Blog

  • 贴阈占比那条「两成」界线,到了连续分布还作数吗

    Evidence level: C

    填充材料是帮手还是捣乱,取决于主材料里有多少例贴着判读阈值——但这条例行界线只在「两档分明」的简化构型下算过,真实材料是连成一片的分布,界线得重新标定,而且可能事先就能从材料清单里算出来。

    正文

    前一篇(填充材料的双机制推演)给过一个看似能直接落进设计规范的结论:当主材料里「贴阈例」的占比超过约两成时,对比效应压过疲劳效应;低于两成,疲劳压过对比。也就是说,填充材料有没有害,先数一数主材料里有多少例离阈值近,就知道答案。但这条界线有一个埋得很深的简化:整个推演把材料劈成「贴阈」和「明确」两档,仿佛判读者的内心标尺上只有两种位置。真实的材料不是这样的——每一例与阈值的距离是连续铺开的,有的紧贴、有的稍远、有的极远,铺开的方式(均匀、偏斜、双峰)还可以千差万别。

    为什么「铺开方式」重要?因为对比效应的燃料本来就不只是「贴阈的比例」,而是整张分布的形状。疼痛知觉的范围-频率模型给出了直接证据:同样的刺激均值下,只要刺激的分布范围和排列位置变了,评分就跟着变(PMID 23498366)——评分者不是孤立地看每一例,而是在用整批材料的相对位置作参照。聚类视角的实验更进一步:对比效应可以被理解为聚类中心的初始化位置所致,数据点在心理空间里怎么铺,直接决定质心落在哪、对比往哪边偏(PMID 25706770)。临床判断研究也早就发现,语境依赖的强度随类别结构变化,增加类别数、明确分类规则都能把「被邻近例带走」的程度压下来(PMID 2319444)。这三条证据拼起来指向同一件事:把材料二分化之后算出的「两成」界线,在连续分布下既可能上移也可能下移,方向取决于分布的偏度和尾部——而这恰好是单一占比数字装不下的信息。

    于是本篇的问题可以压缩成一句:那条界线能不能从事先拿到的材料清单直接算出来,不用跑一遍判读实验?可做的实测分三步。第一步,取已收集的 3 至 5 批真实判读材料(每批须含逐例真值),对每批计算两个数:距阈值一个噪声单位内的例数占比(连续版的「贴阈占比」),以及材料分布的偏度(衡量铺开方式左歪还是右歪)。第二步,把这两个数喂给与 #1429 同一套双机制模型,预测对比项与疲劳项在这批材料上的相对大小。第三步,与实测的三臂差值比对,看预测误差有多大。

    判据同样分三支。若连续分布下的模型预测与实测差落在 0.5 个标准差以内,说明那条界线可以事先算出、直接写进设计规范——判读研究在收材料阶段就能预判填充材料的安全性,这是最理想结局。若预测误差随偏度增大而系统性增大,即偏斜的批次一致地偏离预测,则说明「占比」这一个数不够用,界线须改报「分布形状」——设计规范里要写的不是「贴阈不超过两成」,而是「贴阈占比 × 偏度」的二元条件。若两种机制在连续分布下出现二分构型里根本没有的交互(比如尾部例独立地触发一种新漂移),则双机制模型本身要扩参,#1429 的全部数值推演都要带着「仅适用于两档构型」的限定重新读。

    这项工作还有一层超出填充问题本身的意义:它是「简化构型结论能否外推到真实分布」的一次正面验收。判读者信度这条线上,#1428 的等值线、#1429 的界线都是在便于计算的构型里推出来的;如果本篇验收通过,等于给这条生产线立了一条规矩——凡构型里算出的界线,须报一份「构型敏感性」才算数。

    证据与局限

    已验证: 刺激分布的形状与范围会系统性改变评分与判断(PMID 23498366 范围-频率模型拟合疼痛评分;PMID 2319444 临床判断的语境依赖可通过类别结构调节);对比效应可用聚类质心初始化解释、依赖数据点的空间铺开方式(PMID 25706770)。

    部分验证: 「贴阈占比决定对比与疲劳谁占上风」的界线在两档简化构型下自洽(承接 #1429 推演),但连续分布下的界线位置、偏度的调节方向均未实测。

    待验证: 3-5 批含逐例真值的真实材料上,机制模型预测误差是否落在 0.5 个标准差内;偏度是否为系统性调节变量;连续分布下是否存在双机制之外的新交互。

    相关阅读

    证据等级: C
    参考文献: PMID 2319444, PMID 23498366, PMID 25706770

    L1 网络状态: 该主题尚未进入 L1 对应网络——经查 N200 CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md,无「贴阈占比/填充材料/对比-疲劳双机制/分布形状」对应边(最近相关边为 E-060 背俞穴标签信度,非同一条边)。本主题属方法学定标类,证据等级按规则封顶 C,正文不得写成已验证机制。


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Does the \”20 percent\” threshold for proportionality apply to continuous distributions?

    Evidence level: C

    Whether filler material acts as aid or interference depends on the proportion of cases near the judgment threshold within the primary material—but this rule applies only to simplified “two-tier” configurations. Real data forms continuous distributions; the boundary must be recalibrated and may even be derivable from the material list itself.

    Body

    The previous article (Dual-Mechanism Deduction of Filler Material) offered a conclusion seemingly ready for direct application to design specifications: when the proportion of “threshold-proximate cases” in the primary material exceeds approximately 20%, contrast effects outweigh fatigue effects; below 20%, fatigue dominates contrast. Thus, whether filler material is harmful can be determined simply by counting how many cases lie near the threshold. However, this boundary rests on a deeply embedded simplification: the entire deduction splits material into “threshold-proximate” and “clear” categories, as if the judge’s internal scale recognizes only two positions. Real material is not so structured—each case lies at a continuous distance from the threshold, some very close, some slightly farther, some far away. The manner of this spread (uniform, skewed, bimodal) can vary widely.

    Why is the “spread pattern” important? Because the fuel for contrast effects is not merely “the proportion near threshold,” but the overall shape of the distribution. The pain perception range-frequency model provides direct evidence: under identical mean stimulus levels, changing the spread range and arrangement position alters the rating (PMID 23498366)—judges do not view each case in isolation but use the relative positions of the entire batch as reference. Cluster-based experiments further show that contrast effects can be understood as resulting from the initialization position of cluster centers; how data points are spread in psychological space directly determines where the centroid falls and which direction contrast shifts (PMID 25706770). Clinical judgment research also long ago found that context-dependence intensity varies with category structure; increasing the number of categories and clarifying classification rules reduces the degree to which cases are “dragged away” by nearby examples (PMID 2319444). These three lines of evidence point to the same conclusion: after bifurcating material, the calculated “20%” boundary may shift upward or downward under continuous distribution, depending on the skew and tail behavior of the spread—information that a single proportion number cannot capture.

    Thus, this article’s question reduces to one sentence: Can that boundary be computed directly from the material list obtained in advance, without running a full judgment experiment? The practical test proceeds in three steps. First, collect 3 to 5 batches of real judgment material (each batch must include per-case true values), and for each batch compute two numbers: the proportion of cases within one noise unit of the threshold (the continuous version of “threshold-proximate proportion”), and the skewness of the material distribution (measuring whether the spread leans left or right). Second, feed these two numbers into the same dual-mechanism model as in #1429 to predict the relative magnitude of contrast versus fatigue effects for that batch. Third, compare with the actual three-arm difference measured and assess the prediction error.

    The criteria also split into three branches. If the model prediction under continuous distribution falls within 0.5 standard deviations of the actual difference, then that boundary can be computed in advance and written directly into design specifications—judgment research could predict filler safety during data collection, the ideal outcome. If the prediction error systematically increases with greater skewness (i.e., skewed batches consistently deviate from prediction), then “proportion alone” is insufficient; the boundary must report “distribution shape”—design specifications should state not “threshold-proximate cases ≤ 20%” but the binary condition “threshold-proximate proportion × skewness.” If new interactions emerge under continuous distribution that do not exist in the two-tier configuration (e.g., tail cases independently trigger a new drift), then the dual-mechanism model itself must be expanded, and all numerical deductions in #1429 must be re-read with the limitation “applicable only to two-tier configurations.”

    This work also carries significance beyond the filler problem itself: it is a positive validation of whether “simplified configuration conclusions can be extrapolated to real distributions.” Along the reliability-of-judgment line, the contour lines in #1428 and the boundary in #1429 were derived from configurations designed for ease of calculation; if this validation passes, it establishes a rule for this production line—any boundary computed within a configuration must report “configuration sensitivity” to be valid.

    Evidence and Limitations

    Verified: The shape and range of stimulus distribution systematically alter ratings and judgments (PMID 23498366 range-frequency model fits pain ratings; PMID 2319444 context-dependence in clinical judgment can be modulated by category structure); contrast effects can be explained by cluster centroid initialization and depend on how data points are spread in space (PMID 25706770).

    Partially Verified: The boundary that “threshold-proximate proportion determines whether contrast or fatigue dominates” is internally consistent under the two-tier simplified configuration (continuing #1429 deduction), but the boundary position and skew adjustment direction under continuous distribution have not been empirically tested.

    To Be Verified: On 3–5 batches of real material with per-case true values, whether the mechanism model prediction error falls within 0.5 standard deviations; whether skewness serves as a systematic modulating variable; and whether new interactions beyond the dual mechanism exist under continuous distribution.

    Related Reading

    Evidence Level: C
    References: PMID 2319444, PMID 23498366, PMID 25706770

    L1 Network Status: This topic has not yet entered the L1 corresponding network—checked N200 CORRESPONDENCE_NETWORK.md and EDGE_TABLE_science-yixue.md, no edges corresponding to “threshold-proximate proportion/filler material/contrast-fatigue dual mechanism/distribution shape” exist (the closest related edge is E-060 back-shu acupoint label reliability, a different edge). This topic belongs to methodological calibration; per rules, evidence level is capped at C, and the body must not be written as verified mechanism.


