Zero-Probability Accident Chapter Four | Four Small Changes No One Noticed
On the third day after the earthquake, Xu Anhe walked into the Nuodeng Major Accident Investigation Agency for the first time. Her employee ID and company computer were still impounded. Before entering the digital evidence review room, she was issued another visitor badge that could only be used that day. Security checked her name, required her to lock her phone in a locker, and then Su Yingzhen accompanied her inside. There were no paper files stacked to the ceiling in the review room, and no one was busy cracking passwords. A long table stood in the center, with four screens arranged side by side on the wall. Digital forensics personnel had turned the files provided by Veer Insurance, the intermodal center, and external data service providers into read-only copies. Every time a file was opened, compared, or exported, the operator and timestamp were recorded automatically. “Today, we’re only dealing with how the numbers changed,” Su Yingzhen said. “Not who’s guilty.” Xu Anhe sat in the assigned seat. “The two things can’t be completely separated.” Su Yingzhen pushed the evidence list in front of her. “Then let’s do the half we can prove first.” The first screen displayed the original result from before the accident: 72,000. The second showed the result received by the agency: 59,040. The two files had the same model name and formula version, and both approval records pointed to Xu Anhe. If you only looked at the covers, they appeared to be the same document. Only after expanding the internal fields did the differences begin to emerge, one by one. The forensics specialist compared the model code first. No changes. Then they searched to see whether anyone had entered 59,040 directly into a result field. There was no such field. The final number had been generated after the system recalculated it using passenger records, time intervals, and classification rules. Users could not directly rewrite it on the report page. “So no one opened the report and deleted 72,000?” Su Yingzhen asked. “Not directly, at least not so far,” Xu Anhe said. “The result itself wasn’t changed. The data used to generate it was replaced.” Su Yingzhen pressed her. “How many places were changed?” The forensics specialist called up the difference list on the third screen. Four. None of the rows on the list were marked with a warning. Each item had the gray icon used for ordinary operations, and the creator, processing status, and description fields were all complete. If the two versions had not been placed on the same screen and compared field by field, they might simply have remained scattered across routine records in different departments, each passing its own inspection. The forensics specialist opened the first row of the difference list. The first item concerned the sampling range for the peak period. The original version retained the full ninety minutes. The later data adjusted how one segment of passenger movement was counted, according to an operations memo. The stated reason was to prevent brief activity from affecting the overall peak-period determination. Taken on its own, the wording was not unreasonable. The forensics specialist disabled this difference. The 59,040 on the screen rose back up by 3,600. “Who proposed this memo?” Su Yingzhen asked. The forensics specialist pulled up the document source. “The intermodal center’s operations analysis unit.” Su Yingzhen followed up. “And who actually made the change?” “The external data service provider updated the sampling settings according to the assignment. Both sides have work records.” Su Yingzhen divided a page in her notebook into two columns: proposal and execution. The forensics specialist switched to the second difference: transfer-passenger classification. Some passengers who only passed through the transfer level and did not remain there for an extended period had been moved from the primary exposure classification to the mobile-passenger category. The classification issue mentioned in the phone call before the earthquake had finally appeared in the formal data. The change document did not say “reduce the risk value.” It said only “reflect actual usage patterns.” The equivalent difference was 2,880. “The risk is genuinely different between passing through and remaining in place,” Xu Anhe said. Su Yingzhen turned to Xu Anhe. “So this change was correct?” Xu Anhe replied, “You can’t judge it by the name alone. We need to confirm whether the passengers removed from the original classification were still reflected in evacuation demand in some other way. If they were simply moved to another column and the model also read that column, the result might be reasonable. But if moving them there meant they were no longer counted, the people at the site wouldn’t disappear just because of that.” Su Yingzhen asked, “Do you know which of the two it is now?” “The latter.” Xu Anhe pointed to a field the model did not read. “But I still don’t know whether it was a design error or someone knowingly required it to be used despite understanding the result.” Su Yingzhen recorded “error” and “knowing” separately. Xu Anhe looked at the document’s creation time again. It was shortly before the phone call before the earthquake began. “This is the classification method the contact asked me about,” Xu Anhe said. Su Yingzhen looked up. “Did he know this adjustment had entered the formal data?” Xu Anhe shook her head. “I don’t know. He only said the operations unit had requested a reorganization, and he also promised to send the version history. Maybe he had only discovered that someone had proposed a change. Maybe he had already seen the result.” Su Yingzhen closed the document. “We can’t fill in testimony he never left just because he’s missing.” “I know.” Xu Anhe moved her gaze back to the creation time. The statement sounded cold, but it was necessary. Xu Anhe could not rewrite an unfinished phone call as a warning simply because she hoped the other person had tried to stop something. Nor could she put him in the position of the person who made the changes merely because his work had involved the classification issue. The forensics specialist switched the comparison screen to the third set of differences. This one involved the removal of duplicate records and was the most like a routine data-cleaning adjustment of the four. The intermodal center had gate, ticketing, video, and zone-sensor records running simultaneously. The same passenger could be recorded by different devices within a few minutes, so the system was designed to remove duplicate data. The later version broadened the criteria for determining duplicates, causing some records that originally represented different risk zones to be treated as a single passage. The equivalent difference was 3,240. “Removing duplicate data is something the model validation process normally does anyway, isn’t it?” Su