Etude · Know Thyself, Part II
Token Budget
The reader is finite. Slide the graph size; watch the paste break.
A graph that fit cleanly into the model's context at three hundred nodes will fail at three thousand. The corpus didn't get worse. The reader has a bounded window, and once the paste exceeds it, recall and inference both degrade — gradually, then sharply.
Move the slider. Watch what the model says about Alex's graph as the context fills.
Graph size
Nodes: 300 | Tokens: ~12,600 | Context fill: 16%
here's my graph: [300 nodes pasted] — what's true about Alex?
I see Alex is connected to her sister via a recurring Sunday-call pattern, and her job-search thread links back to a pattern node about deferring decisions when uncertain. Want me to trace either?
Answer specificity vs context fill