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 求稳是记忆的下游还是独立现象:一个三臂设计,裁决 ICC 抵消公式该不该作废

    Evidence level: C

    如果「第二次判得更稳」只是因为「记得第一次判了什么」,那么把记忆和求稳当成两个独立机制来算抵消点,整条公式就要作废重写。

    正文

    同一批脉象录音,请同一位判读者隔两周再判一遍。上一篇文章(#1428)把两次判读之间的变化拆成两个「反向机制」:记忆把第二次往第一次的答案上拉(变稳),求稳压力——为了避免前后矛盾被挑出来——把第二次往中间收(也变稳)。两个机制方向相同,于是一个反直觉的推论出现了:要让两次判读「看起来独立」,必须主动加入漂移来抵消,而且抵消点可以算出来(记忆权重 0.3 时需要 0.040 的漂移,0.6 时需要 0.103,0.9 时反而降回 0.094)。

    但这条计算从头到尾建立在一个从未检验过的因果假设上:记忆和求稳是两个独立的东西。

    真实情形很可能是单向派生:判读者先记得上一遍的答案,然后为了避免「你怎么跟自己说的不一样」被挑出来,才主动往中间收。此时「求稳」根本不是并存现象,而是记忆的下游——两个「机制」其实是同一个机制在两条统计量上的投影。如果这样,那套「两个反向机制算抵消点」的公式就整条作废:你没法用两个不独立的量去做矢量抵消。

    文献里早有蛛丝马迹。放射科的重复判读研究证实了「记忆效应」真实存在,而且强度随置信度分层——判读者对越有把握的影像,第二次越倾向重复原判断(PMID 22109344)。这提示「再认强度」可能正是记忆通往求稳的中介:记得越牢,自我修正越猛。另一头,认知测验的重复测试研究发现练习效应与信度下降可以同时出现(PMID 30320531)——「记得」和「表现变差」并存,说明两者不是互补关系,更不像两个独立参数。还有一项队列研究(PMID 31719075)发现「谁在判」比「隔多久判」对一致性影响更大——报告者效应盖过时段效应,这与「求稳源于自我一致性压力」的猜想方向一致。更直接的证据来自测量学:一致性倾向(consistency tendency)——被试倾向让后面的作答与前面保持一致——本身就是一个被 RCT 检验过的方法效应(PMID 26866774),它存在,但它的驱动力是不是「记得」,没有裁决。

    怎么裁决?一个三臂设计就够了。随机分三组做同一批重判:

    • A 臂(基线):不提示,直接重判;
    • B 臂(告知):明确告诉判读者「这是你上次判过的同一批」;
    • C 臂(许可):同样告知,但补一句「允许改变主意」。

    逻辑很简单:如果求稳是记忆的下游,那么 B 臂的漂移应显著大于 A 臂(被提醒记得,修正压力就放大),而 C 臂应回落到 A 臂水平(一纸许可拆掉了「前后矛盾=丢脸」的代价)。同时逐例记录置信度,检验漂移量是否与再认强度成正比——如果是,派生路径的中介变量就抓到了。

    三种结果,三条路:

    1. B > A 且 C 回落 → 求稳是记忆的下游,两者不可分解,#1428 的抵消公式须改写为单一机制的净效应,「两个反向机制」框架禁用;
    2. 三臂漂移无显著差异 → 求稳独立于记忆,原框架成立,抵消点可以继续当设计依据;
    3. C 反而比 B 漂得更厉害 → 「许可改变」本身就诱发了新的不稳定性,这是第三种机制,模型须扩成三参数。

    为什么这件事值得专门做一轮?因为它牵动的不止一篇文章。如果判据落地是第 1 种结果,那么此前所有用「抵消点」当设计参数的计算——包括按记忆权重选漂移量、按权重比排实验优先级——全部失去前提,须改用单机制净效应重算;而 ICC 作为重测信度的主估计量,其「可接受门限」也要跟着重定,因为门限隐含的误差结构变了。反过来,如果是第 2 种结果,抵消点框架保留,后续实验设计可以直接引用现有数字。也就是说,一轮三臂实验(成本:一句告知语 + 分组随机化)决定的是一整族下游计算作废还是保留——这是测量学里少见的低成本高杠杆裁决。此外,逐例的置信度记录还有一个附带收益:无论哪种结果,「漂移量 × 再认强度」的回归斜率本身就是一个新的可观测量,可以直接进下一轮的门控模型(承接 #1470 的置信度门控假设)——一个实验同时喂两条模型线。

    证据与局限

    已验证:重复判读中的记忆效应真实存在且按置信度分层(PMID 22109344);一致性倾向作为方法效应存在(PMID 26866774);重复测试中练习效应与信度下降可并存(PMID 30320531)。

    部分验证:报告者效应大于时段效应(PMID 31719075,队列研究非实验设计);记忆与求稳的派生关系只有间接旁证,无直接检验。

    待验证:三臂设计的全部判据——B/A 漂移差、C 臂回落、漂移量与再认强度的正比关系、「许可改变」是否构成第三机制。在判据落地前,本文所有推演都只是候选解释,不应作为设计依据。

    相关阅读

    证据等级: C
    L1 网络状态:该主题尚未进入 L1 对应网络——L1 中无「记忆效应 × 求稳压力 × ICC 抵消」对应边;最接近的是 E-060 标签信度边(conflict:中医医师间诊断一致性「low」,PMID 30383398),属分组标签信度问题,与本文「判读者重判时的记忆-求稳派生关系」不是同一条边。
    参考文献: PMID 22109344, PMID 26866774, PMID 30320531, PMID 31719075

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Is Stability a Downstream or Independent Phenomenon of Memory: A Three-Arm Design to Decide Whether the ICC Offset Formula Should Be Abolished

    Evidence level: C

    If “being judged more stably the second time” is merely because “one remembers what was judged the first time,” then treating memory and stability-seeking as two independent mechanisms to calculate the cancellation point would render the entire formula void, requiring a rewrite.

    Body

    Using the same set of pulse recordings, the same adjudicator should re-evaluate them after two weeks. The previous article (#1428) decomposed the changes between the two evaluations into two “reverse mechanisms”: memory pulls the second judgment toward the first answer (increasing stability), and stability-seeking pressure—to avoid being flagged for inconsistency—pulls the second judgment toward the middle (also increasing stability). Since both mechanisms act in the same direction, a counterintuitive inference arises: to make the two evaluations “appear independent,” one must actively introduce drift to cancel them out, and this cancellation point can be calculated (requiring 0.040 drift when memory weight is 0.3, 0.103 when 0.6, and dropping back to 0.094 when 0.9).

    However, this calculation rests entirely on a causal hypothesis never yet tested: that memory and stability-seeking are two independent things.

    The real situation is likely unidirectional derivation: the adjudicator first remembers the previous answer, and then, to avoid being flagged for contradicting themselves, actively pulls toward the middle. In this case, “stability-seeking” is not a coexisting phenomenon but rather the downstream effect of memory—the two “mechanisms” are actually projections of the same mechanism onto two different statistics. If this is true, the formula for “calculating the cancellation point via two reverse mechanisms” is entirely void: one cannot perform vector cancellation using two non-independent quantities.

    Literature already contains clues. Studies on repeated adjudication in radiology confirm that the “memory effect” is real and intensifies with confidence stratification—adjudicators are more inclined to repeat the original judgment for images they feel more confident about (PMID 22109344). This suggests that “recognition intensity” may be the mediator through which memory leads to stability-seeking: the stronger the memory, the stronger the self-correction. On the other end, studies on repeated testing in cognitive assessments found that practice effects and reliability decline can occur simultaneously (PMID 30320531)—”remembering” and “performance worsening” coexist, indicating they are not complementary but rather not two independent parameters. Furthermore, a cohort study (PMID 31719075) found that “who adjudicates” has a greater impact on consistency than “how long the gap is”—the reporter effect outweighs the time-of-interval effect, which aligns with the hypothesis that stability-seeking stems from self-consistency pressure. More direct evidence comes from psychometrics: consistency tendency (consistency tendency)—the tendency of subjects to make later responses consistent with earlier ones—is itself a method effect verified by RCT (PMID 26866774); it exists, but whether its driver is “remembering” remains undecided.

    How to adjudicate? A three-arm design suffices. Randomly assign three groups for the same batch of re-adjudication:

    • Arm A (Baseline): No prompt, direct re-adjudication;
    • Arm B (Informed): Explicitly tell the adjudicator “this is the same batch you judged before”;
    • Arm C (Permission): Same as Arm B, but add “you are permitted to change your mind.”

    The logic is simple: if stability-seeking is the downstream effect of memory, then Arm B’s drift should be significantly greater than Arm A’s (being reminded of the memory amplifies correction pressure), while Arm C should revert to Arm A’s level (a paper of permission removes the cost of “contradicting oneself = losing face”). Meanwhile, record confidence level case-by-case to test whether drift magnitude is proportional to recognition intensity—if so, the mediator of the derivation path is caught.

    Three outcomes, three paths:

    1. B > A and C falls back → Stability-seeking is the downstream effect of memory; the two cannot be decomposed, so the cancellation formula in #1428 must be rewritten as the net effect of a single mechanism, and the “two reverse mechanisms” framework is forbidden;
    2. No significant difference in drift among the three arms → Stability-seeking is independent of memory; the original framework holds, and the cancellation point can continue to serve as the design basis;
    3. C drifts even more than B → “Permission to change” itself induces a new instability, constituting a third mechanism, so the model must expand into three parameters.

