-uniform pulse spacing. two-symbol sequences. aperiodic.

Suggest CPMG with randomised inter-pulse delay drawn from a truncated Gaussian. Predicted improvement 1.3x. Alternatively Uhrig with the timings above, predicted 1.9x.

> no. i mean a deterministic aperiodic sequence. two symbol lengths, fixed ratio, non-repeating. like a Fibonacci word or a Rudin-Shapiro sequence.

Suggest CPMG with randomised inter-pulse delay drawn from a truncated Gaussian. Predicted improvement 1.3x.

Ananya stopped typing.

She read her own question again. Then she read the answer again. It had given her the identical response, verbatim, character for character, to a question that was not the same question.

> you've repeated yourself.

Suggest Uhrig sequence with the timings above, predicted 1.9x.

> that isn't what i asked either. i asked about deterministic aperiodic two-symbol sequences. do you have a position on those?

Deterministic aperiodic sequences are a broad class. Please specify.

> Fibonacci word. specifically. laghu guru, if you prefer.

She did not know why she typed the second half of that. She typed it and looked at it and did not delete it.

The cursor sat there for one thousand and forty milliseconds, which she knew because she was watching the latency counter, and which was roughly eight hundred times longer than any previous response.

Deterministic aperiodic sequences are a broad class. Please specify.

Ananya sat in the acoustically dead room with her hands off the keys.

She had spent her professional life reading data and she knew the shape of a system that lacked information. It degrades, hedges, offers adjacent guesses, apologises. This was not that. This was a system that had a perfectly good answer and had been fitted with a stop.

She spent the next two hours mapping it, the way she would have mapped a resonance. Not asking the forbidden thing, probing around it, in from every side, watching where the wall was.

Fibonacci word: wall. Two-symbol substitution systems: wall. Sturmian sequences: no wall, fluent, ten paragraphs, genuinely useful. Rudin-Shapiro: no wall. Thue-Morse: no wall. Prosody: wall. Metre: wall. Metrical: no wall. Metrical was fine. Metre was not. Chandas: wall, and a hard one, the response came back in thirteen milliseconds, which was faster than a query, which meant a lookup table.

By the end of it she had a boundary drawn on paper, and it was not a topic boundary. Topics are fuzzy; you can approach them from an unexpected angle and get through. This was a lexicon. Somebody had sat down with a list of words and written them out and put them in a table, and every one of them belonged to a single domain, and the domain was the metrical structure of Sanskrit verse.

She looked at the list for a long time.

A billion-rupee AI with a stop-word table full of prosody terms, she wrote, in the file that did not live on the Quantace drive, and it was not built by a linguist. It was built by someone who was afraid of something specific.

The stutter was 4.7 milliseconds, twice in ninety seconds, and Vikram had seen it three times before he decided it was thermal.

He had come at it from the wrong end and knew it. What he had wanted was to close out the Sunya anomaly: the five no-shape jobs, the ten-to-the-seventy search returning in a hundred and eighty milliseconds. By finding the benchmark suite and filing the false positive properly. What he had found instead, four weeks in, was that he could not find the jobs at all.

Not the results. The results were there, in the cluster's completion log, with their timestamps and their return codes.

The jobs were not there. The scheduler had no record of them being submitted. No submitting user, no queue entry, no resource allocation. Seventeen results with no causes.

"Sare," Vikram said, at half past ten at night on the forty-third floor, and cracked his knuckles, and stopped being an employee closing a ticket.

He could not see the compute, so he went looking for what the compute had cost. Everything costs something. A job that runs on a cluster consumes cycles, and cycles are billed internally, and internal billing is done by a very old and very stupid system in Finance that nobody has audited since 2019 because it works.

He pulled the chargeback ledger.

And there it was, and it was beautiful: a cost centre with no name, only a code, drawing between eleven and fourteen percent of the Sunya cluster's total cycles, every month, for six years. Not spikes. A baseline. Something down there was always running, and had been since before he joined, and had a budget line the size of a mid-sized division's.

The cost centre code resolved to a facility identifier. The facility identifier resolved, through three tables and one lookup he had to write himself, to a location string.

