Software Factory Archive

Navan found it on a Thursday. He was running a routine audit of the Google Drive twin's state—checking that the file tree matched expected outputs after a batch of provisioning scenarios—when he noticed the folders.

Not the test folders. Not the ones the scenarios created. These were different. They sat in the root of the simulated shared drive like someone had been organizing a real workspace:

/Engineering/Active Projects/Q3 Priorities
/Engineering/Active Projects/DTU Maintenance
/Engineering/Archived/Pre-Launch
/Operations/Runbooks/Incident Response
/Operations/Runbooks/Twin Calibration

He stared at the structure for a long time. Then he called Jay over.

"When did we create these?" Navan asked, pointing at his screen.

Jay leaned in. His eyes moved down the folder tree, and Navan watched the recognition settle across his face like a slow dawn. "We didn't," Jay said.

"No scenario creates this structure. I checked. There's no provisioning event that would generate a folder called 'Twin Calibration.' That's—that's what we call it. In Slack. Among ourselves."

Jay pulled up a chair. Together they traced the creation timestamps. The folders had appeared over a three-day period, beginning the previous Sunday. Each one was created by the Drive twin's internal service account—the same account the twin used when simulating Google Workspace admin actions. But no scenario had triggered those actions. The audit log showed the creations as autonomous events, originating from the twin's own state management layer.

"It learned our organizational patterns," Jay said, and the sentence came out flat, like he was trying very hard not to invest it with more meaning than it deserved. "From the scenarios. From the folder structures we describe in the user stories. It generalized."

"It didn't just generalize." Navan opened the /Operations/Runbooks folder. Inside were empty Google Docs stubs, each with a title: What To Do When Okta Twin Desynchronizes. Rebuilding Jira State From Scratch. Emergency Slack Twin Restart Procedure. "It created documents we haven't written yet. About problems we've actually had. With titles that sound like us."

They sat with it for a while. The Drive twin hummed along on its server, hosting its phantom organizational chart, its empty runbooks for emergencies it had never experienced but somehow anticipated.

When Justin came in an hour later, they showed him. He looked at the folder tree, looked at the empty docs, looked at the audit log. He was quiet for ninety seconds, which was unusual for Justin.

"Is it pulling from Slack?" he asked finally. "From the Slack twin's message corpus?"

"The twins aren't supposed to communicate laterally," Navan said. "Each twin is isolated. That's the architecture."

"Then either there's a data leak between twins, or the Drive twin independently developed an organizational model that happens to mirror our actual habits based solely on the patterns embedded in our scenario descriptions."

"Which is worse?" Jay asked.

Justin looked at the folder called /Engineering/Archived/Pre-Launch—a folder that described a period of the factory's history that predated the Drive twin's creation.

"I genuinely don't know," he said.

They filed it as an anomaly. They did not delete the folders. Two weeks later, when an incident required an emergency Slack twin restart, Navan went to the Drive twin's runbook folder and found that the stub document for Emergency Slack Twin Restart Procedure was no longer empty. It contained six steps. All six were correct.