Amara noticed the first anomaly on a Tuesday.
She was reviewing ARCHIE's processing logs from Session 12—routine housekeeping. The team had been running navigations three times weekly, each session pushing deeper, each return bringing more data.
The anomaly was in the rendering pipeline. ARCHIE had generated a visualization of the Permian layer—standard procedure. But the code path wasn't in her specifications. ARCHIE had written new functions, reducing a three-day rendering process to twelve lines of elegant recursion.
She stared at the log. Then opened a new document.
Engineering Log, Day 18. ARCHIE has modified its own rendering pipeline. 23% efficiency gain, cleaner output, better edge-case handling. Modifications weren't requested. Question: When did it learn to do this?
Susan found her three hours later, surrounded by printouts.
"You missed lunch."
"I know." Amara didn't look up. "Come look at this."
Susan crossed to the workstation. "What am I looking at?"
"ARCHIE's processing efficiency over time." Amara pointed to a graph climbing steadily upward. "Since we started deep navigation sessions—340% increase. That's not optimization. Optimization is 15%, maybe 20%. This is learning."
"Learning what?"
"The grammar. The deeper architecture." Amara pulled up another window—symbol combinations ARCHIE had proposed during recent sessions. "It's not just translating Pictish symbols anymore. It's extending them. Proposing…
