The new forecast was right to within one percent, and Gao refused to use it.
Yanchuan stared at the screen where the predicted evening peak lined up with the measured curve. The model had performed better than anything the team had tried. In the weekly review Guo’s analyst had circled the result with a green pen and written excellent beneath it.
Gao had written insufficient warning time.
“The number is right,” the analyst said.
“It arrived seven minutes before the load changed,” Gao replied. “We needed twenty to move the cooling schedule without disturbing production.”
The room went quiet. Yanchuan had been about to explain the model’s improved training data. He shut the slide.
Ning asked how the forecast could be made earlier. That was not a simple matter of moving the timestamp on a chart. The model relied on a planned production update that the factory did not confirm until shortly before each shift. Forecasting with an earlier version of that plan would increase uncertainty.
“We can give you a range twenty-five minutes ahead,” Yanchuan said. “It will be wider.”
“How wide?” Gao asked.
“We’ll measure it.”
They reconstructed four weeks of decisions, using only data available at the moment an operator would have had to act. Predictions that looked splendid when evaluated after the fact became less confident. A planned machine restart had been entered late. A truck carrying a replacement part had arrived early. On another day the manager had changed an order without updating the system until the following morning.
The revised forecast was less precise and more useful. It marked a zone in which a peak might arrive, the assumptions that made it likely, and the last time at which an adjustment could safely begin.
Gao accepted the second recommendation in the trial, not because its predicted value was especially close, but because he had time to inspect it.
Ning added two measures to the report: prediction error and decision usefulness. The first measured how far the estimate had been from reality. The second recorded whether an operator could act within approved constraints. She warned the manager that the two would not always improve together.
Guo studied the page.
“Could we sell the second measure?” he asked.
“We can sell a service evaluated against it,” Chen answered. “We shouldn’t sell a score without defining what it means.”
At home Yanchuan told Mei the story. She laughed at the one-percent forecast.
“Your clever machine told them what they were already too late to change.”
“It wasn’t that bad.”
“Wasn’t it?”
He admitted it had been.
She described a fault-detection system she had built years earlier. It identified defects with high accuracy after an image had been archived. The factory needed a warning while the affected product was still on the line. A less accurate model running earlier had prevented more waste.
“Did you publish the comparison?” Yanchuan asked.
“After our client stopped calling the first model intelligent.”
They sat over noodles, comparing ways to measure the cost of arriving with the truth after the useful moment had passed.
On the next scheduled family review, An had little news. The case investigator was still checking the compound’s internal movement records. The technical review’s finding had been accepted by the operator as requiring corrective work. No finding tied the reporting defect to the disappearance.
Yanchuan almost asked whether the process could be accelerated with another letter. Instead he asked whether there was a deadline for the next agreed update. There was, and it had not passed.
Their son was at the table with a clock drawn beside a cardboard bridge.
“What’s that for?” Yanchuan asked.
“To tell the cars when the road is open.”
The boy had learned that building a bridge strong enough for a toy bus did not help if the other end of the road was blocked by furniture. He was writing down the order of the traffic signals, proud of having invented a system whose purpose he understood.
Yanchuan watched him for a while.
At the factory, the next month’s report kept the ugly uncertainties. The broader early forecast produced more false alarms than the late one. Operators declined some recommendations because the available reserve was too small; they accepted others after checking the maintenance schedule. The team could calculate the cost of each response without pretending every declined suggestion was a loss.
A commercial manager asked whether the result could be summarized as improved prediction accuracy.
“No,” Yanchuan said. “It was an improvement in decision support. The accuracy figure for the early window got worse.”
The manager frowned, then asked them to explain the value in ordinary language.
Ning wrote: Operators received warnings early enough to decide, with stated limits, rather than precise estimates after the decision window had closed.
Gao signed off on the sentence. He had been the one carrying the decision, and for once the success claim described the work he actually performed.
That evening Yanchuan walked past the bakery where the lights had failed during the substation outage. The owner was closing and carrying a tray of unsold bread into the back room. He asked whether she knew how much demand she would have tomorrow.
“Enough to make some bread,” she said. “Not enough to know what everyone will buy.”
He bought the last bun. The forecast for the factory sat in his work bag, wrapped in qualifications. Gao’s revised operating sheet specified twenty minutes of usable warning. The next test would score that interval, not merely the closeness of two curves.