The job title was "Ethical AI Analyst," which was a phrase Ben Chen had typed into a job board without thinking about it, and then accepted the offer without thinking about it either, because when you're thirty-one and your parents own a restaurant in Monterey Park and your student loans are still being paid off (his father had refinanced the Golden Wok in 2021 to cover Ben's remaining balance at Berkeley), you don't get to be picky about job titles.

OmniCorp was a mid-tier tech company in South San Francisco, the kind of place that had enough funding to survive but not enough to become famous. Their flagship product was Omnisight, a behavioral analytics platform that companies licensed to understand their customers better. Ben's department—four people, including his manager David—was tasked with ensuring that Omnisight's algorithms met "ethical standards," which in practice meant running a battery of tests to check for bias and then writing a report that senior management would either file away or ignore, depending on the quarter's revenue targets.

Ben's job was to optimize the user risk scoring module. The module analyzed user data—search history, location patterns, purchase behavior, social connections—and produced a risk score from 1 to 100. Higher scores indicated higher risk of fraudulent activity, noncompliance, or other undesirable behaviors. The scores were used by OmniCorp's enterprise clients to make decisions about lending, hiring, insurance, and a variety of other things that affected real people's lives.

Ben thought of the module as a black box. He wrote code for it. He tested it. He optimized it. He did not think about what happened to the scores after OmniCorp sent them out.

He should have.

---

Sarah Miller's email arrived on a Tuesday at 11:47 PM. Ben was still at his desk—overtime was common, and he didn't mind it much, because the office was quiet and the coffee was free and he could work without his mother asking why he wasn't visiting more often.

The email had no subject line. The body read:

"Mr. Chen, I know you work on the risk scoring module. I got a score of 87 last month and I didn't do anything to deserve it. I was denied a mortgage, I was flagged by two airlines, and my rental application was rejected because the landlord ran an 'enhanced risk check.' I'm a product manager. I have a Stanford CS degree. I have never been late on a payment. I know what your system does. It's not about risk. It's about silencing people. If you're the kind of engineer who cares about what your code does, we need to talk. I'm not asking you to leak anything. I'm asking you to look. Just look at the internal logs. See how the scores are generated. You'll see what I mean. —Sarah Miller"

Ben read it twice. Then he closed the email and went home.

He dreamed about it. Not in any interesting way—there were no vivid images or dramatic scenarios. Just the same dream he had every night: he was walking through a corridor of doors, and each door had a number on it, and he couldn't remember which number was his.

The next morning, he logged into the internal system and pulled the logs for Sarah Miller's user ID. He had permission to do this—it was part of his job to review edge cases. He told himself this was just a routine review.

What he saw took forty-five minutes to process.

Sarah's risk score of 87 was not based on any financial behavior. She had perfect credit. No late payments. No delinquencies. The factors that had driven her score up were: (1) she had searched for "whistleblower protections" three times in the month before her score was generated; (2) she had visited the website of a regulatory agency that had recently investigated OmniCorp; (3) she had a social connection—a LinkedIn contact—who had filed a complaint against one of OmniCorp's enterprise clients; (4) she had worked at a company that had lost a discrimination lawsuit two years prior.

None of these factors had anything to do with financial risk. They were political risk factors. Behavioral compliance factors. The system wasn't predicting fraud. It was predicting dissent.

Ben sat at his desk for a long time. The office was empty. The fluorescent lights hummed. He could close the logs and go get coffee and pretend he had never seen this.

He did not close the logs.

---

He started the notebook in October. A simple Moleskine, black cover, $18 at a bookstore in Daly City. He bought it on a Saturday morning and sat in his apartment in Daly City (one bedroom, shared with a roommate who was rarely home, rent $2,800 a month) and wrote the first entry.

October 3rd. Reviewed Sarah M's case. Score 87. Factors: whistleblower searches, regulatory website visits, social connection to complainant, former employer discrimination case. None are financial risk indicators. System is scoring political behavior, not financial risk. I have verified this against 12 other high-score cases. Pattern is consistent. The risk module is being used to identify and penalize users who engage in protected activity or have associations with protected activity. This is not fraud detection. This is behavioral suppression.

He wrote for twenty minutes. Then he closed the notebook and went to work.

At work, he optimized the risk scoring module. He improved its accuracy by 2.3 percentage points. David praised him in the team meeting on Thursday.

Ben bought a second notebook in November. He wrote 47 entries in the first week. By the end of the month, he had documented 34 cases where Omnisight's risk scores correlated with protected activity: union organizing, whistleblower complaints, regulatory filings, political donations, medical records requests (three cases involved users who had sought information about assault victims' resources; all received scores above 80).

He did not show the notebook to anyone. He did not email it to himself. He kept it in a locked drawer at home, beneath a stack of winter clothes he rarely wore.

He continued working. He continued optimizing. He continued eating takeout from his parents' restaurant on Wednesday nights when his mother texted him to come home.

---

David called him into the office on a Wednesday in January. The office was glass-walled and overlooked the parking lot. David sat behind his desk, which was the kind of desk that cost more than Ben's first car.

"Ben," David said. He used a tone Ben had heard before: the tone of someone about to say something difficult but wanting to frame it as casual. "I wanted to check in. How are things going?"

"Fine."

"Good. Good. The team says you've been doing great work on the risk module. The 2.3 improvement is solid."

"Thanks."

David leaned back. He folded his hands on the desk. He looked at Ben with the patient expression of a man who was about to deliver a message he had received from someone else and hoped the recipient would understand without further discussion.

"Ben, you're a smart guy. One of the smartest on the team. And I know you're curious about what the data means. What the scores are used for. And I'm going to be straight with you: the enterprise clients decide how to use the data. Our job is to build the tool. Whether someone uses a hammer to build a house or break a window isn't really on us."

Ben said nothing.

David continued. "You have stock options vesting in April. That's forty thousand dollars, maybe more if the quarter is good. You're close to something real, Ben. I don't want you to throw that away over something you can't change."

He paused. Ben waited.

"The system works the way it's designed to work. It's not perfect, but it's legal. And it's profitable. And you're part of it now. So. Make a choice."

Ben looked at David. He thought about the forty thousand dollars. He thought about his parents' restaurant, which had survived the pandemic but was struggling with rising costs. He thought about Sarah Miller's score of 87.

"I'll make a choice," he said.

David nodded, satisfied. "That's all I ask."

---

Ben went home that evening and opened the notebook. He wrote:

"January 15th. David called me in. Told me to stop asking questions. Said I have stock options worth forty thousand dollars. Said the system is legal. Said I'm part of it now. I am. I am part of it. I write this not because I think it will change anything. I write it because someone, someday, might need to know that people who built this knew what it did. The notebook is in the locked drawer. Three hundred and twelve pages. I will keep writing."

He closed the notebook. He made dinner from a recipe his mother had sent him. He ate alone at his kitchen table. He checked his phone—no new emails from Sarah. No new messages from anyone.

He went to work the next day. He optimized the risk scoring module. He improved its efficiency by 1.7 percentage points. David nodded at him in the hallway.

Ben went home. He opened the notebook. He wrote another entry.

He would keep doing this. Not heroically. Not dramatically. Just: wake up, go to work, write in the notebook, sleep, repeat. A small, steady act of recording in a world that preferred things to be forgotten.

The notebook sat in the drawer. Three hundred and twelve pages. Maybe no one would ever read it. Maybe it would rot in the drawer while Ben aged and forgot why he had started writing. Maybe the next time his building had a fire, the notebook would burn along with everything else.

But tonight, it was there. And he would write tomorrow.