Chapter 320: Traces and Reversal

9:20 a.m. Lin Chen arrived at the office building housing Zhao Qiming’s company forty minutes early. The elevator mirror reflected his silhouette: the dark gray suit Su Man had helped him buy last week fit well, but a slight stoop from years of hunching over a desk was still visible along the shoulder line. He glanced down at his left foot. Custom silicone insoles inside his leather shoes compressed the pain into a dull, muffled throb, like the vibration of a distant pile driver echoing through concrete. He adjusted his breathing, shifted his weight evenly onto his right foot, and his left hand unconsciously rubbed the edge of a black hard-shell folder.

9:45 a.m. He sat in the waiting area outside the conference room. Sound-absorbing carpet lined the corridor, swallowing every footstep. On his lap, the folder held no polished PowerPoint slides—only printed fragments of raw logs, scanned signatures from annotating physicians, raw A/B test curve graphs, and that dog-eared mistake notebook with colorful index tabs sticking out from the page corners. Su Man messaged him on WeCom: “Do you need me to come up? I can monitor the legal and commercial terms on-site.” He replied: “No need. I’ll handle the technical explanations, you watch the backend for the terms. Too many people dilutes focus.” He knew that in capital negotiations, emotional friction was the greatest risk. One person, one set of data, one bottom line. That was enough.

10:00 a.m. The glass door swung open. Zhao Qiming’s assistant ushered him in. The conference room was spacious, with Zhao Qiming seated at the far end of a long table. Flanking him were two risk control officers and an external medical consultant. The air carried a faint scent of coffee and dry paper. No pleasantries. Zhao Qiming raised a hand to gesture: “Sit. Let’s begin directly.”

Lin Chen pulled out a chair and sat down. He laid the folder flat on the table and slid the first document forward. “A variance of zero point three percent. It doesn’t fall within the core delivery scenarios, but since you raised it, I must provide the source.” He opened the mistake notebook, pointing to a hand-drawn tree diagram on one page, annotated in red pen with timestamps and parameters. “In July 2021, the provincial hospital’s pediatric outpatient system was upgraded. The chief complaint field was changed from free text to a structured dropdown. During historical data cleaning, we preserved the original text, but during model training, the noise rate from unstructured expressions in patients under six years old reached thirty-four percent. The annotation team recommended down-weighting to prevent the model from overfitting to colloquial descriptions, which would degrade its generalization ability for the adult cohort.”

He pulled out a second document and flipped to the cross-validation page. “This is the comparison before and after down-weighting. Core metrics: adult pulmonology recall rate increased by one point two percent, while pediatric specificity decreased by zero point three percent. Clinically speaking, pediatric initial screening already requires manual review. The model’s positioning is for auxiliary triage, not diagnostic replacement. For that zero point three percent, we chose to log the trace rather than flatten it with a smoothing algorithm.”

Zhao Qiming didn’t look at the documents. His gaze rested on Lin Chen’s face. “Capital demands scalable certainty. Leaving a trace implies uncontrollability. If the due diligence report states ‘active down-weighting,’ auditors will ask why data augmentation wasn’t performed before training, or why synthetic data wasn’t used to fill the gaps.”

Lin Chen met his gaze. His voice was low but every word was clear. “Mr. Zhao, medical data isn’t e-commerce clickstream. Uncontrollability is the norm; what we control is the fault-tolerance mechanism. Our mistake notebook records seventeen weight adjustments. Each one has clinical justification, a testing baseline, and a rollback contingency. The variance you see is a boundary the system actively exposes, not a flaw.” He paused for a second, his finger lightly tapping a page in the notebook. “If you need absolutely smooth curves, there are ready-made packaging teams on the market. But you’re investing in a vertical application, not data beautification. Packaged certainty won’t clear the procurement catalogs of Grade-A tertiary hospitals. Doctors look for interpretability, not pretty curves.”

