Chapter 321: Roots and Ruptures

The elevator descended to the second basement level. Lin Chen stepped out of the car. As his left foot touched the ground, a dull ache in his arch traveled up his Achilles tendon. He paused for two seconds, shifted his weight, and pushed open the glass door.

The office area was brightly lit. The flowchart from the morning’s due diligence meeting remained on the whiteboard, the marker lines long dried. Su Man sat at her desk, a draft compliance checklist waiting to be broken down on her screen. Hearing his footsteps, she looked up and handed him a printed schedule.

“Fifteen days, split into three phases.” Her voice was kept low, careful not to disturb the annotators resting next door. “First three days: align the historical data lineage graph. Middle seven days: run differential privacy anonymization and hash chain mapping. Final five days: internal cross-validation and material packaging. The tech team is in place, but we need to reallocate computing resources. We’ll have to suspend forty percent of the training cluster’s load to make way for the traceability task.”

Lin Chen took the schedule, his eyes scanning the milestones. No hesitation. “Stop. Suspend the training tasks, freeze the model versions. Traceability takes priority.” He walked to the whiteboard, picked up the eraser, wiped away the morning’s flowchart, and drew three parallel lines. “Data lineage isn’t about drawing relationship diagrams; it’s about finding breakpoints. The logs from the early outsourced cleaning, the raw shards from the provincial hospital’s annotation team, our self-built feature library—there are three migration records in between. Each one could have lost metadata. Tonight, we audit the first one.”

Su Man nodded and turned to coordinate server permissions. Lin Chen returned to his desk and pulled a black anti-static bag from the bottom drawer. Inside were two mechanical hard drives phased out in 2015, their label paper yellowed and curled at the edges. They were pulled from second-hand servers bought with the final payment settled by Old Zhao when the company was just starting out. They held the original cleaning logs for the very first batch of one hundred thousand outpatient text records.

He connected the drive enclosure. The indicator light flickered on with a faint hum. The system recognized it, mounted it. Folders unfolded layer by layer, the dates frozen eight years in the past. He opened the first CSV file, and the screen scrolled through dense fields: patient_id, symptom_raw, clean_flag, timestamp. He quickly dragged to the bottom, checking the continuity of the timestamps.

Continuous. He opened the second. Continuous. The third: a gap.

Lin Chen’s fingers paused on the trackpad. From November 14 to November 21, 2015, seven days of logs were missing. Not file corruption—they were never written in the first place. He pulled up the server migration records from that period and found a single remark: Power outage in the server room. Backup power did not cover the NAS. Partial write queue lost.

He leaned back in his chair and closed his eyes. Under the desk, his left foot twitched slightly. He reached down to press his knee, waiting for the spasm to pass. The dual compliance audit required data lineage verification; a gap meant a broken traceability chain. The Health Commission’s auditors wouldn’t care about “power outages.” They only cared whether the hash chain was closed.

Su Man walked over with two cups of warm water and set them on the corner of his desk. “Stuck?”

“Seven-day gap in the early logs.” Lin Chen opened his eyes, his voice steady. “The raw fields are still there, but the cleaning timestamps and operator IDs are gone. If we report it as is, the lineage graph will leave a conspicuous gap here.”

“Backfill the timestamps?” Su Man frowned. “If we use the current system time to fill them in, the hash values won’t match. The audit tools will flag it the moment they run.”

“No backfilling.” Lin Chen shook his head. “Fabrication is a red line. If it’s discovered, the due diligence gets vetoed outright, and the valuation drops to zero.” He opened his drawer, took out a mistake notebook, and flipped to a blank page. “We find an alternative path. The hospital’s IT department has audit logs. For those seven days, the outpatient system access volume, API call records, even the registration machine transaction logs—they’re all on the intranet servers. They don’t directly map to our cleaned data, but they prove the data was actually generated during that period. We use the audit log’s time window as an outer wrapper, apply differential privacy with salted hashing internally, and mark the gap segment as ‘historically migrated and untraceable,’ isolating it separately. In the traceability report, we clearly define the boundaries. No cover-ups, no sugarcoating.”

Su Man was silent for a few seconds. “The auditors will ask why we don’t just fill the gap with synthetic data.”

“Because medical data isn’t e-commerce clickstream.” Lin Chen picked up a pen and wrote in the notebook: Do not patch breakpoints; isolate and label. Use audit logs as temporal anchors. Clear boundaries beat a false closed loop. “Synthetic data would contaminate the feature distribution. We’re submitting a clinical decision support system, not a demo. The audit demands explainable gaps, not a flawless illusion.”

“Understood.” Su Man turned to contact the old liaison at the provincial hospital’s IT department. “I’ll pull the audit transaction logs. You run the hash mapping script. The computing power has already been switched over.”

Lin Chen nodded. The screen lit up, a terminal window popping open. He typed the first line of code: import hashlib. The rhythmic clacking of the keyboard echoed through the quiet office. He lifted his left foot slightly, keeping it suspended to avoid pressing on the sore spot. His right hand controlled the mouse while his left hand touch-typed. The script loaded, read the raw fields, salted them line by line, and computed the SHA-256 hashes. The progress bar crawled upward.

1:00 AM. The coffee cup was empty. The script hit the gap segment and threw a warning. Lin Chen didn’t stop. He pulled up the compressed audit log package sent by the hospital’s IT department, extracted it, and pulled the timestamp sequence. He wrote the mapping logic: cross-validating the feature hashes of the gap segment with the audit log’s time window to generate an outer wrapper file. The code executed successfully, outputting the log: Isolated segment marked. Lineage chain closure rate: 93.7%.

He leaned back in his chair and let out a long breath. His stomach felt hollow and anxious, but he didn’t move. His left foot had gone completely numb, like a piece of wood that no longer belonged to his body. He flipped the mistake notebook to a new page and wrote: Phase one of traceability complete. Breakpoints isolated. Compliance documentation framework established.

His phone screen lit up. A new message popped up on Enterprise WeChat. Sender: Science and Technology Division, Provincial Health Commission. Attachment: AI Medical Algorithm Filing Material Checklist (Trial Version).pdf.

Lin Chen opened it. He checked it item by item. The first twelve matched the preliminary briefing. He flipped to the end, and his gaze froze.

Article 13: The first batch of training data must be accompanied by a sealing certificate for the original physical media, including hard drive serial numbers, sealing timestamps, and dual signatures. The submission window is compressed to seven days. Failure to submit by the deadline will be deemed an automatic forfeiture of filing eligibility.

Seven days. Not fifteen.

Lin Chen stared at the line. The cold glow of the screen reflected on his face. He looked down at the black anti-static bag on the corner of his desk. On the label paper of the two old drives, the serial numbers were handwritten. He picked one up and turned it over. A string of laser-etched codes was engraved on the metal casing. It differed from the number on the label by a single digit.

He set the drive down and picked up his pen. In the notebook, he wrote: Media serial numbers mismatch. Sealing certificate missing. Seven-day window.

Outside the window came the sound of the early-shift street sweeper. Dawn had not yet broken. He closed the notebook, stood up. His left foot hit the ground; the pain was dulled, but his stride remained steady. He walked to the window, watching the grayish-white light creeping over the city’s edge. Seven days. Enough to break down the steps.