Software Factory Archive

The whiteboard had been there since before Jay and Navan joined. It hung on the wall behind Justin's standing desk, a four-by-three-foot expanse of white surface that, in any normal office, would have been covered with sprint plans and architecture diagrams and passive-aggressive notes about cleaning the kitchen.

Justin's whiteboard had one thing on it. An equation. Or something that looked like an equation. It had Greek letters and summation notation and a few symbols that Jay was fairly certain Justin had invented, because they didn't appear in any mathematical notation Jay had ever encountered.

The equation changed. Not often—maybe once every two or three weeks. Justin would stand in front of it during a quiet moment, erase a term, add a different one, stare at it for several minutes, and then go back to whatever he'd been doing. He never explained what he'd changed or why.

Jay first asked about it in August. "What's the equation?" he said, nodding toward the whiteboard.

"It's the satisfaction function," Justin said.

"The satisfaction metric? We have that in code already."

"The code is the implementation. This is the idea."

Jay waited for more. More did not come.

Navan tried in September. He took a different approach: he photographed the whiteboard, transcribed the equation into LaTeX, and tried to work through it. He got about two-thirds of the way before hitting a symbol that looked like a lowercase phi with a subscript that was either a "t" or a "tau" or possibly just a smudge.

"Justin, what's this symbol?" Navan asked, showing him the photo.

"That's the trajectory weight."

"Weight of what? Against what?"

"Against the probability that the trajectory represents a real user's experience versus an artifact of the scenario's construction."

Navan blinked. "You're weighting trajectories by how realistic they are?"

"By how realistic they probably are. It's probabilistic."

"Everything with you is probabilistic."

"Everything is probabilistic. I just write it down."

By November, the equation had grown. A new term had appeared on the right side involving what looked like a product of conditional probabilities nested three levels deep. Jay and Navan had started a private Slack channel—on the twin, not the real Slack, because they were paranoid about rate limits now—called #whiteboard-watchers, where they posted photos of each change and tried to reverse-engineer what Justin was thinking.

Their best theory, as of January, was that the equation described a way to measure user satisfaction that accounted not just for whether scenarios passed, but for the probability distribution over all possible user behaviors, including behaviors no scenario had ever modeled. It was a metric that tried to answer: "If a user did something we've never thought of, what's the probability they'd still be satisfied?"

It was, Jay admitted to Navan late one evening, either the most ambitious quality metric ever conceived or the most elaborate piece of office decoration in Silicon Valley history.

"Could be both," Navan said.

In early January, Justin erased the entire right side of the equation and rewrote it in a different form. Jay photographed it. The channel went wild—which was a strange thing to say about a channel with only two members, but the volume of messages was impressive.

The new form was simpler. Cleaner. Three terms where there had been seven.

"He's converging," Navan said, and neither of them was entirely sure if he meant Justin or the equation.

The whiteboard is still there. The equation is still changing. Some things at the factory take longer than a single night to resolve, and that's fine. Not everything needs to compile on the first pass.