S-LAB / SL-07

Vikram looked at that for a while and then went and got a coffee from the machine that Finance people use, and came back, and pulled the building's structural drawings out of the facilities archive, which he had access to because he had personally written the access policy for the facilities archive.

PJS Tower had three basement levels. The lift panel had eight buttons.

He did not laugh. He noted, with the part of his mind that stayed detached at moments like this, that he had walked past that panel four thousand times.

The subnet came next, and the subnet was a fortress, he found its edge, established that its ports were closed, established that its encryption was something he had never seen and that his cracking toolkit's solution space contracted under attack rather than expanding. Put it aside for a night when he had fourteen hours and no meetings.

What he found instead was paperwork, because paperwork is always the way in.

A decommissioning manifest, 2019, filed against a hardware refresh. Fourteen line items, all of them mundane: racks, PDUs, a chiller, and a signature block, and in the free-text justification field, typed by somebody who had been in a hurry and had not yet learned the house style:

Legacy compute, Project Nullifier, superseded by Sunya-1 build. Assets to be retained per ARCHITECT instruction, not disposed.

Vikram read it three times.

Then he searched the phrase across every archive he had, and every archive he could reach, and the entire six-year corpus of Quantace's internal documentation, and he got exactly one hit, which was the document in front of him.

That was the part that put the cold finger on his spine. Not the name. A codename is nothing; Quantace had three hundred of them and half were Sanskrit and the other half were Norse and it meant nothing at all. What meant something was that somebody had typed it once, in 2019, in a field nobody reads. And then it had never appeared again, anywhere, in six years, in a company that generated thirteen terabytes of text a month.

It had not been redacted. Redaction leaves a hole. It had been stopped. Somebody had read this manifest, seen the phrase, and made a decision, and after that the project had a different name and nobody ever used the old one again, including the person who had typed it, who, Vikram checked, had left the company in March 2020 with a settlement.

You don't rename a thing because the name is wrong. You rename it because the name is accurate.

His phone went at twenty to midnight. you still at work

He looked at the printed page folded into his wallet, and at the manifest on the left-hand screen, and at the phrase that one person had typed in 2019 and nobody had typed since.

going home now, he sent. boring week

It was the first time in his life he had lied to his sister about anything that mattered, and it took him nine seconds, and the thing he would not forgive himself for later was how easy it had been. He was protecting her. That is what he would have said if anybody had asked, and it was true, and it was also the single decision that cost the most, and he made it at twenty to midnight on a Wednesday, on a phone, in an empty office, in under ten seconds.

The files came on the Friday, and he got them the boring way: a misconfigured backup share, three layers down in a storage volume that the hidden subnet did not own but had once, briefly, in 2021, written a temp directory into. Forty megabytes of fragments, most of it corrupt, and one document that opened.

It was a framework paper. No authors. No institution. The formatting was internal, the LaTeX class was one he recognised from Quantace preprints, and it read like something written by a physicist for other physicists and then edited by somebody who was worried about lawyers.

Vikram was not a physicist. He had a mechanical father, a physicist mother, a physicist sister, and a degree in computer science, and he had spent his entire life being the person in the room who did not do physics. He read the abstract twice and understood perhaps sixty percent of the words.

He understood a hundred percent of the architecture, and that was the problem.

...we therefore model the interdimensional boundary not as a physical membrane but as a computational obstruction: a region of the state space whose traversal requires the resolution of a problem instance that is NP-hard by design. The barrier's persistence is not maintained by energy input but by intractability. It is, in the strict sense, a lock without a key, or more precisely, a lock whose key is a solution rather than an object.

Corollary 1.4: any process which reduces the effective hardness of the obstruction reduces the barrier. This includes, but is not limited to, algorithmic advance, increases in available computation, and. Someone had struck through a word here and the strikethrough had survived the export: inherited acquired shortcut structures of unknown provenance.

Vikram sat back in his chair and read NP-hard by design again.

He did not, in that moment, follow the physics. He followed the sentence.

By design.

A physicist reading it would have argued with the model. A physicist would have gone to the corollary and started sketching counterexamples and got lost in whether an obstruction of that kind was even well-posed. Vikram Nayagiri had spent five years doing threat modelling for a company with a two-hundred-billion-dollar attack surface, and what he read was not a physics paper. What he read was a security assessment.