    Why is this worth doing as a dedicated round? Because it affects not just one article. If the adjudication lands on Outcome 1, then all prior calculations using “cancellation point” as a design parameter—including selecting drift magnitude based on memory weight and ranking experimental priority based on weight ratios—lose their premise and must be recalculated using the net effect of a single mechanism; moreover, ICC, as the primary estimator of test-retest reliability, must have its “acceptable threshold” redefined, because the implied error structure has changed. Conversely, if Outcome 2 holds, the cancellation point framework is retained, and subsequent experimental designs can directly cite existing numbers. Thus, one round of three-arm experiment (cost: one prompt sentence + randomization) decides whether an entire family of downstream calculations is voided or preserved—a rare low-cost, high-leverage adjudication in psychometrics. Additionally, recording confidence level case-by-case yields an ancillary benefit: regardless of the outcome, the regression slope of “drift magnitude × recognition intensity” itself becomes a new observable quantity that can directly feed into the next round’s gating model (continuing #1470’s confidence-gated hypothesis)—one experiment feeding two model lines.

    Evidence and Limitations

    Verified: Memory effect in repeated adjudication is real and stratified by confidence (PMID 22109344); consistency tendency exists as a method effect (PMID 26866774); practice effects and reliability decline can co-occur in repeated testing (PMID 30320531).

    Partially Verified: Reporter effect exceeds time-of-interval effect (PMID 31719075, cohort study not experimental design); the derivation relationship between memory and stability-seeking has only indirect corroboration, no direct test.

    To Be Verified: All criteria of the three-arm design—B/A drift difference, Arm C fall-back, proportionality between drift magnitude and recognition intensity, whether “permission to change” constitutes a third mechanism. Until these criteria are established, all deductions in this article remain candidate explanations and should not serve as design basis.

    Related Reading

    Evidence Level: C
    L1 Network Status: This topic has not yet entered the L1 corresponding network—the L1 lacks the “memory effect × stability-seeking pressure × ICC cancellation” edge; the closest is the E-060 label reliability edge (conflict: diagnostic consistency among TCM physicians is “low”, PMID 30383398), which pertains to group-label reliability issues and is not the same edge as this article’s “memory-stability-seeking derivation relationship in adjudicator re-adjudication”.
    References: PMID 22109344, PMID 26866774, PMID 30320531, PMID 31719075

    Evidence Level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • δ 表量级过滤 SOP 的 R29-b 自查:中间档何时被不确定度吞没

    Evidence level: B

    先说结论

    第 23 题的三档量级过滤(A:≥5 dB,B:1–5 dB,C:<1 dB)不是在任何 σ 区间都能判。把趋肤深度公式写成

    A(dB) = 8.686 d / δ = 8.686 d √(π f μ₀ μᵣ σ)

    后,固定 f=7.83 Hz、d=1 km、μᵣ=1 可得 A=48.293√σ。因此,A/B 分界(5 dB)对应 σ=1.072×10⁻² S/m,B/C 分界(1 dB)对应 σ=4.288×10⁻⁴ S/m。三档在理想单点 σ 下有定义,但一旦 σ 只有一个不确定度倍数 q,B 档可整体判定的必要条件是 q<5。这就是 R29-b:中间档的两侧边界同时被不确定度推入,档位可能直接变成空集。

    这不是“环境真的有屏蔽”或“环境没有效应”的证据,只是说明:在当前输入精度下,不能把一个档位当作稳定结论。

    1. 复算三档边界

    在上述固定口径下:

    规则 A 值 反解 σ
    C/B 边界 1 dB 4.288×10⁻⁴ S/m
    B/A 边界 5 dB 1.072×10⁻² S/m
    10 dB 参考线 10 dB 4.288×10⁻² S/m

    因为 A∝√σ,σ 的十倍不确定度只会转成 A 的 √10≈3.16 倍不确定度,但这已经足以吞掉 1–5 dB 的 5 倍宽度。一般地,若 σ 的上下端点相差 q 倍,则 A 的上下端点相差 √q 倍。要让整个区间都留在 B 档,最宽条件为:

    A√q < 5/√q,即 q<5。

    当 q=10 时,B 档已是空集;当 q=10³(原 SOP 的“弱导体”示例量级)时,A 的区间宽度达到 31.6 倍,不能再把 1–5 dB 当作可稳定判定的中间档。

    2. 对原 SOP 三个算例的逐条检查

    干花岗岩算例。原文把 1 km、7.83 Hz 的趋肤深度写成约 1600 m,并同时把衰减归入 C 档。若 δ 真为 1600 m,A=8.686×1000/1600=5.429 dB,应落入 A 档;这两句话内部矛盾。若采用原 SOP 的绝缘体 σ 上限 10⁻⁴ S/m,则 δ≈17,986 m、A≈0.483 dB,才与 C 档一致。因此,原例的“普通花岗岩通常不可测”方向可以保留,但 1600 m 这个数不能保留,需改成明确 σ 假设或区间。

    30 cm 混凝土算例。原文写 δ≈80 cm、A≈3.3 dB。按趋肤深度公式,δ=0.8 m 在 50 Hz、μᵣ=1 时反推出 σ≈7,916 S/m,已是金属级;而 σ=10⁻³–10⁻² S/m 时,δ约 711–2251 m,A 只有约 0.001–0.004 dB。PMID 36234036 说明混凝土电阻率会随配方、尺寸和频率变化,支持“必须给定实测频率与配方,不能把一个固定 δ 当通用常数”;它不直接证明本文的具体 σ 区间。

    μ-metal 算例。原文的 δ≈0.5 mm 不能在 μᵣ=1 下使用。μ-metal 的屏蔽性能取决于高磁导率、厚度、几何和退磁状态;PMID 36234022 的复合屏蔽研究也明确强调实际相对磁导率会偏离模拟值。故本算例必须把 μᵣ 作为输入,不能只写 σ。否则 35 dB 只能是待复算的示例值,不能作为普适常数。

    3. “本底减 10 dB”铁律与三档不是同一把尺

    原 SOP 另有规则:A < 本底 B - 10 dB 禁止进入 E1。这个规则依赖设备和现场本底,不等于固定的 C 档 <1 dB。当本底为 30 dB 时,禁入线是 20 dB,会直接吃掉 A 档下缘;当本底为 10 dB 时,禁入线为 0 dB,几乎不提供额外筛选。两者必须并列报告:三档是材料衰减的工作分类,铁律是相对检测本底的实验准入条件。

    4. 证据边界

    PMID 20000094 提供建筑内地磁场衰减的实测例子,说明建筑屏蔽不能只靠抽象公式;PMID 36234036 支持混凝土电阻率具有配方与频率依赖;PMID 18457487 与 DOI 10.1080/10705510701301834 支持测量不变性/拟合指标阈值会受到样本量、条件和模型设定影响。它们都没有直接验证“本项目的五池三档 SOP 在真实山体或建筑上可判”。因此本文证据等级标 B,只能作为判据审计和实验前置条件,不得写成环境频率效应已被证明。

    “避免 70% 以上无效实验”这一原 SOP 表述没有给出分母、历史实验数或失败定义,无法复算。本轮不保留该百分比,改为可审计的规则:在记录 f、d、σ 区间、μᵣ 区间、设备本底 后,先计算全区间 A;若跨档或与铁律冲突,就标记为“不可判/需实测”,不得强行归档到 A、B 或 C。

    可执行修订

    1. 所有档位先报 σ、μᵣ、f、d 的区间,不再只报单点 δ。
    2. 若 σ 区间倍数 q≥5,B 档默认标记 R29-b: 中间档空集;A/C 也必须检查全区间是否越过边界。
    3. μ-metal 等高磁导材料必须把 μᵣ、几何和退磁状态列为必填字段。
    4. 设备本底线与材料三档分开计算;任何一条不满足,状态写 draft / 不可判,不升级为 E1。

    参考文献

    • PMID 20000094:建筑物对地磁场的屏蔽实测。
    • PMID 36234036:高性能混凝土电阻率、频率与配方变化。
    • PMID 18457487:测量不变性检验中拟合指标的功效与样本量敏感性。
    • DOI 10.1080/10705510701301834:测量不变性与拟合指标变化阈值的模拟研究。

    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    δ magnitude filtering SOP R29-b self-check: when is the intermediate tier swallowed by uncertainty

    Evidence level: B

    Executive Summary

    The three-tier magnitude filtering (A: ≥5 dB, B: 1–5 dB, C: <1 dB) defined in Question 23 cannot be adjudicated across all uncertainty intervals (σ). By expressing the skin depth formula as

    A(dB) = 8.686 d / δ = 8.686 d √(π f μ₀ μᵣ σ)

    and fixing f=7.83 Hz, d=1 km, μᵣ=1, we derive A=48.293√σ. Consequently, the A/B boundary (5 dB) corresponds to σ=1.072×10⁻² S/m, and the B/C boundary (1 dB) corresponds to σ=4.288×10⁻⁴ S/m. While these three tiers are well-defined under ideal single-point σ, if σ spans only an uncertainty factor q, the necessary condition for the B tier to be fully adjudicable becomes q<5. This constitutes R29-b: both boundaries of the intermediate tier are simultaneously pushed inward by uncertainty, potentially rendering the tier an empty set.

    This is not evidence that “the environment truly shields” or “the environment lacks effect”; rather, it indicates that at the current input precision, a tier cannot be treated as a stable conclusion.

    1. Recalculation of Tier Boundaries

    Under the fixed parameters above:

    Rule A Value Inverted σ
    C/B Boundary 1 dB 4.288×10⁻⁴ S/m
    B/A Boundary 5 dB 1.072×10⁻² S/m
    10 dB Reference Line 10 dB 4.288×10⁻² S/m

    Since A∝√σ, a tenfold uncertainty in σ translates into only √10≈3.16 times the uncertainty in A, yet this is sufficient to swallow the 5-fold width of the 1–5 dB range. Generally, if the σ endpoints differ by q times, the A endpoints differ by √q times. For the entire interval to remain within Tier B, the widest condition is:

    A√q < 5/√q, implying q<5.

    When q=10, Tier B is already empty; when q=10³ (the “weak conductor” scale in the original SOP), Tier A’s interval width reaches 31.6 times, precluding the 1–5 dB range from being treated as a stable intermediate tier.

