The Simulated Year

The service cost more than Julian was willing to admit to Noa. What it promised, in return, was a full narrative simulation of the next twelve months under either decision: take the transfer to the Lisbon office, or stay. Fed his calendar, his spending history, and three years of unguarded text messages, the model produced two documents titled, with a bluntness he found almost funny, "Julian A" and "Julian B."

这项服务的价格，比朱利安愿意向诺亚承认的要高。作为回报，它承诺提供一份完整的叙事模拟，推演他在两种决定下的未来十二个月：接受调往里斯本办公室的安排，或者留下来。输入了他的日程、消费记录，以及三年来毫无防备的短信之后，模型生成了两份文件，标题直白得让他几乎觉得好笑，叫“朱利安A”和“朱利安B”。

Julian A, the one who moved, apparently took up rowing within four months. It was a detail so specific and so unlike him that he read it aloud to Noa twice, each time in a different voice of disbelief. "I have never rowed anything in my life," he said. "Why does it think I row now?"

搬走的朱利安A，似乎在四个月内就迷上了划船。这个细节如此具体，又如此不像他，他把它大声读给诺亚听了两遍，每次都换一种难以置信的语气。“我这辈子从没划过任何东西，”他说，“它为什么觉得我现在划船了？”

"Confidence intervals," Noa said, not looking up from her own copy. "Somewhere in your data there's a correlation between relocation and joining one (1) sport. You're statistically a rower now. Congratulations."

“置信区间，”诺亚说，眼睛没离开自己手里的那份，“在你的数据里，某个地方有一个‘搬家’和‘加入一（1）项运动’之间的相关性。从统计上看，你现在是个划船运动员了。恭喜。”

It would have been easy to dismiss the whole report as expensive nonsense on the strength of the rowing alone. What kept Julian reading past the joke was a paragraph further down, unfunny in a way the rowing hadn't prepared him for. Julian A's projected contact frequency with his father dropped to roughly one call a month by month nine, against a modeled baseline of one call a week for Julian B. The model didn't moralize about it. It simply reported the number, the way it had reported the rowing, as if both facts belonged to the same category.

光凭划船这一条，就很容易把整份报告当成昂贵的胡扯。可真正让朱利安在笑话之后继续读下去的，是再往下的一段，那一段不好笑，是划船那个玩笑没法让他预先做好准备的那种不好笑。朱利安A预测的和父亲联系的频率，到第九个月降到了大约一个月一通电话，而模型设定的基线是，朱利安B一周一通。模型没有对此作任何说教。它只是报告了这个数字，就像报告划船一样，仿佛这两件事属于同一个类别。

Had the simulation weighted every input equally, it would have recommended Lisbon without hesitation — higher salary, wider social network, a nine-percent higher projected life-satisfaction score by month twelve. It did recommend Lisbon, in fact, in a summary line at the very top that Julian had read first and then spent the rest of the evening quietly trying to argue with.

要是这次模拟对每项输入都一视同仁，它会毫不犹豫地推荐里斯本——更高的薪水，更广的社交圈，到第十二个月预计生活满意度高出百分之九。事实上，它确实推荐了里斯本，在最上方的一行总结里，朱利安最先读到的就是这一行，然后那天晚上剩下的时间，他都在悄悄地试着跟它争辩。

"It's not wrong about the money," he admitted finally.

“它说钱的事没错，”他终于承认。

"No," Noa agreed. "It's just not weighing your dad's chemo the way you would."

“没错，”诺亚表示同意，“只是它没像你那样去衡量你爸爸的化疗。”

That, more than anything in either document, was the sentence that settled it. The model had no way of knowing that a monthly call was not, for Julian, interchangeable with four weekly ones divided differently across the calendar. Presence had a texture numbers didn't capture — one his father would notice missing long before any dataset would. He turned down the transfer the following morning and emailed his regrets before he could reread Julian A's suspiciously enviable rowing physique one more time. Over breakfast, he told Noa that he was choosing the version of the year no algorithm had rated highest.

比起两份文件里的任何内容，这一句，才是最终让他下定决心的话。模型无从知道，对朱利安来说，一个月一通电话，并不能等同于把四个每周一次的电话换一种方式分摊在日历上。陪伴有一种数字没法捕捉的质地，他父亲会远比任何数据集更早察觉它的缺失。第二天早上，他谢绝了调动，在他又一次重读朱利安A那令人怀疑的、令人羡慕的划船体格之前，就发邮件表达了歉意。吃早餐时，他对诺亚说，他选的是没有任何算法给出最高评分的那个版本的一年。

"Rowing's probably overrated anyway," she said.

“划船大概也被高估了，”她说。

"Statistically," he said, "we'll never know."

“从统计上说，”他说，“我们永远不会知道。”