TARGET
目标事实Target fact
始终保留在上下文中，内容不变。It remains in context unchanged.
01 · 桌面放得下，为什么还是找不到01 · It fits on the desk, so why is it still hard to find
目标事实一直在文档堆里，窗口也没有超限；当八份措辞相近的材料陆续加入，模型还能稳定找到它吗？The target fact remains in the document pile and the window never overflows. After eight similarly worded documents arrive, can the model still find it reliably?
02不断加入干扰文档，看检出率曲线下降并尝试三种恢复策略Add Distractor Documents, Watch Detection Fall, and Try Three Recovery Strategies
高阶 · 信息衰减expert · information decay
状态 AMetric A
—
—
状态 BMetric B
—
—
状态 CMetric C
—
—
所有概率、命中率、得分与输出均为固定种子的教学模拟，不代表真实模型。拖回主变量即可复核结论。All probabilities, hit rates, scores, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.
03把刚才的变化拆成 4 个零件Break the Change into Four Parts
Context RotDISTRACTORS
干扰文档Distractors
措辞相近但与问题无关的材料逐渐增加。Similarly worded but irrelevant material accumulates.
DETECTION
检出率Detection rate
固定种子教学曲线衡量目标事实被使用的概率。A fixed-seed teaching curve tracks target-fact use.
RECOVERY
恢复策略Recovery strategy
删减、重排、摘要改善信息密度。Pruning, reordering, and summarizing improve information density.
04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction
MECHANISM
usefulness ≠ context_capacity; detection = f(relevance_density,
order, compression)
干扰数是唯一主变量；恢复按钮只改变有效信息密度或位置，因此蓝色恢复点可与原始红线直接比较。Distractor count is the main variable. Recovery changes only
density or order, so the blue recovered point can be compared
directly with the red baseline.
上下文增长Context grows加入相似干扰similar distractors accumulate
利用率下降Use degrades检出曲线下滑detection curve falls
主动整理Curate删减 / 重排 / 摘要prune / reorder / summarize
05加入 8 份干扰，但保持“无恢复”Add Eight Distractors with No Recovery
主动制造失败Create a failure边界实验Boundary experiment
所有内容仍在容量内，却不能稳定找到目标事实。这个失败专门区分容量上限与信息利用。Everything still fits, yet the target fact is no longer used reliably. This failure separates capacity from utilization.
尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction
first.
✗ 窗口没超限，就不会腐化。No overflow means no context rot.→ ✓ 利用准确性可在容量内随干扰增长而下降。Accurate use can decline within capacity as distraction
grows.
✗ 把所有材料摘要后一定恢复。Summarizing everything guarantees recovery.→ ✓ 摘要也可能丢掉细节和来源边界。Summaries can lose detail and provenance
boundaries.
✗ 一次曲线下降就证明所有模型都一样。One declining curve proves every model behaves
identically.→ ✓ 曲线是固定种子教学模拟，真实系统必须独立评测。This is a fixed-seed teaching simulation; real systems require
their own evals.
06 · 一句话带走06 · One line to keep