Software Factory Archive

The conference room was too cold, the projector was slightly out of focus, and Navan was presenting to approximately sixty people who had chosen his talk over "Scaling Kubernetes Service Meshes" in the room next door. He tried not to think about the fact that the service mesh talk probably had better attendance.

"The Digital Twin Universe," he said, advancing to his second slide, "is a collection of behavioral clones of third-party services. Okta, Jira, Slack, Google Docs, Google Drive, Google Sheets. We use them to test at scale without depending on real APIs."

He could see the skepticism settling over the audience like fog. He'd expected it. He pressed on.

He explained the architecture. Each twin modeled the behavioral surface of its corresponding service—not the internal implementation, but the observable behavior. API responses, error conditions, state transitions, timing characteristics. He showed diagrams. He showed code. He showed the satisfaction metrics before and after the DTU was deployed.

The question came during Q&A, from a man in the third row wearing a company t-shirt Navan didn't recognize. "This is interesting," the man said, in the tone of voice that meant he didn't think it was interesting. "But aren't these just fancy mocks?"

The room shifted. Navan could feel it—a collective leaning-forward, the audience scenting blood. This was the question everyone had been thinking. He'd known it was coming. He'd practiced the answer in the mirror at the hotel that morning, and then again in the elevator, earning a concerned look from a woman carrying a tote bag full of vendor swag.

"A mock returns what you tell it to return," Navan said. "You define the inputs and the outputs, and the mock faithfully replays them. If you define ten cases, the mock handles ten cases. The eleventh case, the one you didn't think of, fails silently or throws an error that doesn't match what the real service would do."

He advanced to a slide showing two graphs side by side. One showed a mock's coverage: a flat plateau at whatever coverage the developer had manually defined. The other showed the Okta twin's behavioral coverage over time: a curve that climbed as new edge cases were discovered and modeled.

"A behavioral clone doesn't return what you tell it to return. It returns what the service would return, based on a model of the service's actual behavior. When we discover a new edge case—like, say, how Okta handles group renames during active sessions with claim transformations—" He saw a few knowing nods. Good. "—we add that behavior to the clone. The clone gets more accurate over time. A mock stays exactly as limited as the day you wrote it."

He paused. The room was quiet. The skeptic in the third row was still listening.

"Mocks are assertions about what a service does," Navan said. "Behavioral clones are theories about what a service is. Assertions are brittle. Theories can be refined. That's the difference. And in our experience, it's the difference between a test suite that gives you confidence and a test suite that gives you certainty."

The silence held for a beat. Then someone in the back row started clapping, which was unusual for a technical talk and mildly embarrassing. But it spread. Not a standing ovation—this wasn't that kind of conference—but a genuine, sustained applause that told Navan his answer had landed.

The skeptic in the third row was clapping too. After the talk, he came up and asked for the slides. Navan gave him something better: the URL to the factory's published research.

"Read the section on the Okta twin," Navan told him. "Start with the group rename incident. That's where it clicks."

He walked out of the conference room into the hallway, where Jay was waiting with two cups of coffee.

"How'd it go?" Jay asked.

"Someone asked if they were fancy mocks."

"And?"

Navan took the coffee. "I changed their mind."