Zero-Probability Accident Chapter Three | The Version Approved in Her Name
On the morning of the second day after the earthquake, Veer Insurance opened only part of the building.
The elevators were still being inspected, so employees had to enter the temporary office area through the emergency stairwell. Yellow signs indicating completed structural inspections were posted in the corners, and barricade tape had been pulled across several cracked partitions. The model validation group, which normally required an employee badge for entry, now had two security guards and an equipment registration desk.
Xu Anhe handed over her phone and personal flash drive, then logged in to the designated computer with an IT staff member accompanying her.
Her own work computer had been impounded. No one was allowed to touch the desktop, hard drive, or login state; even the mouse remained at the angle it had been left in when everyone evacuated after the earthquake. The company's legal department reminded everyone that this did not mean anyone had been determined to be at fault. It meant that after a major public accident, any change could compromise evidence.
Incident updates continued popping up in the lower-right corner of the temporary monitor. The search and rescue operation at the intermodal center was still underway, and the confirmed casualty count continued to rise. Some personnel from the risk management office had been rescued, while others had been taken to the hospital. The contact who had promised to send the version history remained listed as missing on the roster released the night before.
Xu Anhe did not open the list again.
She moved the cursor to the center of the workspace and forced herself to finish reading the evidence-preservation notice first. The notice required the start time, equipment used, and accompanying personnel to be recorded for every review. No personal notes could be taken out of the designated area. These procedures turned a case that could originally be opened in a few seconds into one that required several rounds of confirmation every time. She did not rush them, because she understood better than anyone that the more desperately one searched for an answer, the more likely one was to leave traces of the operation on that answer.
At 10:12 a.m., the major-incident response lead sent a new document into her review area.
The major-incident response lead pointed to the document title. “The client's official risk summary submitted to the Major Incident Investigation Bureau.”
The document could be read but not downloaded. Xu Anhe looked at the facility name first, then the data date, model version, and approval time. All three matched the case she had processed before the earthquake.
Then she saw the exposure count for passengers during the 90-minute peak period.
59,040.
It was neither the green marker on yesterday's screen nor a vague statement that the conditions had not been met. This time, there was a definite number.
She moved the company's approved calculation tool to the left side of the screen and entered the reference code for the raw data she had saved. The system loaded a read-only copy from the evidence-preservation area and reran the model.
The result was still 72,000.
The IT staff member repeated the procedure once, as required. They did not carry over her temporary settings or read her personal preferences. The result was also 72,000.
Xu Anhe picked up a pen and paper:
72,000 − 59,040 = 12,960. 12,960 ÷ 72,000 = 18%. 60,000 − 59,040 = 960.
The revised result was 18% lower than the original and only 960 people below the mandatory escalation review threshold.
She checked the three lines of calculation again, then turned the paper toward the IT staff member. They did not look only at the final result but reentered every keystroke from the beginning. The numbers appeared on the screen in sequence, with no rounding and no hidden decimal discrepancy. They arrived at exactly the same `12,960`, `18%`, and `960`.
Confirming that the arithmetic was correct only made the problem clearer. If the difference had been just a few people, it might have come from different data update times. What had disappeared now was nearly one-fifth of the original result, and it had landed precisely on the side that did not require escalation review.
She did not write “fabrication” beside “960.”
A figure that landed exactly below the threshold was certainly worth noting, but it could also have resulted from different data-loading times, disaster-related corrections that had not been re-versioned, a synchronization error in the backup system, or a problem with the copy she had saved herself. Suspicion was not a conclusion. It was only a reason to decide what to investigate next.
What could not be avoided was reproducibility.
If the model version, input data, and execution conditions were all the same, the result could not be 72,000 on one side and 59,040 on the other. At least one of the two documents was falsely labeled.
Xu Anhe turned to the legal staff member accompanying her. “Can I look at the client's raw fields one more time?”
“No,” the legal staff member replied. “The bureau has already taken control of the official data. For now, you can only view the summary they approved for release.”
Xu Anhe pointed to the two different results. “I need to know where the discrepancy occurred.”
The legal staff member closed the authorization list. “You can raise that during the interview.”
At 1:00 p.m., investigators from the Norden Major Incident Investigation Bureau arrived at Veer Insurance.
The formal interview room had no one-way mirror and no excessively bright desk lamp like the ones in dramas. Two synchronized recording devices, a statement of rights, and an evidence number for every item submitted sat on the table. Xu Anhe's supervisor and the company attorney sat on one side. Across from them was a single investigator.
She looked to be in her early thirties, with a chin-length bob. The cuffs of her dark jacket bore the marks of having been repeatedly folded up. She confirmed the recording time first, then placed her identification card on the table.
The investigator turned the front of the card toward Xu Anhe. “Su Yingzhen, Criminal and Digital Evidence Division, Major Incident Investigation Bureau.”
She made no small talk, and her first question was not whether Xu Anhe had altered the data.
Su Yingzhen placed each document in a fixed position and read the evidence number aloud to the recording devices before opening it. The plastic evidence bag made a faint rasp against the tabletop, but no one in the room spoke. Xu Anhe looked at her supervisor and the company attorney. Neither avoided her gaze, and neither gave her any signal. From the moment the recording began, every word spoken by everyone in the room would become a record that could be examined again.
Su Yingzhen opened the first interview record. “Please explain why 72,000 and 59,040 cannot both be correct.”
“If they came from different data or different settings, both could be correct,” Xu Anhe said. “The problem is that the documents claim they came from the same data, the same model version, and the same approval.”
Su Yingzhen turned a page in her notebook. “No model terminology.”
Xu Anhe paused. “The same recipe, the same batch of ingredients, and the same scale won't produce nearly one-fifth less without someone changing something. If the finished product weighs differently, we have to confirm whether the ingredients were changed, the steps were altered, or the label was put on the wrong product.”