Yingzhen asked. “It is,” Xu Anhe replied. “Then where’s the problem?” Su Yingzhen asked. Xu Anhe pointed to two merged anonymous records. “The same person appearing on the ticketing level and the transfer level in the same minute could mean the device clocks were out of sync, or it could mean they actually walked from one area to the other. Before merging them, we have to confirm the spatial and temporal conditions. This version treated the later record as a duplicate simply because the passenger ID was the same.” Su Yingzhen looked at the two merged records. “So the data became cleaner, but one segment of the actual route was deleted.” “Right,” Xu Anhe said. Finally, the forensics specialist opened the fourth row. This difference came from the data version selected by the automated integration system. The model received multiple data versions every day and selected its formal input according to completeness, timing, and approval status. Before the accident, the system selected a version marked “cleanup complete.” That data happened to contain the first three changes and was also missing a batch of records whose status confirmation was not yet complete. The equivalent difference was also 3,240. The forensics specialist added, “At the moment, it appears the system selected it automatically according to the existing rules. We found no indication that a user manually specified it at the last moment.” “Automatic selection doesn’t mean no one was responsible,” Xu Anhe said. “We need to see who configured the selection criteria, and why that data was marked as usable.” Su Yingzhen gave her a look. “We’re not reaching a conclusion on that today.” The four screens rearranged themselves. The original data was on the far left, followed by the four differences applied in sequence. A red 60,000 review threshold ran across every chart. After the first item was applied, 72,000 became 68,400. After the second, it dropped to 65,520. After the third, 62,280 remained. The value was still above the red line. Only when the fourth was added did the screen change to 59,040. A green marker lit up. Xu Anhe wrote out the complete calculation on paper: 72,000 − 3,600 = 68,400. 68,400 − 2,880 = 65,520. 65,520 − 3,240 = 62,280. 62,280 − 3,240 = 59,040. The four changes reduced the total by 12,960. A result that had originally exceeded the review threshold by 12,000 ended up 960 below it. “What if we remove one of them?” Su Yingzhen asked. The forensics specialist restored the four changes one at a time. Without the first, the result was 62,640. Without the second, it was 61,920. Without the third or fourth, it was 62,280. None of the results remained below 60,000. To prevent the order of operations from affecting the result, the forensics specialist reapplied the four differences in different sequences. The intermediate values changed depending on the order, but the final result always returned to 59,040. Xu Anhe confirmed that this meant the four reductions could be added together in the current model. It did not mean that the four people who carried them out had calculated them together. Su Yingzhen instructed the forensics specialist to record “calculation reproducible” and “personnel jointly aware” as two separate entries. The former could already be repeatedly verified in the review room. The latter remained blank. The review room was quiet for a while. The four screens on the wall looked like four modest adjustments in succession, each with documentation, work records, and a seemingly adequate explanation. Yet if any one of them were removed, the intermodal center would have required escalation review. “Can this prove that the four people colluded?” Su Yingzhen asked. “No,” Xu Anhe said. Su Yingzhen continued, “Can it prove that someone knew in advance the final result would be 59,040?” “Not yet,” Xu Anhe replied. Su Yingzhen circled the four differences with her pen. “Then what can this set of numbers prove right now?” Xu Anhe looked at the red line. “All four changes were indispensable. Together, they produced a result that happened to avoid escalation review.” Su Yingzhen proposed another possibility. “It could also be that four departments each made a reasonable decision, and the result happened to cross the line.” “Possible,” Xu Anhe admitted. Su Yingzhen looked at her. “Do you believe it was a coincidence?” Xu Anhe said, “I don’t need to believe in one answer first. I need to know whether anyone could see the other three changes when each of the four was proposed.” Su Yingzhen did not respond. She signaled the forensics specialist to call up the access and timestamp records. The first change was proposed by the operations analysis unit and carried out by the external service provider. The second came from a classification confirmation between passenger services and risk management. The third was applied by data-quality personnel. The fourth was selected by the automated integration system according to its status at the time. The four changes were not completed at the same time, nor were they made on the same operating screen. In the existing records, no frontline employee had personally handled all four. More troublesome was that each person’s change was minor. Placed back in its original work context, one person had been cleaning duplicate data, another had been correcting passenger classifications, and another had merely accepted the version automatically selected by the system. If only one link in the process was checked, it would be difficult for anyone to see the final red line being crossed. Su Yingzhen arranged the four records in a row. Pointing to the records side by side, Su Yingzhen asked, “If the four people each did something that appeared reasonable, does that mean no one committed a crime?” “No,” Xu Anhe said. “We need to confirm separately what each of them knew, what instructions they received, and whether anyone knew the combined result in advance.” Su Yingzhen continued, “Can you prove now that someone calculated it in advance?” “No,” Xu Anhe replied. Su Yingzhen pulled back her pen. “Then don’t write the answer into the question yet.” Xu Anhe did not argue. The forensics specialist reordered the records by creation time. Four work tracks extended forward separately across the screens, with no overlap. At least based on the screens currently before them, no one had opened all the data at once, and no one had issued a direct instruction to “bring the result below 60,000.” Su Yingzhen tapped the final 59,040 with her pen. “If no one handled all four sets of data at the same time, who knew the final number would be this?” Su Yingzhen asked. Xu Anhe looked at the four records, each normal on its own. She had been looking for whoever changed 72,000. Only now did she realize that the real question might not be who had personally altered all the data. It was who had let each person take only one step, until together they crossed the same red line.