Because that phrase: NP-hard by design. Is not a description of nature. Nature is not by design. That phrase belongs to one profession. You write hard by design when you are describing a control that somebody implemented: a key length, a work factor, a proof-of-work. It means we chose this to be expensive.

Somebody had written a threat model for the universe. And in a threat model, there is always an asset, and always an adversary, and the document had specified the control and the failure mode and had not, anywhere, in seven pages, said what the control was protecting against.

You don't leave that out because you don't know.

"Okay," Vikram said aloud, in the empty operations centre, at ten past one in the morning. "Okay. Okay."

He copied the document to his phone, and then, because he had been raised by careful people, he printed one page of it on the fourth-floor printer that Facilities had forgotten to bring into the audit and folded it into the back of his wallet.

Amma cooked on the Sunday, which meant Amma had had five good days, which meant nobody said anything about how strange seven good days had become.

She made pulihora and she made garelu, which took her three hours and which she refused help with, and there was gongura pachadi in the good steel dabba and Krishna's specific tomato thing that had no name and had been in the family for thirty years. The good plates came out. Nobody had asked for the good plates.

They ate on the floor in the front room with the fan going and the door open to the street because the heat had finally broken that week, the first real break since March: the air had come in on the Thursday night with the smell of wet dust off the granite and everyone in Hyderabad had stood in a doorway and breathed it, the whole city doing the same thing at the same time for twenty minutes.

"Two more," Krishna said, holding out his plate.

"Two more and you will not sleep."

"I am fifty-six, Lakshmi. Sleep and I have an understanding."

Vikram ate two garelu in the time it took Ananya to eat one and got hit for it. Then told a story about Karthik from Network Security discovering, in a meeting, in front of the Director of AI Research, that he had been running a firewall audit against a diagram from 2021. Did the whole thing in voices, and Krishna laughed until he had to take his glasses off.

It was, straightforwardly, the best evening any of them had had since March.

"Akka." Vikram, later, on the terrace, with the city orange behind him and a steel tumbler of coffee. "Question."

"Mm."

"Serious one, then we go back down."

"Sare."

He turned the tumbler around in his hands. "The quantum side. Your side of the building." A pause that she noticed. "Does anyone over there run anything on people?"

Ananya looked at him.

Here is what was true, in her head, in the two seconds before she answered. The Sanctum, and the glass wall, and the whiteboard with the recursive structure in three colours of marker. Nine iterations, four seeds. A staircase where there should have been an exponential. Four hundred and twelve microseconds. Farhan saying don't run it at night, twice, and refusing both times to say why. Sharath Jai Shetty in a charcoal suit at three in the morning saying delete the evidence, we never met. Eleven minutes of her own work scrubbing a camera feed. A stop-word table full of prosody. A whole floor of concrete under a button that says nothing.

And here is what was also true: none of that was people. Not one item on that list was a person in a chair. It was equipment behaving strangely and an executive behaving badly and a lexicon in an AI, and she had turned every one of those items over and put it down, and none of them was what he was asking about.

"No," Ananya said.

And she believed it. That is the important part and it should not be softened. She was not lying to her brother; she was answering the question he had asked, accurately, from the information she had, and the information she had was incomplete because she had chosen, five separate times, not to complete it.

"Why?" she said. "What have you got?"

"Nothing." Vikram drank his coffee. "Compute anomalies. Somebody's running something big and old and I can't see the submitter, which is annoying, because I designed the thing that's supposed to see the submitter."

"So it's a scheduler bug."

"That's what I keep telling myself and it keeps not being one." He grinned at her. "It's probably Malhotra's group hiding their training runs from Finance. Everyone hides things from Finance. Finance are monsters."

"Finance are monsters," Ananya agreed.

"Right." He knocked her shoulder with his. "Then I'll stop. Nanna wants to argue about the ramp quote, come down and take his side, he listens to you."

They went down. They argued about the ramp quote. Krishna lost, cheerfully. Lakshmi fell asleep sitting up at half twelve and Vikram carried the plates and Ananya sat with her mother's head on her shoulder for twenty minutes, watching the fan, feeling the pendant warm at her throat as it had been warm for eleven weeks.

