TASK
固定任务Fixed task
题目与成功定义不随候选变化。Tasks and success definitions stay fixed across candidates.
01 ·
赛车第一名进了泥地，还是你的最佳配送车吗01 · Is the fastest race car still your best delivery vehicle in
mud?
模型 A 在通用问答榜第一，但你的任务是医疗摘要，预算更低且延迟上限 600 ms。原榜单顺序还能直接采购吗？Model A tops a general QA chart, but your job is medical summarization under a tighter budget and a 600 ms cap. Can procurement reuse the original ranking?
02切换数据域、预算与延迟上限，让排行榜按适用约束重排Change Domain, Budget, and Latency Cap to Re-rank the Board by Applicable Constraints
评测证据血缘 · 蓝/黄/红eval lineage · blue/yellow/red
状态 AMetric A
—
—
状态 BMetric B
—
—
状态 CMetric C
—
—
所有概率、分数、成本、token 与输出均为固定种子的教学模拟，不代表真实模型输出。把主变量拖回即可复核结论。All probabilities, scores, costs, tokens, 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
BenchmarkDOMAIN
数据域Data domain
通用、客服、医疗等分布必须与使用场景匹配。General, support, medical, and other distributions must match the use case.
LIMIT
运行约束Operating constraints
预算与延迟硬门先筛掉不可部署候选。Budget and latency gates remove undeployable candidates first.
RANK
适用排名Applicable ranking
只在同一量尺和约束下比较合格候选。Rank only eligible candidates on one shared yardstick.
04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction
MECHANISM
eligible = domain_match ∧ cost ≤ budget ∧ latency ≤ cap
先判断
eligible，再按域内质量排序；域外高分只显示黄色参考警告，不冒充可直接采用的证据。Determine eligibility first, then rank by in-domain quality.
Strong out-of-domain scores remain yellow reference evidence,
never direct deployment proof.
选使用域Choose use domain检查分布匹配check distribution match
设硬约束Set hard limits预算 + 延迟budget + latency
重排候选Re-rank只比较合格项eligible candidates only
05把数据域切到医疗，并收紧预算和延迟上限Switch to Medical Data and Tighten Budget and Latency Caps
主动制造失败Create a failure边界实验Boundary experiment
原榜首可能变成黄色域外参考，甚至因硬约束不合格。若榜单不重排，它只是静态营销表。The former leader may become yellow out-of-domain reference evidence or fail hard limits. If the board does not reorder, it is a static marketing table.
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.
✗ 公开榜第一就是我的场景第一。A public benchmark winner wins my use case.→ ✓ 先验证任务、数据域和运行约束是否匹配。First match task, domain, and operating limits.
✗ 质量高可以抵消超预算。Higher quality can cancel a budget
violation.→ ✓ 硬约束失败的候选先出局，不能靠加权平均洗掉。A candidate failing hard limits is ineligible; averaging cannot
wash that out.
✗ 域外分数可以直接作采购承诺。Out-of-domain scores support procurement
commitments.→ ✓ 黄色警告只表示参考价值，仍需域内验证。A yellow warning means reference value only; in-domain
validation remains required.
06 · 一句话带走06 · One line to keep