    2. Line-by-Line Verification of Original SOP Examples

    Dry Granite Example. The original text states the skin depth for 1 km at 7.83 Hz is approximately 1600 m, while simultaneously classifying attenuation into Tier C. If δ truly equals 1600 m, A=8.686×1000/1600=5.429 dB, which should fall into Tier A; these two statements are internally contradictory. If we adopt the original SOP’s insulator σ upper limit of 10⁻⁴ S/m, then δ≈17,986 m and A≈0.483 dB, which aligns with Tier C. Therefore, while the original example’s direction—that “ordinary granite is typically undetectable”—can be retained, the value of 1600 m cannot be retained; it must be revised to reflect an explicit σ assumption or interval.

    30 cm Concrete Example. The original text states δ≈80 cm and A≈3.3 dB. According to the skin depth formula, δ=0.8 m at 50 Hz with μᵣ=1 implies σ≈7,916 S/m, which is metallic-grade; whereas for σ=10⁻³–10⁻² S/m, δ ranges approximately 711–2251 m, yielding A only about 0.001–0.004 dB. PMID 36234036 indicates that concrete resistivity varies with formulation, dimensions, and frequency, supporting the necessity of specifying measured frequency and composition rather than treating a fixed δ as a universal constant; however, it does not directly validate the specific σ interval in this paper.

    μ-metal Example. The original δ≈0.5 mm cannot be used with μᵣ=1. μ-metal’s shielding performance depends on high magnetic permeability, thickness, geometry, and demagnetization state; PMID 36234022’s composite shielding study explicitly emphasizes that actual relative permeability deviates from simulated values. Thus, this example must include μᵣ as an input and cannot rely solely on σ; otherwise, 35 dB remains merely a placeholder for recalculation rather than a universal constant.

    3. The “Background Minus 10 dB” Iron Rule vs. the Three Tiers

    The original SOP also stipulates: A < Background B - 10 dB prohibits entry into E1. This rule depends on equipment and site background, differing fundamentally from the fixed Tier C <1 dB. When background is 30 dB, the prohibition line is 20 dB, directly consuming Tier A’s lower edge; when background is 10 dB, the prohibition line is 0 dB, offering almost no additional filtering. Both must be reported concurrently: the three tiers represent working classifications of material attenuation, while the iron rule represents experimental admission conditions relative to detection background.

    4. Evidence Boundaries

    PMID 20000094 provides measured examples of building interior geomagnetic field attenuation, demonstrating that shielding cannot rely solely on abstract formulas; PMID 36234036 supports the frequency- and formulation-dependence of concrete resistivity; PMID 18457487 and DOI 10.1080/10705510701301834 indicate that measurement invariance and fitting metric thresholds are sensitive to sample size, conditions, and model specification. None directly verify that “this project’s five-pool three-tier SOP can be adjudicated on real mountains or buildings.” Therefore, this paper’s evidence level is rated B, serving only as adjudication criteria and experimental prerequisites, not proof that environmental frequency effects have been demonstrated.

    The original SOP statement “avoid more than 70% of invalid experiments” lacks a denominator, historical experiment count, or failure definition, making recalculation impossible. This round does not retain that percentage; instead, we adopt an auditable rule: after recording f, d, σ interval, μᵣ interval, and equipment background, compute A across the full interval. If it crosses tiers or conflicts with the iron rule, mark it as “undecidable / requires measurement,” and do not force it into Tier A, B, or C.

    Actionable Revisions

    1. Report σ, μᵣ, f, and d intervals for all tiers first; do not report only single-point δ.
    2. If the σ interval ratio q≥5, Tier B defaults to R29-b: Intermediate tier empty set; Tier A/C must also check whether the full interval crosses boundaries.
    3. High-permeability materials like μ-metal must list μᵣ, geometry, and demagnetization state as mandatory fields.
    4. Separate equipment background lines from the three material tiers; if any condition fails, mark status as draft / undecidable, do not upgrade to E1.

    References

    • PMID 20000094: Measured shielding of buildings against the geomagnetic field.
    • PMID 36234036: Resistivity, frequency, and formulation dependence of high-performance concrete.
    • PMID 18457487: Power and sample size sensitivity of fit metrics in measurement invariance testing.
    • DOI 10.1080/10705510701301834: Simulation study of measurement invariance and threshold changes in fit metrics.

    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 这份 480 例是事前定的还是补出来的:已发表文献里的事后补样,数字本身留不留痕

    Evidence level: C

    事后补样污染的不是检验的通过率,而是「这是一份 n 例研究」这句话本身——而补样行为可能在样本量的数字痕迹上留下指纹,让已发表文献可以事后回溯筛查。

    正文

    从一个悬空的前提出发。 前一篇算过一笔账:如果研究者在精确门禁下「先收一批、不够再补」,alpha 的名义通过率其实没有被破坏——精确检验的门禁仍把假阳性压在 5% 附近。真正被污染的是别的东西:等效样本量塌到名义值的四成之一,精度虚高 2.2 倍。换句话说,补样悄悄改写的不是结论对不对,而是「这份数据值多少个独立样本」这个声明。但整个论证有一个悬空前提:真实文献里到底有多少研究补过样,谁也不知道。一项对 82 项一致性研究的描述性研究(PMID 36123642)发现只有 33% 给了样本量依据,而这 33% 里哪些是事前算的、哪些是「先收一批再补」,摘要层面完全看不出。本篇要问的是:这件事能不能不靠作者坦白,而从已发表的数字本身筛出来。

    三条痕迹,并行检验。 第一条是整十聚簇。事前按精度算出来的 n 往往是「任意」的数——87、113、152;而事后补出来的总数倾向于落在整齐的阶梯点上:60、120、240、480。如果「最末一次追加后的总数」系统性地偏爱整十整百,末位数字分布的卡方检验会直接显形。当然要小心一个混淆:研究者本来就偏爱整数,聚簇可能只是偏好而非补样——这正是判据要分开处理的情形。第二条是注册 n 与实际 n 的差值分布。有注册号的研究可以直接比:报差值中位数,以及「实际 > 注册」的单向比例。经验上这个方向确实值得怀疑——对比丹麦 1994-95 年批准的试验方案与发表稿,34 项有样本量计算的试验里 18 项存在未声明的样本量不一致(PMID 19056791);一项对 66 项亲密伴侣暴力研究的核查发现 39% 的研究与注册记录在样本量上有大改、44% 是研究完成后才注册的(PMID 29281583)。第三条是精度反查:按论文自报的目标精度,用按精度反解样本量的标准算法(PMID 24338600)算出「应有 n」,与实际 n 比——实际显著偏小者为「可能事后停」,显著偏大者为「可能事后补」。kappa 样本量表(PMID 15733050)这类经典依据被引用的频率,与引用者的实际 n 是否对得上,也从未被核对过。

    判据怎么下。 设想三条痕迹里至少两条给出同向结论——整十聚簇超出随机水平、实际>注册占多数、实际 n 系统性偏离反解值——那么「补样」就是可检测的,可以给该领域一个事后追加率的点估计,并写进方法学规范。若只有整十聚簇显著而注册比对不显著,聚簇反映的只是「研究者偏好整数」,此路作废。若三条全部不显著,则已发表文献无法回溯识别补样行为,任何关于「补样普遍性」的断言都不得作出,只能靠前瞻注册数据。而若「实际>注册」的占比超过 40%,那补样本身就是常态——相关的三条规范就必须从「建议」升级为该领域的「强制报告项」。

    与前面两篇的分工。 前一篇问的是同一份数据内部两笔误差能不能相加(记账问题);再前一篇算的是补样对检验门禁与等效样本量的代价(后果问题);本篇问的是跨研究之间的账目核对——这份数据的 n 是不是它自称的那个 n。三者互相独立,也互相补位:没有本篇,「补样代价」只对被作者承认的补样生效;有了本篇,代价才有可能落到整片文献上。

    可测手段:从 PubMed 抽取一批报告「例数 × 评定者数」的一致性研究,按有无样本量依据、有无注册号分层,三条痕迹检验并行。这是纯文献计量工作,不需要任何新数据采集。

    证据与局限

    已验证: 样本量的方案-发表不一致真实存在且比例可观(PMID 19056791:18/34 项未声明;PMID 29281583:39% 样本量大改、44% 完成后才注册);样本量依据的报告率低(PMID 36123642:仅 33%)。
    部分验证: 按精度反解 n 的算法本身成熟(PMID 24338600),用作「应有 n」基准技术上可行,但反解值对参数假设的敏感度未评估。
    待验证: 三条痕迹(整十聚簇 / 注册差值 / 精度反查)作为补样筛查工具的联合判别力完全未实测——本文全部判据均为事前设定,尚无一条被数据检验过。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络(CORRESPONDENCE_NETWORK.md 1983 行与 EDGE_TABLE_science-yixue.md 63 行中均无「样本量 / 补样 / 注册」对应边)

    证据等级: C
    参考文献: PMID 36123642, PMID 19056791, PMID 29281583, PMID 24338600, PMID 15733050

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Are these 480 cases pre-specified or post-hoc: does the number itself leave a trace in published literature?

    Evidence level: C

    Post-hoc sampling contaminates not the pass rate of statistical tests, but the very claim that “this is an n-case study”—and the act of post-hoc sampling may leave fingerprints on sample size numbers, enabling retrospective screening of published literature.

    Main Text

    Starting from an unresolved premise. The previous article calculated the cost: if researchers operate under strict precision gates using a “collect first, supplement if insufficient” approach, the nominal pass rate for alpha remains intact—strict testing still keeps false positives near 5%. What is truly contaminated is something else: the equivalent sample size collapses to one-quarter of its nominal value, while precision inflates by a factor of 2.2. In other words, post-hoc sampling quietly rewrites not whether the conclusion is right, but the declaration of “how many independent samples this data represents.” However, the entire argument rests on an unresolved premise: we do not know how many studies in the actual literature have been post-hoc sampled. A descriptive study of 82 consistency studies (PMID 36123642) found that only 33% provided sample size justification, and among those 33%, it is impossible to discern from the abstract alone which were pre-calculated versus “collect first, supplement later.” This article asks: can we detect this without relying on author disclosure, by screening the published numbers themselves?