The conference room fell silent. Only the faint hum of the air conditioning vent remained. Zhao Qiming’s fingers tapped lightly twice on the table. The risk control consultant flipped through papers, murmuring something to the person beside him. Zhao Qiming raised a hand to stop him. He picked up the mistake notebook Lin Chen had slid over and flipped to a middle page. Taped there was a yellowed sticky note with a handwritten code snippet from a county-town internet café a decade ago, annotated in red pen beside it: Fault tolerance is not compromise; it’s leaving a way out. The rollback path must be clearer than the forward path.

“The twenty percent discount on valuation was a stress test,” Zhao Qiming said, closing the notebook, his tone even. “But your explanation passes the technical hurdle. The ledger for medical compliance can’t just calculate algorithms; it must account for clinical pathways. Your logic can go into the due diligence report.” He turned to his assistant. “Issue the Term Sheet at the original valuation. But add a performance clause: within six months of product launch, you must secure formal procurement contracts from two Grade-A tertiary hospitals. If you miss the deadline, equity dilutes by five percent.”

Lin Chen didn’t agree immediately. He looked down at his left foot. The toe was slightly numb, and the edge of the silicone insole had already worn into fine creases. He thought of the firelight from his mother boiling medicine at the stove, of the last uncolored star in Xiao Man’s sketchbook, of the crisp notification chime when Old Zhao made his first transfer. Capital’s terms were never charity; they were exchange. He looked up. “Six months. Agreed. But the acceptance criteria for the procurement contracts must strictly follow the clinical efficacy report we submit. No last-minute demands, no delaying payment milestones under the guise of ‘internal hospital procedures.’ Write these into the annex.”

Zhao Qiming studied him for three seconds, then nodded. “It’ll go in the annex. Legal will send it to you this afternoon.”

11:20 a.m. The meeting concluded. Lin Chen walked out of the office building. The early autumn sun was glaring. Standing on the steps, he slowly unbuttoned his suit jacket. His phone vibrated. A message from Su Man: “Terms synced with legal. Valuation locked. Congratulations.” He replied: “Received.” No extra words.

He walked to a corner convenience store and bought a bottle of mineral water. Twisting off the cap, he tilted his head back and took a drink. The cool water slid down his throat, suppressing the heat in his stomach. He leaned against the glass storefront, watching the silhouette of the hospital building across the street. The procedural pipeline was cleared, but the real battle had only just begun. A procurement contract wasn’t finished with a signature. Clinical deployment, hospital system integration, physician training, compliance audits—every step was a slog. Capital provided a ticket to enter, not a shield.

His phone vibrated again. This time, it was a call from the head of the provincial hospital’s annotation team. Lin Chen answered.

“Director Lin.” The caller’s tone was urgent, the sound of flipping papers audible in the background. “We just received an internal notice from the Health Commission. Starting next month, all AI-assisted diagnostic systems involving patient data must pass a newly added ‘dual review for algorithm filing and data traceability.’ The ethics approval you submitted might not be enough. The filing window is only fifteen days. The material checklist hasn’t been issued yet, but internal briefings say they’ll be scrutinizing data lineage and the interpretability of feature engineering. That batch of historical cleaning logs of yours will need de-identification mapping redone.”

Lin Chen’s fingers tightened slightly around the water bottle. The plastic emitted a faint creak. He watched the traffic on the road, traffic lights cycling, pedestrians hurrying past. His voice remained steady. “Understood. Send me the checklist the moment it’s out. We’ll start preparing the traceability graph tonight. For the de-identification mapping of the historical logs, we’ll use differential privacy combined with a hash chain. We won’t touch the raw fields. Fifteen days is enough to break it down into steps.”

He hung up. Twisting the cap back on, he tossed the empty bottle into a recycling bin. His left foot hit the ground; the pain remained, but his stride didn’t falter. He pulled out the mistake notebook, flipped to a fresh page, and set his pen down. The handwriting was neat: Compliance upgrade. Dual-review window: 15 days. Core: data lineage tracking and algorithm interpretability. No evasion. Break it into steps.

Wind swept through the street, lifting a few fallen leaves. He turned and merged into the flow of pedestrians.