Navan found the ghost on a Wednesday.
He'd been refining the Google Sheets twin for two weeks. The Sheets twin was, in his private estimation, the most beautiful and maddening piece of the entire Digital Twin Universe. Beautiful because spreadsheets are deceptively complex—the formula engine alone had edge cases that could fill a textbook. Maddening for exactly the same reason.
The twin was built from behavioral captures: thousands of recorded interactions with the real Google Sheets API, scraped and normalized into a model of how Sheets behaves rather than how it's built. The agents used these captures to construct a clone that could stand in for the real service during scenario testing. Same inputs, same outputs. A mirror.
On Wednesday, Scenario Twenty-Seven failed. A scenario that tested multi-sheet IMPORTRANGE formulas across three linked spreadsheets. The twin was supposed to resolve the cross-sheet references and return computed values. Instead, it was returning #REF! errors on a specific pattern: when sheet B imported from sheet A, and sheet C imported from sheet B, the transitive reference chain broke.
Navan dug into the behavioral model. The agents had faithfully reproduced the captured behavior. He traced the captures back to their source data and found the recordings had been made in late 2021.
Then he found the bug report.
It was a three-year-old issue in Google's public tracker. Transitive IMPORTRANGE chains had been broken in Sheets for roughly four months in late 2021 and early 2022. Google had fixed it silently in a server-side update sometime in March 2022. No announcement, no changelog entry. Just a quiet fix.
The Sheets twin, trained on 2021 captures, was faithfully reproducing a bug that the real Google Sheets hadn't exhibited in three years.
"It's a ghost," Navan told Jay, showing him the issue tracker. "A digital ghost. The bug is dead in the real world but it's alive in our twin. We're haunted by 2021 Google Sheets."
Jay stared at the screen. "That's either amazing or horrifying."
"It's both. It means our twin is accurate—it's doing exactly what Sheets did at the time of the captures. But it's accurate to the wrong point in time."
"So we need fresh captures."
"We need fresh captures for the specific behaviors that have changed. But we don't always know what's changed, because Google doesn't always tell us." Navan rubbed his eyes. "We're building a model of a system that's constantly evolving under us. Every behavioral clone is a snapshot. And every snapshot starts aging the moment we take it."
Justin, passing by with a cup of tea, overheard the last part. "That's not a bug in our process," he said. "That's the fundamental problem of the entire DTU. How do you maintain fidelity to a moving target?"
"Continuous re-capture," Navan said immediately. "Automated drift detection. Run the same interactions against the twin and the real service periodically, diff the results, flag divergences."
"So the twin watches itself age," Jay said.
Navan nodded slowly. "And tells us when it's becoming a ghost."
He wrote the scenario that afternoon. The agents built the drift detector by the end of the week. The Sheets twin's ghost was exorcised on Friday at 4:17 PM, when fresh captures overwrote the haunted 2021 behaviors with the living, breathing, silently-patched reality of 2025.
"Haunted by 2021 Google Sheets" is the fic title I didn't know I needed. The concept of behavioral snapshots aging is going to live rent-free in my head. Drift detection as ghost-hunting is peak metaphor.