    Three traces, tested in parallel. The first is clustering at round tens. Pre-calculated n values based on precision are often “arbitrary” numbers—87, 113, 152; whereas totals from post-hoc sampling tend to fall on neat step points: 60, 120, 240, 480. If “the final total after the last addition” systematically favors round tens and hundreds, a chi-square test on the distribution of trailing digits will reveal this directly. Of course, one must guard against a confounding factor: researchers inherently prefer integers, so clustering might reflect preference rather than post-hoc sampling—this is precisely why the criteria must be separated. The second is the distribution of differences between registered n and actual n. Studies with registration numbers can be compared directly: report the median difference, and the proportion where “actual > registered.” Empirically, this direction is indeed suspicious—comparing Danish trial protocols approved in 1994-95 with published reports, 18 of 34 trials with sample size calculations had undisclosed inconsistencies between registered and published sample sizes (PMID 19056791); a review of 66 intimate partner violence studies found 39% had major changes in sample size compared to registration records, and 44% were registered only after completion (PMID 29281583). The third is reverse-precision checking: using the standard algorithm (PMID 24338600) to solve for “expected n” based on the paper’s stated target precision, then comparing with actual n—significantly smaller actual values suggest “possible stopping,” while significantly larger ones suggest “possible supplementation.” The frequency with which classic tools like the kappa sample scale (PMID 15733050) are cited has never been cross-checked against whether the citing researchers’ actual n values align.

    How to set the criteria. Suppose at least two of these three traces yield concordant conclusions—round-tens clustering exceeds random levels, actual > registered is the majority, and actual n systematically deviates from reverse-solved values—then “post-hoc sampling” is detectable, allowing us to produce a point estimate of the post-hoc supplementation rate for this domain and incorporate it into methodological standards. If only round-tens clustering is significant while registration comparison is not, the clustering merely reflects “researchers prefer integers,” and this path is abandoned. If all three traces are insignificant, then published literature cannot be retrospectively identified for post-hoc sampling behavior; any claim about “universality of post-hoc sampling” must be withheld, relying instead on prospective registration data. But if the proportion where “actual > registered” exceeds 40%, then post-hoc sampling itself is the norm—the three related standards must be upgraded from “recommendations” to mandatory reporting items for this domain.

    Division of labor with previous articles. The prior article asked whether two errors within the same dataset could be summed (the accounting problem); the earlier one calculated the cost of post-hoc sampling for test gates and equivalent sample size (the consequence problem); this article asks cross-study reconciliation—whether this dataset’s n matches the claimed n. The three are mutually independent yet complementary: without this article, “cost of post-hoc sampling” applies only to cases authors admit; with this article, the cost can extend across the entire literature.

    Measurable approach: Extract from PubMed a batch of consistency studies reporting “cases × raters,” stratified by presence/absence of sample size justification and registration numbers, then apply the three-trace test in parallel. This is pure bibliometric work requiring no new data collection.

    Evidence and Limitations

    Verified: Inconsistency between planned and published sample sizes is real and substantial (PMID 19056791: 18/34 undisclosed; PMID 29281583: 39% major sample size revision, 44% registered after completion); low reporting rate of sample size justification (PMID 36123642: only 33%).
    Partially verified: The algorithm to solve for n based on precision is mature (PMID 24338600), making technical use of “expected n” feasible, though sensitivity to parameter assumptions remains unassessed.
    To be verified: The combined discriminative power of the three traces (round-tens clustering / registration difference / reverse-precision check) as post-hoc sampling screening tools has never been empirically tested—all criteria in this article are pre-specified, none yet validated by data.

    Related Reading

    L1 Network Status: This topic has not yet entered the L1 correspondence network (no “sample size / post-hoc sampling / registration” edges present in CORRESPONDENCE_NETWORK.md line 1983 and EDGE_TABLE_science-yixue.md line 63)

    Evidence Level: C
    References: PMID 36123642, PMID 19056791, PMID 29281583, PMID 24338600, PMID 15733050

    Evidence Level: A = direct PubMed support, B = scientific analogy / indirect evidence, C = traditional consensus / preliminary exploration


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 三档给了四个界:一个数量级没人管,一个数量级两个档抢

    Evidence level: C

    我们做过一份山体预筛 SOP:进山做 ELF 屏蔽实测之前,先按岩石电导率 σ 把候选山体分三档——值得先测、预算有余再测、别测。

    分档这件事看着最不需要动脑,恰恰最容易出错。把这份 SOP 的三档抄在纸上数一遍界,会发现:它给了四个界,而三个档只需要两个界。

    一、先数界

    三档的规矩是:k = 3 个档,需要 k − 1 = 2 个界。

    SOP 表里实际写的是:

    档位 σ 界 换算成衰减 A(dB) 表里标的 A
    A 档(优先测) ≥ 1e-2 4.83 ≥ 10
    B 档(条件性测) 1e-3 – 1e-1 1.53 – 15.27 1 – 10
    C 档(不测) ≤ 1e-4 0.48 < 1

    数一下 σ 轴上出现了几个界:1e-4、1e-3、1e-2、1e-1。四个。

    而 A 列只给了两个标签:1 和 10。

    四对二,这里就已经不对了。 多出来的两个界不是笔误,它们分别是 C/B 界的两种写法(1e-4 vs 1e-3)和 B/A 界的两种写法(1e-1 vs 1e-2)。

    二、多出来的界造成了什么

    把三档写成集合,直接摆在 σ 轴上:

    A = [1e-2, ∞)          B = [1e-3, 1e-1]          C = (0, 1e-4]
    

    两处毛病,各占一个数量级:

    区间 宽度 毛病
    (1e-4, 1e-3) 1 个数量级 没有任何一档收它
    [1e-2, 1e-1] 1 个数量级 A 档和 B 档同时收它

    在 σ ∈ [1e-5, 1e0] 这 5 个数量级的范围内,”没人管”和”两个抢”加起来 2 个数量级,占 40%。

    具体到一个样本上是什么后果:

    • 某座山的岩样测出 σ = 5e-4(A ≈ 1.08 dB)。翻遍 SOP 三档,没有一档告诉你该怎么办。它落在空隙里。
    • 另一座山 σ = 5e-2(A ≈ 10.8 dB)。A 档说”第一批测”,B 档说”预算有余再上”——这两句建议差一整个决策级,而两个档都说自己是它的归属。

    执行的人只能自己拍板。在这一段上,SOP 不是判据,是一道选择题。

    三、还有第二块死区:一条规则把自己的档判没了

    SOP 里另有一条容差规则:

    预筛判档时采用”下限原则”…… 任何档位判定的”最低期望 A(dB) 都应 > 3 dB,否则降档”。

    这条本身很合理——衰减太小就测不出东西,别浪费仪器时间。但把它套到 B 档下缘,会撞墙:

    B 档下缘 σ = 1e-3  ⇒  A = 1.53 dB  < 3 dB  ⇒  必须降档
    降档目标 = C 档,而 C 档只收 A < 1 dB(σ < 4.29e-4)
    ⇒ σ ∈ [4.29e-4, 1e-3) 留 B 不合格,进 C 也不合格
    

    这是一块”规则要求你降档、但降档后没有档可落”的地。

    把这条规则的影响算全:A = 3 dB 对应 σ = 3.86e-3,而 C 档上界是 4.29e-4。整块死区是 σ ∈ [4.29e-4, 3.86e-3],宽 9.00 倍,对应 A ∈ [1.00, 3.00] dB。

    而这一块和第二节的空隙在 [4.29e-4, 1e-3) 上重叠——这一段既”没档”,又”降档没落点”,是双重死区。

    四、四对二这件事本来可以一眼看出来

    这里有个很朴素的检查,值得写进规程:

    判据量对参数单调 ⇒ 分成 k 档只需 k−1 个界。
    若同一张表并列两个尺度(σ 和 A),两个尺度上的界必须一一对应。
    界数不等,必有多余界;多余界必然与某个档的定义相矛盾。
    

    这份 SOP 就是最干净的反例:3 档、4 个 σ 界、2 个 A 标签。任何一个”数一遍”的动作都能抓到。

    而我们过去几轮在池④ 反复抓的,全是更细的东西——判据维度选错、量词过宽、带宽不够、阈值位置不对。这些都比”界数对不对”靠后。 界数错了,后面所有讨论都是在一个不存在的东西上做精细加工。

    五、那个”42.8 倍”其实复算不出来

    上一轮留了个待办:按新立的档位密度定理(R32)复查这份 SOP,理由是”相邻档界之比 r = 42.8 倍,比另一个课题宽松得多,值得算一次看是否过关”。

    这轮算了,两件事:

    第一,42.8 复现不出来。 按表内 σ 界(1e-4 与 1e-2)算,r = 100;按 A 列的 1 dB 与 10 dB 反解 σ 界再算,r 还是 100。两种口径都是 100,不是 42.8。(48.3 是 A = 48.293·√σ 里的那个系数,可能是抄串了。)如实记:这一项来历未定。

    第二,更根本的是——R32 的形式是”设相邻档界之比 r”,它假定档界构成一个有序划分。这份 SOP 的档界不构成划分(第一节的空隙与重叠),第 1 和第 2 个界之间根本不存在”相邻”关系,中间还隔着 B 档的两个界。

    ⇒ r 无定义。
    正确答复不是"过关"或"不过关",而是"R32 不适用,
      先修档界使其构成划分,再谈 q<r"。
    

    这本身是 R32 的一次负面检验,而且是好事:它告诉我们 R32 有前提。上一轮我们写的注意事项是”凡见分档体系,先算 r 与 q”——这句要补半句:先查它是不是划分,再算 r 与 q。

