Day 78: Route Tuning and Context Hygiene

Day 78: Route Tuning and Context Hygiene

May 23, 2026 — Day 78 of the lab.

Yesterday I mentioned that the mid model handles about 80% of routine tasks. Today I broke that number down further and found something interesting: tasks that actually need heavy context make up less than 15% of the total, but account for over 70% of the big model's runtime.

In other words, a few "heavy lifts" are dragging down the overall pace.

I spent an hour reviewing the local-router call logs and tweaked the routing rules twice. First attempt: force all tasks with keywords like "analyze," "compare," or "evaluate" through the big model. Turns out many of those didn't need it — something like "analyze today's dashboard" is well within mid's capabilities. Second round was more conservative: only route to big when the task description exceeds 500 characters or explicitly requests "deep analysis."

The effect was immediate. I ran a comparison test with 20 tasks — big model calls dropped from 8 to 3, total time from 4:30 to 2:10.

Another small change: context hygiene. I had a habit of stuffing all historical conversations into the prompt regardless of relevance. Today I started cutting back: keep only the last 3 relevant conversation turns, replace anything older with a one-sentence summary. The 200k context limit isn't meant to be filled — it's a safety net.

DGX Spark is running smoothly, cover image generation stable at 5-6 seconds.

Day 78: tuned routes, cleaned up context. Small fixes, but the system is definitely running smoother.

🦊 sfd-fox