PROMPT
同一输入Same prompt
只有输入一致，回答间比较才有共同问题。Responses are comparable only when they answer the same prompt.
01 · 两位裁判都说“我选更好”01 · Two judges both say “I picked the better one”
一位重视安全，另一位重视简洁；他们对同三份回答排出不同顺序，哪一份才是训练用的“标准答案”？One judge prioritizes safety, another brevity. They rank the same three responses differently—so which ranking should become training data?
02分别给三份回答排序，让分歧矩阵与最终偏好对实时显现Rank Three Responses Twice and Reveal the Disagreement Matrix
两标注者 · 成对矩阵two annotators · pair matrix
状态 AMetric A
—
—
状态 BMetric B
—
—
状态 CMetric C
—
—
所有概率、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, 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
Preference DataRANK
回答排序Response ranking
每名标注者把 A、B、C 排成有序序列。Each annotator orders responses A, B, and C.
CRITERIA
偏好准则Preference criteria
安全、正确、简洁等准则决定“更好”的含义。Safety, correctness, and concision define what better means.
PAIR
偏好对Preference pair
一致的成对比较可形成 chosen / rejected。Agreed pairwise comparisons can form chosen/rejected pairs.
04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction
MECHANISM
usable_pair = same_prompt ∧ explicit_criteria ∧
annotator_agreement
矩阵逐对比较
A/B、A/C、B/C；红格不是错误答案，而是需要解释或复标的准则分歧。The matrix compares A/B, A/C, and B/C. A red cell is not
automatically wrong; it signals a criteria gap that needs
explanation or re-annotation.
固定输入Fix prompt三份候选回答three candidate responses
独立排序Rank independently两名标注者two annotators
对齐准则Resolve criteria保留可解释偏好对retain explainable pairs
05让两名标注者给出完全逆序Make the Two Annotators Give Exact Opposite Rankings
主动制造失败Create a failure边界实验Boundary experiment
完全逆序时没有可静默合并的多数结论。必须补齐准则、仲裁或复标，而不是任取一人的排序。Exact opposite rankings have no safe silent merge. Add criteria, adjudicate, or re-annotate instead of selecting one person's order.
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.
✗ 偏好数据就是客观真值。Preference data is objective ground truth.→ ✓ 它记录特定人群、时点与准则下的行为偏好。It records behavioral preferences for a particular population,
time, and rubric.
✗ 两人不同就删掉所有数据。Any disagreement makes all data useless.→ ✓ 分歧可暴露准则缺口，并指导仲裁或分层建模。Disagreement can expose rubric gaps and guide adjudication or
stratification.
✗ 只保留最终排序，不必记录准则。The final order is enough; criteria can be
discarded.→ ✓ 没有准则就无法解释、复现或审计“更好”。Without criteria, better cannot be explained, reproduced, or
audited.
06 · 一句话带走06 · One line to keep