FREQUENCY

"A thing that is hidden leaves a shape where it is hidden. The shape is not the thing. The shape is the part of the hidden thing that cannot be hidden, because hiding itself has a weight, and the weight bends the things around it in the manner of all weights."

By the second Friday Maya had stopped pretending she was going to do the proper tests tomorrow. Tomorrow had been Wednesday twice, and Thursday once, and a Saturday afternoon she had told no one about except Gerald. She'd been running tests at her kitchen table for almost two weeks, in the hours when work stopped interrupting her ability to think about work. The tools she had were telling her what she already knew and nothing more. The tools weren't the problem. The thing she was looking at was the problem. The thing she was looking at didn't fit any of the shapes her tools had been built to name.

So she had gone slower.

She'd pulled four more batches across those two weeks, on a personal VPN she had set up in her third month at Lumen after reading a specific sentence in the employment handbook about remote access auditing and deciding that remote access auditing was the kind of thing an engineer with nothing to hide might still want to hide. The VPN had been there for more than a year because Maya had wanted it to exist. She hadn't expected, at the time, ever to use it for what she was using it for now. She was using it to pull training batches she had already been cleared to review, which wasn't, strictly, a violation of anything. It was, loosely, a small private map of what she had started to think of as the shape of the problem. The shape was the same in every batch. The pattern didn't migrate. Whoever was planting it had chosen a place and was committed to the place.

She'd sat through two of Petra's Thursday reports in the last two weeks without flinching, which had been the hardest part. The second one had been worse than the first. Petra had stood at the front of the conference room explaining the training-pipeline status for the next two model iterations, and Petra's hand had been holding a pen she wasn't using, and Maya had watched the pen-without-writing and had silently catalogued the fact that the tic was getting more frequent, not less, and Maya had said nothing, had asked no questions, had produced a small engineered facial expression that meant I am interested and I am tracking and had held that expression for forty minutes without her face understanding what it was doing. Afterward, in the hallway, Phoebe had said you okay, you seem a little zoned, and Maya had said batch day tired, and Phoebe had said I get it and hadn't asked again, because Phoebe had learned by now that Maya's zoning had a quality to it that meant the asking would make things worse. Maya had loved Phoebe for exactly four seconds after that and had placed the four seconds where she placed the things she loved and didn't have time for.

Now it was the second Friday. Eleven-twenty in the morning. Maya had called in a migraine she didn't have and was at her kitchen table with her personal laptop and a mug of the coffee she made at home on the weekends, which was better than the coffee at work but took twice as long because she ground the beans by hand. Her hair was tied back loosely in the way she only tied it back when no meeting was going to ask her to make decisions about how the room read her. The grinding had been on purpose. The grinding had given her forty-five seconds in which she wasn't looking at a screen, and forty-five seconds was long enough for a person to decide whether she was about to do something she couldn't un-do.

She opened the embedding visualization tool. The anomaly was there. It had been there on Wednesday and on Thursday and she had checked it so many times it had become a fact she wasn't going to be able to not know, like a crack in a ceiling becomes a thing you see every time you look up.

She opened a database query tool and searched the published literature for the mathematical signature of what she was looking at. The query ran for eleven seconds and returned a list of forty-three papers. She narrowed by keyword. Fourteen papers. She narrowed by citation pattern, screening for anything cited more than twice but fewer than thirty times in the last decade, which was the citation profile of real work that nobody had noticed. Seven papers. She narrowed by author affiliation to rule out industry-funded research, because industry-funded research was always a wall dressed up as a door. Three papers.

Two were about a harmonic problem in speech synthesis. Not what she was looking for.

The third was from 2003, published in a journal she had never heard of, by an author she had never heard of, titled Recursive Harmonic Sequences and the Modeling of Limbic Response at the Semantic Layer, written by Dr. James Okafor, then of Rawlings University.

Maya read the abstract.

We present a theoretical framework for modeling cumulative affective response in human readers exposed to natural-language outputs of generative computational systems. We propose that the structural rather than semantic features of generated language carry the dominant load in producing such response, and that these structural features can be characterized as recursive harmonic sequences embedded at specific dimensional intervals within the system's internal representation space. We define the relevant intervals as the below-threshold range in which structural features remain undetected by content-aware analysis but produce measurable cumulative shifts in limbic response among exposed readers across populations of sufficient scale. We outline the conditions under which a generative system might be configured, accidentally or by design, to produce such sequences in production output. The implications, both for the design of safe affective-language systems and for the detection of hostile applications, are discussed in §4. Appendix B (unpublished, available from the author by request) describes the population-level consequences predicted by the framework under conditions of sustained exposure.

She read it again.

The methodology section was twelve pages long. She read it slowly. The most important paragraph was on page nine, in the second column, in the small careful prose of a man who had been writing the paragraph in his head for years before he wrote it on paper.

