Chapter 292: Hard Rules and the Desensitization Pool

By the time he returned to the rented coworking space, it was already four-thirty in the afternoon. Three of the voice-activated lights in the corridor were broken, and the one remaining light had a faulty connection, flickering on and off with the sound of his footsteps. Lin Chen pushed open the glass door and set his backpack down at the workstation by the window. When his left foot touched the floor, the dull pain shooting from his ankle into his calf made him instinctively hold his breath. He did not sit down. First, he bent over and pulled half a case of unopened bottled water from under the desk, wedged it beneath the chair, and rested his left foot on it. The numbness from obstructed circulation eased slightly, but the stabbing pain deep in the seams of the bone remained clear.

He opened his laptop. The screen lit up, reflecting the Excel spreadsheet he had left open the night before. The cells for fund allocation had already been marked red. Four hundred thousand had entered the jointly controlled account; after taxes and handling fees, the usable amount was three hundred seventy-two thousand. Eight GPU server instances, prepaid monthly, cost sixty thousand; Su Man's three-person team's salaries plus social insurance, ninety thousand; routine operations and maintenance, travel, and server bandwidth, twenty thousand. Three hundred seventy-two thousand minus one hundred seventy thousand left two hundred two thousand. That two hundred two thousand had to cover the next forty-five days of compliance evaluation, procurement of data desensitization tools, cooperation with the third-party audit, and a buffer for emergencies.

Lin Chen created a new sheet and named it Delivery Checklist V1.0 - Compliance Edition. He was used to breaking complex problems down into executable granularity. The first column was milestone, the second deliverable, the third acceptance criteria, the fourth responsible person, and the fifth risk response plan. His fingers tapped the keyboard in a steady clatter.

Milestone 1: Underlying data desensitization. Deliverables: desensitization rule library + execution scripts. Acceptance criteria: 100% replacement rate for patient names, ID numbers, mobile numbers, and detailed addresses; clinical diagnosis fields retained. Risk response: mixed GBK and UTF-8 encoding in the old system may cause regex matching failure; add an automatic encoding detection module. Milestone 2: Localized inference node deployment. Deliverables: edge server image + offline model package. Acceptance criteria: core diagnostic functions usable without network access, response latency < 2 seconds. Risk response: unstable power in the hospital machine room; configure UPS backup power. Milestone 3: Interface joint debugging and audit logs. Deliverables: API integration documentation + full operation traceability system. Acceptance criteria: provincial health commission evaluation team can export audit trails with one click, with no privilege-escalation vulnerabilities. Risk response: the hospital's old HIS system interfaces may not be open; capture through an intermediate database and add data consistency checks.

After typing the last line, he saved the file and exported it as a PDF. He picked up his phone and called Su Man. It rang four times before connecting. In the background was the low-frequency hum of machine-room fans.

"I've sent you the checklist," Lin Chen said. "For the forty-five-day evaluation period, we need to recut the budget. Desensitization and compliance pre-review take the largest share. On your side, can the model fine-tuning schedule be pushed back by a week?"

"It can't." Su Man's voice was calm, with the slight hoarseness that came after staying up late. "After the clinical director signed off, the pilot departments already started running. The doctors' feedback says the auxiliary diagnosis confidence is around 82%, but when it encounters rare complications, the model still gives conservative recommendations. I need to feed in the newly labeled three thousand medical records this week and run one incremental training pass. Otherwise, when we roll it out hospital-wide, the false positive rate will exceed the standard."

"Incremental training needs compute." Lin Chen looked at the budget sheet on the screen. "The eight GPU instances are all running real-time inference for the pilot departments right now. If we split off part of them for training, inference latency will rise."

"We can stagger it." Su Man paused. "From two to six in the morning, hospital system traffic is at its lowest. I'll move the training jobs to that window. But the desensitization script has to be verified first. What the evaluation team cares about most is data export security. If our model is going to run locally, the training data cannot leave the machine room. The unstructured text in old medical records, such as doctors' handwritten progress notes, contains a lot of abbreviations and colloquial descriptions. The miss rate of the current regex rules is too high."

"How high is the miss rate?"

"About 3.4% in the test set. Mainly because older doctors are used to writing things like 'Wang bed' to mean 'Director Wang's bed,' or abbreviating a patient's name as surname plus bed number. The regex treats 'Wang bed' as a patient name and replaces it directly with '***', which breaks the clinical context."

Lin Chen was silent for a few seconds. A 3.4% miss rate might be acceptable in a technical test, but in a provincial medical safety evaluation it was a hard flaw. The evaluation team would not listen to explanations; they would only read the report.