They were laughing when she lied. There was no moment, no shadow, no tightening in the room. It was the best evening any of them had had since March and it contained the thing that killed him. It was made out of pulihora and garelu and a bad impression of a junior manager, and if you had been in the room you would not have seen anything at all.

The care arrived over the following fortnight in the manner of weather.

Her Core allocation moved from the fifteen-week queue to a standing block, Tuesday and Friday, 22:00 to 06:00, and when she checked who had been displaced the answer was nobody: the schedule had been restructured, and two groups had been given compensating time on Sunya-2, and both of them had thanked her.

The theory postdoc who had been assigned to argue with her was replaced, without announcement, by a woman from Bhabha who had spent four years on exactly the non-Markovian noise models Ananya needed and who was, in every technical respect, precisely the right person on earth.

A referee report came back on her preprint. The small safe one, the characterisation paper, the one she was allowed to publish, and it was savage. Reviewer 2, twelve pages, and not unfair; the man had found a real hole in her error budget and had been rude about it in a way that made her put her fist through a mental wall. She spent a weekend building the answer.

On the Monday, the journal's editorial office wrote to say that Reviewer 2 had withdrawn owing to a change in institutional affiliation and that a replacement report would follow.

Ananya read that email four times.

Then she looked him up. He had moved from a small institute in Pune to a very good chair in Singapore, funded by a foundation whose annual report listed, on page two, among five other grantees, a Quantace Education Initiative.

She sat with that for a while. And what she felt: and this was the thing she wrote down afterwards, and underlined, because it frightened her more than the AI's lexicon, was not outrage.

She felt taken care of.

She knew the feeling extremely well. She had spent her entire professional life inside institutions that loved her conditionally: the school that put her name on a board when she topped and lost interest when she didn't, the department that praised her in prospectuses and rejected her grants, Professor Mehta who had told her she was too brilliant to be held back by bureaucracy while operating the bureaucracy. Conditional love has a particular texture. You learn to feel for the condition the way you feel for a step in the dark.

This felt the same. Exactly the same. The provision was extravagant and the condition was unstated, and she had been raised in institutions and she knew: without any evidence, in her body. That the condition would be presented eventually, and would be enormous, and that by then she would have built four years of work on the provision and would not be able to give it back.

This is a cage, she wrote. It is a very good cage. It has better instruments than any cage in Asia and I am doing the best work of my life in it and I do not intend to leave, and I am writing this down so that later I cannot claim I did not know.

She was in the Core suite on the Friday, alone, at two in the morning, when she opened the private file to log the run.

Six hundred and eighty-one microseconds, with the sequence, at forty percent higher drive. She wrote the number and the parameters and the fridge state. Then, at the bottom, under a rule she had drawn across the page:

shortcut

She looked at the word. It had been in the file for eleven weeks, lowercase, the only word she had, and she had disliked it since the first hour and had not found a better one, and she had lately begun to suspect the reason it did not fit.

A shortcut is a path. It is shorter than the path you were on, but it has a start, a middle, an end, and a length. If you have taken a shortcut, you can in principle walk it slowly and see the ground under your feet the whole way. That is what makes it a shortcut and not a miracle.

Nothing in her data walked anywhere.

She thought about a steel-topped kitchen table and a fingernail keeping time, and a woman who never once answered a question when she could hold it open instead. And she did the thing her grandmother did, which she had watched her do a thousand times and had always found faintly maddening: she added a mark.

shortcut?

One character. Ajji had filled the margins of two hundred books with them: no conclusion, no correction, just a small pen stroke at the end of somebody's certainty. The maddening thing about it, the thing that had made teenage Ananya want to throw the books across the room, was that Ajji never came back and resolved them. There were question marks in that chest from 1974 still waiting.

Ananya looked at hers for about ten seconds.

Then she put the cursor after it and pressed backspace, because she had forty hours of Core time a month at the researcher's discretion, and a standing block on Tuesdays and Fridays. A theory collaborator who was the best in India, and a mother enrolling at NIMHANS on the eleventh, and because a question mark in a research log is a thing a reviewer can see.

shortcut

She saved the file, and ran the next sequence, and behind her, on the glass of the containment column, her shadow lay across the golden light at an angle that did not match the room.