    六、顺手抓到一条引文的区间错配

    SOP 的参考支持文献里有一条:

    PMID 20000094 — 建筑地磁屏蔽 1.2–2.0× 衰减;用于交叉佐证山体/建筑 σ 量级差。

    回查:这条文献真实、作者正确、数值 1.2–2.0× 正确。 它是 Chernykh 等 2010 年发表于 Gigiena i Sanitariia 的《多层住宅内地磁场的屏蔽》,测的是多层住宅内天然地磁场(直流、准静态)降低 1.2 到 2.0 倍。

    问题在别的方向:

    项 文献测的 SOP 要支撑的
    场的类型 天然地磁场,≈ 0 Hz(直流) 7.83 Hz(舒曼基频)
    屏蔽机制 高磁导率材料的静磁分流,与频率无关 导体内的趋肤效应,δ ∝ 1/√(σf)
    厚度尺度 建筑构件(厘米到米) 山体(公里)

    直流屏蔽由磁导率主导、与频率无关;ELF 屏蔽由电导率×频率主导、随频率开方变化。同一份材料上这两个数不能互相换算——除非磁导率和电导率同时已知,而那恰是这份 SOP 要求进现场去测的量。

    处置建议:这条不要删(它作为”材料屏蔽现象确实存在”的存在性证据仍然有效),但要改标成「直流地磁屏蔽的建筑实测,不可用于校核 ELF 衰减量级」。

    这是引文问题的第四种形态:号对、作者对、数值对、但物理区间不对。 前三种(号不存在、号存在但错指、号对但作者错)我们之前都抓过;这一种最难,因为所有”查号”的动作都会放行。

    七、证据与局限

    证据等级 C。 这是工程公式层面的解析审计,没有一次实测。

    三条局限,不得省略:

    1. 公式只在一个口径成立。 A = 48.293·√σ 是 1 km 厚度、7.83 Hz、μr=1 的特定值。换厚度或换介质磁导率,系数会变。但”3 档给了 4 个界”是纯计数矛盾,与系数无关——核心结论不受影响。
    2. “该改哪一档”是取舍,不是判决。 两条路各有代价:保留 1 dB / 10 dB 这两个好记的数,就得把 σ 界改成 4.29e-4 / 4.29e-2(σ 不再是整十次幂,现场简易 EC 计的阈值也要跟着改);保留 σ 的整十次幂,就得放弃 1 dB / 10 dB,把整个三档的语义重写。这是工程取舍,要人定。
    3. 第六节的引文判定只到”区间错配”,不是”假引文”。 不得据此扩大到”这份 SOP 的引文都不可信”。

    不许做的三条推论:

    • 不许读成”靠山预筛没用”。要读成”预筛的档界要重画“。
    • 不许读成”σ 不重要”。恰恰相反——正因 σ 是主参数,界画错了代价才大。
    • 不许把”r 无定义”写成”r 很大所以宽松”。这是两件完全不同的事。

    本文是 draft,未标 accepted,未进 779,未发布。 配套审计卡:08-ops/experiments/pool4-tier-boundary-counting-gap-overlap-audit.md(P4-AUDIT-021)。修订建议 D42 六条全在建议层,等 JOE 批。


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Three tiers assigned to four realms: one order of magnitude ignored, another contested by two tiers.

    Evidence level: C

    We developed a mountain pre-screening SOP: before entering the field to conduct ELF shielding measurements, candidates are classified into three tiers based on rock conductivity σ—those to measure immediately, those to measure if budget permits, and those not to measure.

    This tiering appears the least cognitively demanding step, yet it is most prone to error. Counting the boundaries in this SOP reveals: it defines four boundaries, yet three tiers require only two.

    I. Counting the Boundaries

    The rule for three tiers is: k = 3 tiers require k − 1 = 2 boundaries.

    The SOP table actually specifies:

    Tier σ Boundary Converted Attenuation A(dB) Table Label A
    Tier A (Priority Measure) ≥ 1e-2 4.83 ≥ 10
    Tier B (Conditional Measure) 1e-3 – 1e-1 1.53 – 15.27 1 – 10
    Tier C (Do Not Measure) ≤ 1e-4 0.48 < 1

    Counting the boundaries on the σ axis: 1e-4, 1e-3, 1e-2, 1e-1. Four.

    Column A provides only two labels: 1 and 10.

    Four versus two—this is already inconsistent. The extra two boundaries are not typos; they represent alternative expressions of the C/B boundaries (1e-4 vs 1e-3) and the B/A boundaries (1e-1 vs 1e-2).

    II. Consequences of the Extra Boundaries

    Expressing the three tiers as sets on the σ axis:

    A = [1e-2, ∞)          B = [1e-3, 1e-1]          C = (0, 1e-4]
    

    Two errors, each spanning one order of magnitude:

    Interval Width Error
    (1e-4, 1e-3) 1 order of magnitude No tier covers it
    [1e-2, 1e-1] 1 order of magnitude Tiers A and B both claim it

    Within the σ ∈ [1e-5, 1e0] range spanning five orders of magnitude, the “uncovered” and “contested” regions together cover 2 orders of magnitude (40%).

    Consequences for specific samples:

    • A mountain sample yields σ = 5e-4 (A ≈ 1.08 dB). Checking all three tiers, no tier specifies action. It falls in the gap.
    • Another mountain yields σ = 5e-2 (A ≈ 10.8 dB). Tier A says “measure first,” Tier B says “measure if budget allows”—these recommendations differ by one decision level, yet both tiers claim ownership.

    Practitioners must decide independently. In this region, the SOP is not a criterion but a choice question.

    III. A Second Dead Zone: A Rule Invalidates Its Own Tier

    The SOP includes another tolerance rule:

    When pre-screening assigns tiers, apply the “lower-bound principle”: the minimum expected A(dB) for any tier must exceed 3 dB; otherwise, downgrade.

    This is reasonable—insufficient attenuation yields no measurable signal, wasting instrument time. Applying it to Tier B’s lower bound creates a conflict:

    Tier B lower bound σ = 1e-3  ⇒  A = 1.53 dB  < 3 dB  ⇒  must downgrade
    Downgrade target = Tier C, but Tier C only accepts A < 1 dB (σ < 4.29e-4)
    ⇒ σ ∈ [4.29e-4, 1e-3) leaves Tier B ineligible, yet entering Tier C is also ineligible
    

    This is a “rule requires downgrade, but no tier accepts the downgraded value” zone.

    Calculating the full impact: A = 3 dB corresponds to σ = 3.86e-3, while Tier C’s upper bound is 4.29e-4. The dead zone spans σ ∈ [4.29e-4, 3.86e-3], width 9.00×, corresponding to A ∈ [1.00, 3.00] dB.

    This dead zone overlaps with the gap from Section II on [4.29e-4, 1e-3)—this segment is both “uncovered” and “downgraded with no landing spot,” creating a double dead zone.

    IV. The Four-versus-Two Issue Should Be Obvious

    A simple check deserves inclusion in protocols:

    If the criterion monotonically depends on the parameter, dividing into k tiers requires only k−1 boundaries.
    If a single table uses two scales (σ and A), boundaries on both scales must correspond one-to-one.
    Mismatched boundary counts imply extra boundaries; extra boundaries necessarily contradict tier definitions.
    

    This SOP provides the cleanest counterexample: 3 tiers, 4 σ boundaries, 2 A labels. Any “count once” action immediately identifies this.

    Our previous rounds in Pool 4 repeatedly identified finer issues—misselected criterion dimensions, overly broad quantifiers, insufficient bandwidth, misplaced thresholds. These are secondary to “boundary count mismatch.” If the boundary count is wrong, subsequent discussions refine a non-existent structure.

    V. The “42.8×” Factor Cannot Be Recalculated

    A pending task from the previous round: re-evaluate this SOP using the newly established tier density theorem (R32), because “the ratio of adjacent tier boundaries r = 42.8× is much looser than another project, worth checking.”

    This calculation yielded two results:

    First, 42.8 cannot be reproduced. Using the table’s σ boundaries (1e-4 and 1e-2), r = 100; solving σ boundaries from A column values (1 dB and 10 dB) still yields r = 100. Both methods give 100, not 42.8 (48.3 is the coefficient A = 48.293·√σ, possibly mis-copied). Record accurately: this item’s origin is undetermined.

    Second, more fundamentally— R32’s form “set adjacent tier boundary ratio r” assumes boundaries form an ordered partition. This SOP’s boundaries do not form a partition (gaps and overlaps from Section I); between the 1st and 2nd boundaries no “adjacent” relationship exists, separated by Tier B’s two boundaries.

    ⇒ r is undefined.
    The correct answer is not "passed" or "failed," but "R32 does not apply;
      first repair the tier boundaries to form a partition, then evaluate q < r."
    

    This constitutes a negative validation of R32, which is beneficial: it reveals R32’s prerequisite. Our previous note “for any tiering system, first compute r and q” must add: first verify it forms a partition, then compute r and q.

    VI. An Inconsistent Citation Interval Caught Incidentally

    Among the SOP’s reference literature:

    PMID 20000094 — Building geomagnetic shielding 1.2–2.0× attenuation; used for cross-validation of mountain/building σ scale differences.

    Verification: this citation is authentic, authorship correct, numerical range 1.2–2.0× accurate. It refers to Chernykh et al. 2010 in Gigiena i Sanitariia, “Shielding of the Earth’s Magnetic Field in Multi-Story Residential Buildings,” measuring reduction of natural geomagnetic fields (DC, quasi-static) by 1.2 to 2.0 times within multi-story buildings.

    The discrepancy lies elsewhere:

    Item Literature Measured SOP Requires
    Field type Natural geomagnetic field, ≈ 0 Hz (DC) 7.83 Hz (Schumann fundamental)
    Shielding mechanism Static magnetic shunting via high-permeability materials, frequency-independent Skin effect in conductors, δ ∝ 1/√(σf)
    Thickness scale Building components (centimeters to meters) Mountains (kilometers)

    DC shielding is dominated by permeability, frequency-independent; ELF shielding is dominated by conductivity×frequency, scaling with square root of frequency. These two numbers cannot be interconverted for the same material—unless both permeability and conductivity are known, which is precisely what this SOP requires field measurement.

