The new model beat the old one every time they ran last month’s data through it.
That was exactly the problem.
Ning had prepared an evaluation of the forecasting improvements. The model showed smaller errors, earlier warnings and fewer unnecessary recommendations. Guo’s analyst wanted to use the chart in a proposal to another industrial customer.
Yanchuan looked at the source dates.
“When was the production schedule entered?” he asked.
“The final schedule is in the archive.”
“Was that version available to the operator before the shift?”
The analyst stopped scrolling.
Several archived schedules had been corrected after the events they described. A maintenance entry was completed the morning after a repair. A late truck arrival had been entered with the original expected time and no visible history. The evaluation had supplied the model with facts the factory staff could not have known when they needed to decide.
“It is still historical data,” the analyst said.
“Yes,” Yanchuan replied. “History has more information than the past had at the time.”
They rebuilt a sample of decisions from timestamped versions retained in the source systems. Some records could not be reconstructed reliably. Those were marked unavailable rather than filled with the final value.
The model’s improvement became smaller. It did not vanish. That was enough to justify further testing, but not the promise the first chart had implied.
Chen wanted a name for the mistake.
“Look-ahead information,” Ning said. “Sometimes leakage. Here it’s a mixture of version history and our own assumptions.”
She wrote the explanation in ordinary language: the first test allowed the model to use records that had been added or corrected only after the decisions were made.
Gao understood immediately.
“So you’re letting the machine read tomorrow’s logbook.”
“Some of it,” Yanchuan admitted.
“That’s a very clever machine.”
Nobody laughed until Gao did.
Guo asked for two reports: one showing what the model could do with complete historical information and another showing what it could have done in the moment. The second would guide investment decisions. The first could be useful for diagnosing faults, but it was not an operational forecast.
They also added a test in which the model had to identify that information was missing rather than supply an elegant answer. It declined to recommend a cooling change during an uncertain maintenance state. That made the system appear less active. Gao called it progress.
At dinner Mei said she had made the same kind of mistake in an early image project. Training labels had been corrected after experts inspected damaged products. A system evaluated against the corrected set seemed to know which objects would fail long before the inspection was complete.
“It wasn’t predicting failure,” she said. “It was agreeing with a verdict written later.”
“What fixed it?”
“Recording what we knew at each stage. More work than we wanted.”
Yanchuan thought of the file his brother had sent. In November he had almost interpreted every silence in the family messages as evidence of something sinister. Later information made the old conversations feel different. It did not change what he had known when he answered with a thumbs-up.
He wondered whether his memory of the last call had become cleaner and more meaningful each time he replayed it.
At the scheduled family review An showed a revised chronology with source times distinguished from reporting times. The hotel payment was entered on one date and described by the employer on another. The transport booking had changed before the route was completed. The visitor pass was returned at a time that did not identify its holder’s movement.
“Why did you change the columns?” he asked.
“Because people keep telling us when a record was written as if that were when the event happened.”
He laughed softly. It was the same issue that had occupied him at work.
A new official note said the investigator was comparing an annex witness account with available site records. No verified movement beyond the receiving area could yet be disclosed. An recorded the date of the note and the period it described separately.
There was no secret code in the difference. It was merely one more place where accuracy depended on the right kind of time.
At the factory the final evaluation included every available versioned decision. The confidence bands widened. A few recommendations that had looked successful in hindsight could no longer be reproduced. The report stated the limitation instead of estimating what the missing records might have shown.
The client accepted a longer test with stricter data retention. The team was required to preserve when input became available, who changed it, and whether the recommendation had been actionable before the deadline.
Guo said this would make the models harder to compare with competing products.
The sales team had already sent the original chart to a prospective customer. Nobody volunteered to call them. Chen asked who had approved the attachment; the analyst raised his hand. He had used the version marked final, and looked frightened enough that Yanchuan wanted to say it had been understandable. Chen asked him to draft the correction and offered to review it before it went out. The customer would see the difference between the two charts. They could lose the meeting.
“Then compare the promises, too,” Yanchuan answered.
On the train home he looked out at lights moving across fields. The past was full of events that now seemed inevitable because their outcomes were known. At the time, people had seen only fragments. His brother had sent a question about a report, not a map of his disappearance. The family had acted with the information they had, and sometimes they had been wrong.
He could improve the record without rewriting their choices as if they had known the future.
He opened the old message once, then closed it before the phone battery died.
At work the next morning, the team’s first rule for historical testing was printed on the wall: Only use what could have been known when the decision was due.
Yanchuan added a second line beneath it: Keep the later corrections, but mark them as later.
He checked the timestamp on the next training extract before showing it to Ning. The correction was small enough to miss; this time he did not let it pass.