Token Reservation

Guaranteeing each sub-question a share of the context, so a compound query's better-matching half cannot spend the whole budget on itself.

Why it matters in a memory layer

It is the same argument score fusion makes about slots, one level down: ranking orders within a question and something has to allocate between them. On this corpus the guarantee turns out to be inherited — the retrieval merge already gives each sub-question its best answer — which leaves the interesting problem as padding: a second answer to a question already answered, spending tokens the other question still needs.

Connections

Taught in: The Packing Problem

Used in: What Must Never Be Dropped · Ordering and Formatting · Does This Earn Its Tokens?