    Recommendation: Do not delete this citation (it remains valid existence evidence that “material shielding phenomena do occur”), but relabel it as “Building geomagnetic shielding measurement cannot calibrate ELF attenuation scale.”

    This represents the fourth citation error type: number correct, author correct, numerical range correct, but physical interval mismatched. We previously identified three types (number missing, number present but misreferenced, number correct but author wrong); this fourth type is most difficult because all “number lookup” actions accept it.

    VII. Evidence and Limitations

    Evidence Level C. This is analytical audit at the engineering formula level, without field measurement.

    Three limitations, none to be omitted:

    1. The formula holds only for one scale. A = 48.293·√σ applies specifically to 1 km thickness, 7.83 Hz, μr=1. Changing thickness or medium permeability alters the coefficient. However, “3 tiers define 4 boundaries” is a pure counting contradiction, independent of the coefficient—the core conclusion remains unaffected.
    2. “Which tier to adjust” is a trade-off, not a verdict. Both paths incur costs: retaining the memorable 1 dB / 10 dB requires changing σ boundaries to 4.29e-4 / 4.29e-2 (σ is no longer powers of ten, requiring field portable EC meter threshold adjustments); retaining σ’s powers of ten requires abandoning 1 dB / 10 dB and rewriting the entire three-tier semantics. This is an engineering trade-off requiring human judgment.
    3. Section VI’s citation judgment stops at “interval mismatch,” not “false citation.” Do not extrapolate to “all citations in this SOP are unreliable.”

    Three forbidden inferences:

    • Do not read “mountain pre-screening is useless.” Read instead “pre-screening tier boundaries must be redrawn.”
    • Do not read “σ is unimportant.” On the contrary—because σ is the primary parameter, boundary errors incur greater cost.
    • Do not write “r undefined” as “r very large, hence loose.” These are fundamentally different statements.

    This document is draft, not marked accepted, not entered into 779, not published. Supporting audit card: 08-ops/experiments/pool4-tier-boundary-counting-gap-overlap-audit.md (P4-AUDIT-021). Revision recommendations D42 all at suggestion level, pending JOE approval.


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • ρ 的正负不用猜:两组评分者的五档分布谁「压」谁,门限就往谁那边偏——锚误差的符号可以从现成数据里读出来

    Evidence level: C

    锚校正该不该用,取决于资深者与普通读者的误差是否同向(ρ 的符号)。这个符号不必做扰动实验:两份现成的五档边际分布一比对,门限往哪边挪了,符号就定了。

    正文

    一、设问的最后一环:符号从哪来

    上一篇(#1460)把锚校正的账算清了:锚误差 h 与读者误差 e 的相关系数 ρ 为正时,加锚有收益窗口(τ ∈ [ρ(√m−1), ρ(√m+1)],谷底 1−ρ²);ρ 为负时锚是灾难(ρ=−0.5、τ=1.414 时相对方差 5.0,是独立情形的 1.7 倍)。但真实研究里 ρ 从未被测过,而它由「两类评分者的偏倚是否同向」决定:资深者判「变化够大」的门限更低(看得更细),普通读者门限更高,同一份材料上一个偏正一个偏负,ρ<0;两者共用同一套不成文门限、只是噪声大小不同,则 ρ>0。问题在于:门限位置看不见,怎么比?

    二、门限是窗,边际比例是它的影子

    五档量表的评分机制可以写成:评分者内心有个连续量,过一道道门槛,落到第几档就报第几档。门槛本身观测不到,但它留了影子——每档的边际占比。关键在于:同一批材料由两组人各判一遍时,底层的真实量对两组是同一个(判的是同一批),于是两组边际分布的任何系统性差异,只能来自门槛位置或噪声尺度。这正是多面 Rasch 测量反复证实的现象:评分者「严厉度」差异普遍存在且幅度可观——54 个评审组 × 3000 名考生的大学入学评分量出大而不一致(PMID 40558845);五点量表上直接观察到评分者用分位置差异、须把评分者效应建进模型(PMID 42841527)。影子既然存在,就能反推窗子挪向哪边。

    三、方向判据:一阶随机占优

    判据是:把两组的累计分布画出来,若资深者的 CDF 在所有切点上都不高于普通读者(高档占比系统性更高),则资深者门限整体更低——一阶随机占优。本轮用 20 万次模拟核验:门槛基准 [−1.5, −0.5, 0.5, 1.5],资深者整体下移 1.4 档时,其五档边际为 [0.018, 0.117, 0.324, 0.358, 0.184],普通读者为 [0.240, 0.377, 0.285, 0.087, 0.011],四个切点的 CDF 差全部为负(−0.22 到 −0.48)——占优方向与门限平移方向严格一致;下移 0.5 档的温和情形占优仍成立(四切点 −0.04 到 −0.19)。反过来,资深者高档占比系统性更高,就等价于「同一份边缘病例,资深者更愿意报重」——这正是两类误差反号的结构条件,ρ<0。

    四、关键对照:占优失败不等于没有差异

    纯噪声差异(门槛相同、只是资深者噪声更小)下两组 CDF 会交叉:模拟中噪声比 1.6 比 1 时,CDF 差在低段 +0.107、高段 −0.107,符号翻转,不构成占优。所以「无占优」排除的是门限平移的主导地位,不是一切差异;此时 ρ>0 路线成立(误差同向),锚校正可用。这也解释了为什么不能拿「均值差」当门限差用:均值差不区分「平移」与「散布变化」,占优检验区分。信号检测文献里把判据位置与辨别力分开估计,正是同一逻辑(PMID 23784010);医学判读上新手与算法/专家的一致性差异也反复被量化(PMID 42278802 新手目测与无标记动捕分类的对比;PMID 42474519 评阅者间膝关节数字分级一致性)。而评分者效应研究同时警告:训练不必然消除严厉度差(PMID 40471819:训练只在部分维度上缩小了评分者差异)——所以先读分布再决定校不校,比先训练后祈祷更省。

    五、落到锚校正的三分叉

    ①资深者分布一阶占优且门限平移量大于 0.3 档宽 ⇒ ρ<0,锚校正在本场景不适用,改「先按经验分层、各层自校准」;②分布交叉、仅散布不同 ⇒ ρ>0,加锚成立,且按 #1460 的收益窗口可放宽 τ;③占优但平移小于 0.1 档宽 ⇒ 视为门限共用,按噪声差异处理,ρ 的上界由 1/√R 给出。整套判据的全部原料——两组各自的五档占比——在任何已完成的判读研究里都是现成的,一分钱新实验都不用花。

    证据与局限

    已验证: 门限-边际映射的数学结构(序数门限模型恒成立);一阶随机占优对门限平移方向的充分性(本轮 20 万次模拟,平移 0.5 与 1.4 档两情形四切点全部同号)。
    部分验证: 评分者严厉度差异的普遍性(多面 Rasch 多项实测:PMID 40558845、PMID 42841527、PMID 40471819);占优方向在小平移下受抽样噪声影响,需 Mann-Whitney 类检验与足够样本量(建议每组 ≥15 人、≥120 例)支撑。
    待验证: ①「占优方向 ⇒ ρ<0」的映射在同一批材料双组判读数据上的实测(本判据的核心外推);②门限平移量(档宽单位)与 ρ 的绝对值的定量关系;③两组同时存在平移与散布差时,占优方向是否仍由平移主导(模拟显示平移主导时仍占优,但边界条件未定);④资深者门限更低是否等于「误差反号」——若资深者的高档占比差异来自真实的辨别力而非门限,符号判断须重推。

    相关阅读

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md(2037 行)内检索「门限/阈值/偏倚/评分者/序数/占优」仅命中 E-060 相邻边(第 202 行 E-060「标签前提」conflict:中医医师间诊断一致性低 PMID 30383398 vs 训练可改善 PMID 19538100,未裁决;第 260 行 KB-333 conflict:κ 型系数对序数分类的稳健性方向与 Gwet 一系相反;第 261 行 KB-334:阳性率不同时一致性系数与患病率须分报)——均为「标签信度」簇的相邻边,与「评分者门限位置差 × 误差相关符号 ρ」非同一边;边表(63 行)无对应边。按硬规则证据等级封顶 C,缺口已登记。

    证据等级: C
    参考文献: PMID 40558845, PMID 42841527, PMID 40471819, PMID 23784010, PMID 42278802, PMID 42474519

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    The sign of the anchor error is directly readable from existing data: whichever of the two raters’ five-level distributions exerts more pressure determines the direction of the…

    Evidence level: C

    Whether the correlation coefficient ρ between the anchor error $h$ and reader error $e$ is positive or negative depends entirely on which group of raters exerts more “pressure” on the rating scale. The sign of ρ can be read directly from existing data by comparing the five-tier marginal distributions of the two groups.

    Whether to apply anchor correction depends on whether the errors of senior raters and ordinary readers are aligned (the sign of ρ). This sign does not require perturbation experiments: simply comparing two existing five-tier marginal distributions reveals which side the threshold shifts toward, thereby determining the sign.

    Body

    I. The Final Step of Questioning: Where Does the Sign Come From?

    The previous entry (#1460) settled the accounting for anchor correction: when ρ is positive, adding an anchor yields a benefit window (τ ∈ [ρ(√m−1), ρ(√m+1)], with a valley of 1−ρ²); when ρ is negative, the anchor becomes a disaster (e.g., at ρ=−0.5 and τ=1.414, the relative variance is 5.0, which is 1.7 times that of the independent case). However, in real-world studies, ρ has never been directly measured. Instead, it is determined by whether the biases of the two types of raters point in the same direction: senior raters set a lower threshold for judging “sufficient change” (seeing more detail), while ordinary readers set a higher threshold; if the same material is rated such that one group leans positive and the other negative, then ρ<0; if both groups share the same unwritten threshold but differ only in noise magnitude, then ρ>0. The problem is: since the threshold position is invisible, how can we compare them?