Let H denote the space of structural features that can be encoded in the internal representations of a generative language model trained on a sufficiently large corpus of natural language. We will be concerned with a subset of H, denoted H\, consisting of those features which (a) lie below the threshold of detection by content-level analysis, and (b) admit recursive embedding across multiple representation layers. We argue that H\* is non-trivial, that is, non-empty, for any generative system whose representational capacity exceeds a specific minimum that has been crossed by all production-class systems built since approximately 1996. The features of H\* are not artifacts of training. They are mathematical possibilities that any sufficiently expressive system will be able to encode, whether or not the system's designers intend to encode them.*

The paragraph below the paragraph was the one Maya read three times.

The question we pose in this paper is: given that H\ exists, and given that any sufficiently expressive system can be configured to populate it, what is the predicted limbic response of a population of readers exposed to language generated by such a system across a period of months to years? We do not, in this paper, attempt to answer this question empirically. We provide a theoretical answer, derived from established work on subliminal affective priming and from the harmonic analysis of natural language rhythms developed by Pollard (1987) and Kuznetsova (1991). The answer is in §4. The answer should not be reassuring.*

The paper was theoretical. The author hadn't built the thing he described. The author had reasoned about what it would do if it existed.

It existed now. Maya was looking at it on her personal laptop in her kitchen on a Friday morning in February.

She set the mug down. Her hand wasn't steady enough to keep holding it. She was going to need to sit very still for a while.

She counted her own pulse at her wrist. Faster than she liked. The heat in her hand was specific. She placed the hand on the cool of the table. She was not afraid. She was something with the same shape as afraid that was taking place forty seconds further into the future than afraid usually takes place. She didn't have a word for it. Her mother would have had a word for it. Her mother had a word for most of the states Maya had learned to live inside.

Calyx Systems was in a stucco building on a frontage road off El Camino. Nora Voss stood in the lobby on a Monday morning in June of 1978 with a canvas bag on her shoulder and an advisor's letter in her hand, and the word that described what she felt was inside. Inside the building, inside the work that mattered, inside the kind of place she had decided at fifteen she was going to spend the rest of her life trying to be allowed into. The feeling was clean and uncomplicated.

The man who came to meet her introduced himself as Hal Brenner. He was in his mid-thirties, friendly in the manner of people trained to be friendly with interns. He shook her hand too firmly and apologized for the firmness and walked her through the main lab, which contained, behind a half-glass partition, a minicomputer the size of a refrigerator humming in a way that made Nora aware of her own pulse.

"PDP-10," Hal said. "We share it with a machine at Stanford. You get used to the hum."

She was going to get used to the hum. She didn't know that yet. It wouldn't have changed anything if she had. Nora at twenty had a specific relationship with information she didn't yet have: she wanted it, she chased it, she went toward the thing she didn't know until the thing she didn't know either became the thing she knew or became a thing she could name as an absence. She didn't yet practice patience with her own curiosity.

Hal introduced her to Diane Krol, the other woman on the technical staff. The only other woman. Diane looked up from a terminal, nodded once, and went back to her work. Nora would try harder with her later. She could tell that Diane was worth trying with. She met a senior engineer named Dennis Kaplan, who mispronounced Voss as Vohss, and Marjorie Vance, the admin, who had Voss, V-O-S-S written on a form before Nora had said her own name out loud.

She was shown to a desk in the main lab room and given a three-ring binder she recognized as the kind of binder that contained nothing useful. She opened it anyway because she was twenty and wanted to be seen reading things.

Halfway through the morning, she noticed the door.

It was down a hallway off the main lab, and it had a badge reader installed on it, which wasn't the kind of thing that existed on most research-lab doors in 1978. Most research-lab doors in 1978 used keys on hooks. The badge reader wasn't the thing she noticed first. What she noticed first was that the tour had gone around it, not past it, and that the going-around had been so smooth she had not registered it as going-around until she was already sitting at her desk ten minutes later and the route through the building replayed itself in her head and she understood that she'd been walked in a specific pattern designed to keep her oriented away from the door she had now noticed.

She planted the door the way she had started planting things that interested her. The planting, at twenty, was still a conscious act. She didn't yet have a word for why.

At noon she ate lunch at her desk from a paper bag. Diane came over and said something brief about the PDP's job queue, and Nora answered with the exact correctness that Diane was testing her for, and the conversation went well enough that Nora felt the private thrill of having said something correct to a person she wanted to impress. The thrill was a clean young thing. She sat with it for longer than the thrill needed.

At some point in the afternoon, a tall man in his mid-forties crossed the lab with a folder under one arm, in conversation with someone Nora couldn't see. He didn't look up as he passed. A voice at a nearby desk said, quietly, that's Eisen's group, you won't be on that yet, and Nora planted the name the way she had planted the door, without knowing why, and then went back to her binder. The binder was still not useful. She was going to ask Hal about the PDP's backup tape schedule in the morning, because she wanted to know where the data that passed through the machine actually lived, and wanting to know was the kind of question she had learned, at Stanford, was the kind of question that impressed senior people enough that they started telling you things they would not otherwise have told a summer student.

