Context Prioritization — References

Ranking theory from information retrieval, the fusion problem, and the evidence that position within a context changes how material is used.

Why position is part of the decision

• Lost in the Middle: How Language Models Use Long Contexts (Liu et al., 2023), arXiv:2307.03172 — the measured U-shaped relationship between a passage's position in a long context and the model's ability to use it, which is the reason ordering is specified rather than inherited.
• Murdock, 1962, on the serial position effect in free recall (Journal of Experimental Psychology) — the original human result that gives the primacy and recency vocabulary now used loosely about models. Cited by name, as the source of the terms rather than as evidence about model behaviour.

Ranking and fusion

• Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods (Cormack, Clarke and Buettcher, SIGIR 2009) — the rank-space combination method and the demonstration that it beats score-space combination without needing any calibration between the systems being merged.
• Robertson, 1977, on the probability ranking principle in information retrieval (Journal of Documentation) — the result that ranking by probability of relevance is optimal, and explicitly the independence assumption it rests on, which is exactly the assumption a context window of overlapping passages violates. Cited by name.
• The Use of MMR, Diversity-Based Reranking for Reordering Documents and Producing Summaries (Carbonell and Goldstein, SIGIR 1998) — the standard formulation of trading relevance against novelty when the selected set, rather than each item, is what matters.

No verified reference is recorded for cross-class allocation between retrieved documents, agent memory, tool output and dialogue history, or for shadow replay of a context selection policy. The retrieval literature ranks within a single corpus; the heterogeneous case is engineering practice.