Su Yingzhen asked, “So you can prove someone changed the data?”
“No,” Xu Anhe replied.
Su Yingzhen continued. “Can you prove 59,040 is false?”
“Not at this point,” Xu Anhe said.
Su Yingzhen tapped the two summaries with her pen. “Then what can you prove?”
Xu Anhe answered, “Under the conditions stated in the document, 59,040 cannot be reproduced. To obtain a different result, at least one of the model, the input, or the version designation must be different.”
Su Yingzhen wrote a line in her notes. “Which one do you think is most likely?”
Xu Anhe shook her head. “I can't choose the answer that best fits my own position before investigating.”
“I asked for your assessment, not your conclusion,” Su Yingzhen said.
Xu Anhe looked at the evidence list on the table. “Start by checking the data load and its version linkage. Before the earthquake, someone at the intermodal center asked for the peak period and passenger classification to be redefined. He promised to send me the complete version history, but it never arrived.”
Su Yingzhen confirmed, “That person is currently missing?”
“Yes,” Xu Anhe replied.
Su Yingzhen asked, “Did you preserve the call?”
Xu Anhe pushed the evidence number for the contact records toward her. “No recording. Only the contact times, work messages, and my follow-up requests afterward.”
Su Yingzhen drew a line through her notes. “So at present, the only evidence of the classification request is your statement.”
“Yes,” Xu Anhe said.
Su Yingzhen's tone held no accusation. She was simply drawing a boundary around the evidence. For the first time, Xu Anhe understood that the investigator was doing something similar to what she did: not immediately choosing the most plausible story, but first marking off the boundaries of what could be proved.
“The local copy you voluntarily handed over yesterday is favorable to you,” Su Yingzhen said. “But it also proves that you accessed the complete data before the accident. You understand that?”
“Yes,” Xu Anhe replied.
Su Yingzhen asked, “Why hand it over anyway?”
Xu Anhe said, “Because submitting only an explanation that benefits me can't prove the data is genuine.”
Su Yingzhen closed the first notebook and removed another printed document from the evidence bag.
She turned the new document toward Xu Anhe. “Next, let's look at the verification record.”
The risk summary for 59,040 appeared at the top of the document. Below it were an employee account, a verification record ID, and an approval time. Xu Anhe saw her employee number first, then 9:46 a.m.
That was when she had submitted her escalation review recommendation.
The ID was hers too.
At the bottom, it said:
Approver: Xu Anhe.
She pulled the document closer and checked it word by word. Her name was not misspelled, nor was this a case of two people sharing the same name. The 59,040 version linked directly to the verification record she had left before the accident.
“I approved the escalation review,” Xu Anhe said.
Su Yingzhen pointed to the verification record ID. “The system record shows that you approved this version.”
Xu Anhe stared at the number on the document. “I've never seen 59,040.”
Su Yingzhen asked, “Has anyone else ever used your account?”
“No,” Xu Anhe replied.
Su Yingzhen continued. “Was the password written on paper, saved in a browser, or stored in a messaging app?”
“No,” Xu Anhe said again.
Su Yingzhen looked at the interview record. “Did you leave the computer before submitting it?”
“No. I was at my desk the entire time before the earthquake,” Xu Anhe said.
Su Yingzhen raised the last possibility. “Is it possible you submitted the approval from the wrong work tab?”
Xu Anhe was about to say it was impossible, but stopped before speaking.
She remembered checking the data period, classification, and version. She also remembered pressing the submit button. But memory could not replace the system record. Since the record pointed to her, she had to accept that her own actions were also within the scope of the investigation.
“I don't believe so,” Xu Anhe said, “but we need to confirm that with the operation logs.”
As she said those words, she felt for the first time that her memory and the two numbers were standing opposed to each other. She clearly remembered the red line, the escalation recommendation, and the moment she pressed submit, yet she also knew that people repeatedly reconstruct fragments after a disaster. She could not ask the investigators to treat the system record as erroneous simply because she sounded certain.
Su Yingzhen nodded once. “There are still other possibilities: someone used the account, the version linkage was wrong, the system malfunctioned after the disaster, or someone manually redirected the process. We won't choose an answer now.”
She put the document back in the evidence bag.
After securing the evidence bag, Su Yingzhen formally informed her: “From this moment on, you are being reclassified from an ordinary witness as a key subject of the investigation. This is not a conviction, but you may not access the case system on your own, contact potential witnesses to discuss their statements, or ask coworkers to obtain data for you.”
The company attorney then read out the internal measures. Xu Anhe's system access and duties were suspended immediately. Company equipment, employee badge records, and all work communications were placed under preservation.
The interview ended close to evening.
Xu Anhe retrieved her personal phone, while her work computer and company communications equipment remained in the evidence room. She walked to the glass door outside the office area and held her employee badge against the reader.
The red light came on.
The door did not open.
A coworker saw her through the glass and instinctively moved toward the door, only to be quietly stopped by a security guard. The coworker stopped where they were, still holding the disaster-case list the two of them had organized together the day before. Not knowing what to say, they could only nod at her.
Xu Anhe nodded too. She did not ask them to retrieve anything for her. From the moment she was listed as a subject of the investigation, familiar coworkers, her desk, and the files had all become evidence she could not contact privately. The more she wanted to prove her innocence, the less she could touch them.
On the other side of the glass, the two investigation monitors were still lit side by side. One preserved the 72,000 she had personally reproduced; the other displayed the 59,040 received by the Major Incident Investigation Bureau.
Only a glass door stood between the two records.
The approver on one of them was clearly listed as Xu Anhe.
Xu Anhe now had to prove more than that the numbers had been altered.
She also had to prove who, exactly, the “Xu Anhe” who approved the record really was.