The camera images never left Arjun’s factory, and the model still improved.
Mei had expected the restriction to make the next revision harder. The previous development plan assumed that she would examine a large collection of annotated images in Shenzhen, inspect the difficult cases and retrain the classifier herself.
Arjun’s legal and engineering teams approved a narrower arrangement. The production images belonged to the factory. The labels sometimes included notes about customers, batches and employees that were not needed to teach a model to recognize the chosen surface mark. Broad export would increase the risk without a clear necessity.
“Then give me the measurements I need,” Mei said.
Leela and the local engineer selected a permitted set of aggregate error counts, example characteristics without identifiers and a small number of explicitly cleared reference images. They performed the first training revision on a local machine. Mei reviewed the method, not the entire underlying dataset.
The first week was frustrating. A summary table did not show why one particular reflection had fooled the classifier.
Mei drafted a request for the original image, deleted it, and drafted it again. She had built the earlier prototype by looking at failed examples herself. Waiting for another engineer to answer a question she could usually resolve in minutes felt like losing control of her own work. She called Leela and asked whether they could agree on a faster local review process rather than an exception to the data rules. Instead of demanding every image, Mei wrote a question describing the pattern she needed to understand. Leela tested it against local examples and replied with the conditions under which the false warning occurred.
They changed the lighting setup.
The error rate improved. No one had to send the entire production archive anywhere.
Guo asked whether keeping data local would prevent them from building a single model for all customers.
“Perhaps a single model isn’t the first thing we should promise,” Ning said.
His analyst proposed sending only extracted features. Mei asked whether those features could still identify customers or sensitive production patterns when combined. The answer required review. They did not call a transformation private merely because it removed the original image.
They designed the system so the factory could perform essential inspection, record corrections and stop recommendations without continuous access to an overseas service. Approved engineering updates could be transferred when the parties agreed on their contents. The local team kept the ability to inspect and refuse them.
Arjun said that made the arrangement slower.
“Yes,” Mei replied. “It also means the factory doesn’t stop inspecting parts because a network connection breaks.”
One afternoon the link did break during a permitted test. The local tool continued to mark parts for review. The shift supervisor received no new model update, but the existing version remained available, with its limitations displayed.
Leela said the exercise had taught her more about the product than the sales demonstration.
At the same time Asha wrote from Kenya about a similar boundary in the maintenance-map project. Her team did not want to export staff names and routes to a commercial partner. They were willing to share aggregated travel-time scenarios for an agreed technical study. The map itself would remain under local control.
Yanchuan noticed how different the two projects were. A defect classifier at one factory and a maintenance-planning framework for a power network might use some common engineering methods. Their data, permissions and public consequences could not simply be merged into one global asset.
Guo wanted a portfolio dashboard. They designed one that reported approved project-level measures without requiring the underlying personal or operational records. It was less detailed than the analyst first requested and more defensible about what each measure meant.
On an evening call An asked why a system advertised as intelligent needed people to approve every transfer.
“Because deciding what can be shared isn’t just a pattern-recognition problem,” Yanchuan said.
“Then remember that when someone asks for our case documents.”
He understood the warning. The family had spent more than a year learning when to share information with investigators, advisers, employers and the public. Their most painful records were not raw material for improving an unrelated product.
He checked the permissions on the family archive and confirmed that people involved in the search had only the access they needed. He did not remove any record from an investigator or disrupt the agreed process. He simply stopped treating convenience as a reason to copy everything into every folder.
At the factory in Pune, the next model revision was evaluated locally. The results were sent as a limited report with the tested conditions and failures recorded. Mei could inspect the method and propose changes. She could not pretend she had personally reviewed every image.
“That limitation belongs in my signature,” she told Yanchuan.
He agreed.
A month later Guo presented the new operating model to his investment team. The slide named the local owners of data, the support obligations and the point at which cross-border access required a new agreement. One investor asked whether those conditions would make the business harder to scale.
“Probably,” Guo said. “They may also keep us from selling a service we can’t responsibly maintain.”
Yanchuan did not know whether Guo entirely believed the statement or had learned it would reassure cautious clients. He cared more that the contract and technical design made the conditions real.
Some knowledge crossed borders as a method, a carefully limited example or a question. Some stayed with the people whose lives and work gave it meaning.
Mei closed the request for raw-image export. The next improvement report would contain only the measures Arjun’s team had authorized for release.