By September, Navan could predict Justin's lunch order with 94% accuracy. Thai food. Specifically, pad thai from the place two blocks east, with medium spice and no peanuts. The 6% margin covered the occasional deviation to a noodle soup on cold days, which in the Bay Area meant days below sixty-five degrees, which meant almost never.
Navan's own order was similarly predictable: a burrito from the taqueria on El Camino. Carne asada, black beans, extra salsa, no sour cream. He ordered the same thing so consistently that the staff had started preparing it when they saw him walk in. He found this efficient rather than embarrassing.
Jay was the outlier. Jay rotated. Monday might be pho. Tuesday might be a sandwich from the deli. Wednesday might be leftovers from something he'd cooked the night before, appearing in a glass container with a blue lid. Thursday might be sushi from the grocery store—the kind with the green plastic grass, which Jay called "aesthetically committed." Friday might be nothing at all, because he'd gotten absorbed in a Hacker News thread and forgotten to eat.
The variance was notable. In a team of three, two were deterministic and one was stochastic. Navan, who tracked things, tracked this. In his notebook, on a page he'd labeled Lunch Entropy, he had a grid showing each person's orders for the past six weeks. Justin's column was a solid block of "Thai." Navan's was a solid block of "Burrito." Jay's was a sequence that defied pattern recognition: Pho, Sandwich, Leftover, Sushi, Skip, Thai, Sandwich, Pho, Pizza, Leftover, Skip, Burrito.
"Your lunch entropy is very high," Navan told Jay one afternoon.
"Thank you," Jay said.
"I didn't mean it as a compliment."
"I'm taking it as one."
The irony was not lost on any of them. They had built a system—a factory, a collection of agents and twins and scenarios—designed to predict and validate complex behavioral patterns across enterprise software. The agents could model Okta's authentication flow with probabilistic precision. The satisfaction metrics could tell you whether a user's journey through six interconnected services would likely succeed. The system could predict behavior at scale with remarkable accuracy.
And yet nobody wrote a scenario to predict Jay's lunch order. Not because the agents couldn't do it—with enough historical data, even Jay's apparent randomness might reveal a pattern. But because some things are better left unpredicted. Some forms of entropy are features, not bugs. Jay's lunch rotation was one of them.
"I could build a model," Navan offered once.
"Please don't," Jay said.
"It would need at least six more weeks of data."
"Absolutely not."
Justin, eating his pad thai, said nothing. But his satisfaction metric, if anyone had been measuring, was very high.
"Some forms of entropy are features, not bugs." That's it. That's the whole post. Also the page in Navan's notebook labeled "Lunch Entropy" is very in character.