The sun was still up when she walked out to the parking lot at the end of the day. The lot smelled of warm asphalt and eucalyptus. She put her canvas bag on the passenger seat of her Datsun and turned to look at the building one more time before getting in, because turning to look at the building was the kind of thing a person did on her first day at a place that was going to matter, and Nora had decided, on the drive down that morning, that she was going to be the kind of person who marked the moments that were going to matter.

She was about to get in the car when she heard it.

The hum.

It was coming from the building. A steady low note, under the HVAC, under the sound of distant traffic on El Camino. A machine running. A machine that didn't stop when the people went home. Nora stood in the parking lot and listened to it for a count of maybe ten seconds. The building had looked, from the outside, like an insurance office. The hum was the part of the building the insurance-office outside was built to not announce. The building had a thing inside it that didn't sleep. The building had chosen, deliberately, to look like it did not.

She got in the car. She drove home with the windows down. The hum stayed with her for six weeks before she stopped hearing it in rooms that were not the building's rooms, and by then she had stopped being able to not hear it, and by then she was inside.

In her kitchen on a Friday morning in 2019, Maya Voss read Dr. James Okafor's 2003 paper for the third time and then set it aside and went to the building's bibliography, because there was a reference there that she had noticed on the first read-through and had been refusing to think about since.

The bibliography listed an unpublished appendix. Cognitive Harmonics: Appendix B (Unpublished Draft), author's personal manuscript, cited with permission. The appendix wasn't included in the journal version of the paper. The appendix was referenced eleven times in the methodology section. Eleven references to a document that wasn't in the document.

Maya looked up from the screen. Her kitchen was quiet the way an apartment is quiet in the middle of a weekday morning. Refrigerator hum. Water in someone else's pipes. A car passing below the window. Gerald was on the kitchen windowsill. A small peace lily in a terracotta pot the size of a coffee mug. Six dark glossy leaves on slender stems, one of them slightly longer than the others and leaning toward the window. The single white flower spathe that had opened three weeks ago still upright in the center. She'd moved him from the bedroom last Sunday when the bedroom light had gotten too direct for him. His leaves weren't doing the curl. Gerald was having a calm week. Gerald was the only thing in the room that was having a calm week.

She read the paper's author affiliation again. James Okafor, then of Rawlings University. She ran a second search on his name. He was still at Rawlings. A Professor Emeritus, listed as semi-retired, last known publication in 2014. The citation count for his 2003 paper was fourteen. Fourteen citations in sixteen years. The citation profile of a man nobody had paid attention to, which was the profile of the safest possible stranger to contact, because nobody who had written a paper that nobody read was the kind of person the machinery that was now running in Maya's embedding space paid attention to.

She thought, for about three seconds, about the possibility that this reasoning was also wrong. She placed the possibility in the same internal space where she had been putting things she couldn't yet name. Fourteen citations, seventeen years, a journal no one reads: seventy percent that nobody's watching him. She opened her email client.

She wrote a draft. She read it. She deleted three sentences from the middle. She read it again. She deleted two more. She read it again. She cut the opening paragraph and started over. The draft she finally sent was four sentences long.

Subject: About your 2003 paper, I think someone stole it Dr. Okafor, I'm a machine learning engineer working on a production emotional-response model in San Francisco. In the last batch of training data I reviewed I identified a structure that matches the recursive harmonic framework described in your 2003 paper on limbic modeling at the semantic layer, section 4, within the dimensional range you characterized as sub-threshold. I'm writing because Appendix B of that paper is referenced in the methodology but isn't included in the published version, and I would like to ask you about it. If I'm wrong about this I will apologize for wasting your time. I don't think I'm wrong. Maya

She read it one more time. She changed sub-threshold to below-threshold because sub-threshold was technically correct but sounded like a person showing off, and Maya didn't want to sound like a person showing off to a man whose paper had been cited fourteen times in sixteen years. She read it again. She moved the cursor to the Send button. She didn't hover. She hadn't been a hoverer since she was eleven, which was the year her mother had told her that hovering was just a slower way of deciding the same thing you were going to decide anyway.

She clicked.

The email went out. The outgoing notification chimed and disappeared. Maya sat at her kitchen table and listened to her apartment being her apartment, and to Gerald being Gerald on the windowsill, and to the refrigerator running in the way a refrigerator runs in a building where nobody is home, and she understood that she had just done a thing that couldn't be un-done.

She put her phone face down on the table.

She didn't close the laptop. She didn't get up. Her mother had taught her to sit with a thing once it couldn't be taken back, to pay for it by staying with it a while before walking away. So she stayed.

The apartment was quiet. Gerald was on the windowsill. Outside, San Francisco was being San Francisco, the kind of ambient noise that stopped being noise after a while and became the thing you lived inside without noticing.

And underneath all of it, the way she had always been able to hear underneath all of it when she paid attention, was the hum of the building. The HVAC. The elevator shaft two floors up. The machine-sound a building makes when it's full of machines, even when none of the machines are being operated by anyone she could see. The sound her mother had told her to listen for, a long time ago, without ever telling her why, because her mother had never said why about anything that mattered.

Maya listened to the hum for a long time. Then she closed the laptop and went to make lunch.

[End of Chapter 2]