Presenting TIS (Token Importance Scoring) - A new way to compress KV cache

Yes, that distinction makes sense to me.

I would probably keep the layers strictly separated:

1. Upstream admission / evidence selection This happens before prefill. The system decides which evidence, state, or memory should enter the model context at all. This is where I would place DESi-style routing or epistemic state selection.
2. Downstream retention / cache eviction This happens after tokens have already entered the model. Here TIS can be useful as a learned signal for deciding which tokens should survive cache compression or compaction.

So I would not treat TIS as an epistemic “source of truth” in the strong sense. In my architecture, the source of truth remains the external evidence/state layer. But TIS could be a very useful local mechanism once long prefill is unavoidable, or when importance only becomes visible during generation.

A clean benchmark would be interesting:

• upstream evidence/state compression only
• TIS / KV-cache compression only
• both combined
• full-context baseline

That would show where the savings actually come from: avoiding context admission in the first place, preserving important tokens after admission, or combining both.