The apprentice had never seen the fault before the software learned to flag it.
At Arjun’s factory Ravi, a junior technician involved in the pilot’s early trials, was assigned to review images marked suspicious by the inspection model. Most warnings were routine. The system displayed the likely surface defect, and an experienced worker confirmed the result.
Ravi had become good at following the suggested categories. He had not become good at noticing a problem the categories did not contain.
Leela realized this when she showed him a part with an unusual wear mark outside the model’s trained range. Ravi waited for the screen to name it.
“What do you think you see?” she asked.
“The model says unknown.”
“I didn’t ask what the model says.”
He turned the part beneath the light, uncertain where to begin. Leela spent twenty minutes explaining how the mark related to tool pressure, material temperature and recent changes in the line. The lesson had always been part of the apprenticeship. The new workflow had quietly removed many occasions to practice it.
Arjun’s manager argued that the system was saving training time.
“Training time for what?” Leela asked. “For pressing accept when a label looks right?”
They redesigned Ravi’s training schedule. Some inspections would be conducted without model suggestions, under supervision, before the tool’s output was shown. Apprentices would record their observations and compare them with the system afterward. The practice was not a competition to prove humans superior to machines. It was a way to preserve the ability to recognize conditions outside the model’s experience.
Mei welcomed the comparison. She wanted to know whether the model’s warnings were teaching workers useful patterns or encouraging them to stop looking.
In the first exercise Ravi made several mistakes. Leela corrected them and explained why.
He laughed at the first correction, then stopped laughing when the same class of mark fooled him again. During the break he asked whether his earlier inspections would be reviewed. Leela said they would examine a sample together, not treat every accepted model label as his personal failure. The model had made different mistakes on the same batch. Both records were kept for learning, not employee ranking.
Guo asked how the extra training would affect the factory’s business case.
“It costs time,” Arjun said. “It also means someone will be able to maintain the process when Leela retires.”
Leela raised an eyebrow at the mention of retirement.
“Don’t schedule my departure without asking me,” she said.
Everybody laughed, including Arjun.
At Henglan the support engineers faced the same problem. New staff learned to consult dashboards but rarely had access to the original equipment conditions that explained a warning. Ning arranged supervised visits and case reviews so technicians could connect a model’s output to the process it described.
Yanchuan volunteered to teach one session. He presented the smooth and untidy plots from Han Rui’s workshop and asked the trainees what could be concluded without knowing the test boundary.
One answered that the smooth plot showed higher performance. Another said the untidy plot showed a fault.
Yanchuan told them both were plausible stories and neither was yet established. Then he showed the conditions and asked them to revise their answers.
The exercise took longer than showing the finished conclusion. That was the point.
That evening Jianming asked how the training had gone.
“They made the same mistakes I made.”
“Did you tell them that?”
“Yes.”
“Good. Teachers who were never wrong are exhausting.”
His father spoke of learning electrical repairs from an older worker who would deliberately leave an obvious part on the bench and ask the apprentice to explain why it was not suitable. The lesson was not merely to recognize the correct component. It was to develop the habit of checking its purpose.
Yanchuan thought of Ravi waiting for a label.
In Pune, the apprenticeship plan gained a practical exercise each week. The model remained available for ordinary production. Training periods were protected from throughput targets so learners would not be punished for taking time to look.
Ravi’s first independent observation was small. He noticed a change in a tool’s sound and asked Leela to inspect it before the model recorded any warning. The resulting maintenance check found a developing issue that did not yet produce the trained surface mark.
Nobody called it proof that people were better than algorithms.
Leela wrote it in his training record as a correctly escalated observation.
A system could make a worker more effective without making the work understandable. Arjun wanted both. The second required teachers, time and a willingness to leave questions open long enough for someone to learn.
Yanchuan filed the lesson beside the pilot’s next technical milestone. He asked Arjun to add supervised practice to the next maintenance budget. Ravi’s name appeared as a trainee, not as a savings assumption.