"Add another layer of rules," Lin Chen said. "Don't rely purely on regex. First run entity recognition and extract contextual features such as 'doctor,' 'nurse,' 'bed,' and 'department.' If 'surname + bed number' appears in the medical order signature field, mark it as medical staff and don't desensitize it. If it appears in the patient information field, replace it. Can the script be ready tonight?"

"Yes. But we need manually reviewed samples," Su Man said. "I'll have the annotation team pull five hundred records and manually check them. I'll give you feedback tomorrow morning."

"Good. I'll adjust the budget on my side. Training compute will be scheduled off-peak, and the desensitization script takes priority. Push the checklist milestones forward in that order."

"Understood." Su Man hung up.

Lin Chen put down the phone and picked up the thermos cup on the desk. The water had gone cold. He took a sip, and a faint spasm rose in his stomach. From the drawer he found half a blister pack of ibuprofen, pressed out two tablets, and swallowed them with the cold water. It would take forty minutes for the medicine to take effect. He could not wait. He had to write out the underlying logic of the desensitization script clearly.

The screen's light reflected on his face. He opened the code editor and created a new file, data_masking_v3.py. His left hand rested on the keyboard, his right foot braced against the desk edge, his left foot suspended in the air. He began writing the class structure for the rule engine. He did not pursue elegance, only stability. Every try-except block had detailed logging; every regular expression came with test cases. He wrote slowly, but steadily. Whenever he encountered an uncertain boundary condition, he stopped and drew a flowchart in the paper notebook beside him. Arrows, decision boxes, exception branches. Just like back in the main room of the house in Qingshi Village, when he had dismantled Python syntax in front of that second-hand computer.

At nine in the evening, the medicine began to take effect, and the stabbing pain in his foot turned into a dull ache. He stood up and slowly walked two laps in the narrow aisle, moving his stiff joints. Returning to his seat, he continued debugging. At eleven, the script ran successfully. He imported a desensitization test set of five thousand records. The progress bar advanced slowly. 98.1% passed. 1.9% entered the exception queue. He opened the queue and checked the entries one by one. Sure enough, most were abbreviations like "Li bed" and "Zhang nurse" being misclassified. He adjusted the contextual weight parameters and ran it again. The pass rate rose to 99.4%. The remaining 0.6% were genuine rare characters and mojibake. He added a fallback logic: fields that could not be recognized would all be replaced with [REDACTED], and the original hash would be recorded for later manual review.

At one-twenty in the morning, he packaged the final script and test report and sent them to Su Man. When the email sent notification sounded, he leaned back in the chair and closed his eyes. His breathing gradually steadied. Funds, technology, and compliance had finally meshed together. There were no miracles, only decomposition, trial and error, and correction.

The phone screen suddenly lit up. It was not a reply from Su Man. It was a system email from the provincial health commission evaluation preparation group. The subject line read: Notice of On-Site Verification for the Safety Evaluation of the AI-Assisted Diagnosis and Treatment System.

Lin Chen opened his eyes and tapped the attachment. The body was short: "After comprehensive assessment, the originally scheduled forty-five-day evaluation cycle has been compressed to twenty-eight days. The on-site verification team will enter your company and the pilot hospital on the 22nd of this month. Please prepare the data desensitization ledger, local deployment environment, and third-party penetration testing report in advance. Failure to do so by the deadline will be deemed automatic forfeiture of evaluation eligibility."

The 22nd. From today, only nine days remained.

The budget had not changed, the compute had not changed, but the time had been cut in half. The desensitization script's coverage rate had to rise from 99.4% to 99.9%. The penetration testing report required an expedited qualified third-party agency, which would cost at least thirty thousand. The money in the jointly controlled account, originally buffered by weekly disbursements, now had to be front-loaded in a single release.

Lin Chen stared at the date on the screen. His left foot began to ache faintly again. He did not move. He only reached out, opened the drawer, and pulled out the thick error notebook. On a new page, he wrote: 1:25 Evaluation cycle compressed to 28 days. Risks: insufficient desensitization coverage, cost overrun for expedited penetration testing, broken fund disbursement rhythm. Countermeasures: redraw the Gantt chart tonight, contact a third-party agency to lock in a slot, apply to Zhao Qiming for 30% early release of the first tranche.

The tip of the pen paused. He added one last sentence: Note: Don't gamble on luck. Calculate probabilities.

Outside the window, the city had already fallen asleep. Only the occasional truck passing over the distant elevated road made a low rumble. Lin Chen closed the notebook and placed his hands back on the keyboard. The light from the screen illuminated the bloodshot veins in his eyes, and also the next line of code waiting to be written. Time waited for no one, but rules could be laid in advance. He pressed Enter and began restructuring the progress chart. The next hard battle would come in nine days. And he was already ready.