A routine reconciliation used more quota than I expected on November 4. I asked why, because the setup already included instructions to delegate suitable work. If those instructions were supposed to help control usage, I needed to understand what had actually run.
The investigation reported long context and coordination happening inline, along with failed routing hooks and a decision service that had been failing silently. That widened the question beyond the model answering the main thread. The surrounding system could change how much work reached it and how much context it received.
The analysis included token activity, but I couldn’t treat that as a verified bill. Read and cache activity were part of the totals. Nor would a lower total by itself establish how much time I saved. I wanted to fix the behaviour without attaching a precise savings percentage that the records couldn’t support.
I asked for an independent Sol 6.1 medium review before changing the system. I also rejected a proposed setting that would load more context automatically. I wanted routing that I could edit later, and preferred to use the existing decision service rather than create another substitute that hadn’t been trained for the role.
My preferences for Gemini changed that day too. I asked to limit its role to direct, fully specified commands and keep more consequential work with other choices. That was a response to reliability in my workflow. It wasn’t a controlled comparison of the whole model family, and I wouldn’t use it as one.
Later, I explored a narrower writing-cleanup role with a fallback and short requests. Those two decisions fit together: reducing a model’s autonomy still leaves room to test it on a bounded job whose output can be checked. The live evaluation for that writing role remained deferred, so it was still a proposed use.
The recovery work reported visible offline warnings, restored delegation nudges and live route probes. Some runtime behaviour remained unconfirmed. I needed that distinction in the record, because an instruction saying delegation should happen had already proved insufficient to explain the actual usage. The next sessions would have to show whether the repaired routes were being used as intended.