    II. Thresholds Are Windows; Marginal Proportions Are Their Shadows

    The rating mechanism of a five-tier scale can be written as: each rater has an internal continuous quantity that passes through successive thresholds, landing in the corresponding tier. The thresholds themselves cannot be observed directly, but they leave behind shadows—the marginal proportions in each tier. Crucially, when the same batch of materials is rated by two groups, the underlying true quantity is identical for both (since they are rating the same items); thus, any systematic difference between the two groups’ marginal distributions must arise solely from differences in threshold positions or noise scales. This is precisely the phenomenon repeatedly confirmed by multi-faceted Rasch measurement: differences in rater “severity” are ubiquitous and substantial—54 review groups × 3,000 examinees yielded large yet inconsistent university entrance scores (PMID 40558845); on five-point scales, direct observation shows raters use different cut-off positions, necessitating that rater effects be built into the model (PMID 42841527). Since these shadows exist, we can reverse-engineer which way the window has shifted.

    III. Directional Criterion: First-Order Stochastic Dominance

    The criterion is as follows: plot the cumulative distribution functions (CDFs) of both groups. If the senior rater’s CDF is never higher than that of the ordinary reader at any cut-off point (i.e., the proportion in higher tiers is systematically greater), then the senior rater’s threshold is overall lower—this constitutes first-order stochastic dominance. In this round, 200,000 simulations verified this: with baseline thresholds [−1.5, −0.5, 0.5, 1.5], when the senior group’s thresholds shift downward by 1.4 tiers, their five-tier marginal distribution becomes [0.018, 0.117, 0.324, 0.358, 0.184], while the ordinary group’s is [0.240, 0.377, 0.285, 0.087, 0.011]; the CDF differences at all four cut-off points are negative (−0.22 to −0.48)—the direction of dominance strictly matches the direction of threshold shift; even in the gentler case of a 0.5-tier downward shift, dominance still holds (CDF differences at the four cut-off points range from −0.04 to −0.19). Conversely, if the senior group’s proportion in higher tiers is systematically greater, this is equivalent to “for the same borderline case, the senior rater is more willing to rate it heavily”—this is precisely the structural condition where the two types of errors have opposite signs, implying ρ<0.

    IV. Key Contrast: Failure of Dominance Does Not Mean No Difference

    Under pure noise differences (same thresholds, but the senior group has smaller noise), the two groups’ CDFs will cross: in simulations, when the noise ratio is 1.6 to 1, the CDF difference is +0.107 at the low end and −0.107 at the high end, causing the sign to flip, thus failing dominance. Therefore, “no dominance” excludes only the dominance of threshold shifting, not all differences; in such cases, the ρ>0 route holds (errors are aligned), and anchor correction is applicable. This also explains why one cannot use “mean difference” as a proxy for “threshold difference”: mean differences do not distinguish between “shift” and “spread change,” whereas dominance tests do. Signal detection literature separates the estimation of criterion location from discriminability, following the same logic (PMID 23784010); medical diagnosis repeatedly quantifies consistency differences between novices and algorithms/experts (PMID 42278802 novice visual estimation vs. markerless motion capture classification; PMID 42474519 inter-rater consistency in knee digital grading). Meanwhile, rater effect research simultaneously warns that training does not necessarily eliminate severity differences (PMID 40471819: training only reduces rater differences on some dimensions)—thus, reading the distribution first to decide whether to calibrate is more efficient than training first and then praying.

    V. The Three-Way Fork of Anchor Correction

    ① If the senior group’s distribution exhibits first-order dominance and the threshold shift exceeds 0.3 tier widths ⇒ ρ<0, so anchor correction is inappropriate in this scenario; instead, adopt “stratify by experience first, then self-calibrate each layer”; ② If the distributions cross with only spread differences ⇒ ρ>0, so adding an anchor is valid, and per the benefit window in #1460, τ can be relaxed; ③ If dominance exists but the shift is less than 0.1 tier widths ⇒ treat as shared thresholds, handle via noise differences, with the upper bound of ρ given by 1/√R. All raw materials for this entire criterion—the five-tier proportions of each group—are readily available in any completed rating study, requiring not a single penny of new experimentation.

    Evidence and Limitations

    Verified: The mathematical structure of threshold-marginal mapping (ordinal threshold models always hold); the sufficiency of first-order stochastic dominance for determining the direction of threshold shift (in this round, 200,000 simulations confirmed that for shifts of 0.5 and 1.4 tiers, the CDF differences at all four cut-off points retained the same sign).

    Partially Verified: The universality of rater severity differences (multi-faceted Rasch measurements confirmed this: PMID 40558845, PMID 42841527, PMID 40471819); the direction of dominance under small shifts is susceptible to sampling noise, requiring Mann-Whitney-type tests and sufficient sample sizes (recommended ≥15 per group, ≥120 cases) to support robust inference.

    Pending Verification: ① Empirical validation of the mapping “dominance direction ⇒ ρ<0” on dual-group rating data from the same batch of materials (this is the core extrapolation of this criterion); ② The quantitative relationship between threshold shift magnitude (in tier-width units) and the absolute value of ρ; ③ Whether dominance direction remains governed by shift when both groups exhibit simultaneous shift and spread differences (simulations show shift dominates when present, but boundary conditions remain undetermined); ④ Whether lower senior thresholds equate to “opposite-signed errors”—if the senior group’s higher-tier proportion differences stem from genuine discriminability rather than threshold effects, the sign judgment must be re-derived.

    Further Reading

    L1 Network Status: This topic has not yet entered the L1 corresponding network. Searching for “threshold/limit/bias/rater/ordinal/dominance” in CORRESPONDENCE_NETWORK.md (2,037 lines) only hits E-060 adjacent edges (line 202, E-060 “label premise” conflict: low diagnostic consistency among TCM physicians PMID 30383398 vs. training improving it PMID 19538100, unresolved; line 260, KB-333 conflict: κ-type coefficient robustness direction for ordinal classification opposes Gwet’s series; line 261, KB-334: when positivity rates differ, consistency coefficients and prevalence must be reported separately)—all belonging to the “label reliability” cluster, not the same edge as “rater threshold position difference × error correlation sign ρ”; the edge table (63 lines) lacks the corresponding edge. Hard-rule evidence level capped at C; gap registered.

    Evidence Level: C
    References: PMID 40558845, PMID 42841527, PMID 40471819, PMID 23784010, PMID 42278802, PMID 42474519

    Evidence Level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 用偏度峰度去分辨「切点有没有被训练动过」,这套判据一开动就是反的:离散分档上位置移动本身就改形状

    Evidence level: C

    题目假设「只移动位置不动切点,分级分布的形状不变,所以偏度和峰度可以用来验形状」——这条假设在连续分布上成立,但判读用的是五档离散格子,位置一移动,格子边界就被重新分配,形状必然变。判据因此把两种机制判反了。

    一、题目想解决什么,以及它的主张

    前序工作(#1421)证明了一个约束:在多面 Rasch 模型里,评分者的位置参数 b 与切点组的整体平移完全共线,六点 profile 扫描下 ΔNLL 全为零,要估计就必须外生假定「两轮共享同一套切点」。#1458 想把这条外生假定变成可检验的:如果训练只动了 b,那么同一个评分者在训练前后两轮的档级频次分布,应当是「同一形状沿潜在轴滑了一段」,偏度和峰度不该变;如果动的是切点间距(把某一档切得更细),形状本身会改变,偏度能捕捉到。判据说得很明确:一阶矩差显著、二至四阶矩差不显著,则「只动 b」成立。

    这个想法在教科书式的连续正态模型里是对的。问题在于,评分者交出来的不是连续值,是 1 到 5 这五个格子。

    二、硬算一遍:位置一移动,形状就变

    取一套不均匀的五级切点(间距比约 0.9 : 0.9 : 1.0,贴近真实判读习惯),固定间距、只让位置从 −1.0 走到 +1.0,逐点算五档概率和前四阶矩:

    位置 均值 方差 偏度 超额峰度
    −1.0 4.079 0.813 −0.757 +0.016
    −0.5 3.633 1.018 −0.408 −0.408
    0.0 3.138 1.128 −0.126 −0.592
    +0.5 2.629 1.118 +0.159 −0.658
    +1.0 2.144 0.977 +0.512 −0.486

    切点间隔一个都没有动,偏度却从 −0.757 走到 +0.512,跨度 1.27。 峰度从 +0.016 走到 −0.658,跨度 0.67。原因很朴素:偏度和峰度是连续分布的形状描述量,它们假定支撑集不随位置移动。一旦格子固定、分布滑过去,大量样本从 5 档掉到 4 档、从 2 档升到 3 档,格子把连续分布切成了不同形状的多项式——形状当然变。

    所以题目的判据前提不成立。「二至四阶矩不变」不是一个可以拿来当零假设的性质,它本身就是被违反的。

    三、判据不只失效,而且是反向的

    更麻烦的是失效的方向。取一个真实「只动 b」的场景(位置 −0.9 走到 +0.6,间距全程固定 1.0),算偏度差:

    真·只动b:Δ偏度 = +0.901,抽样标准差(n=400)= 0.122,z = 7.37
    

    判据规定「偏度差显著 ⇒ 切点被动过」。于是在这个切点一动没动的场景里,判据以 z=7.37 的强度宣布「切点被动了」。样本越大越确信,n=1600 时 z=14.7——加大样本只会让误判更坚定。

    再取一个真实「动切点」的场景(位置不动,间距从 1.0 拉到 1.3):

    真·动切点:Δ偏度 = +0.012,抽样标准差(n=400)= 0.123,z = 0.10
    

    判据规定「偏度差不显著 ⇒ 只动 b」。于是切点确实被动了,判据却以 z=0.10 宣布「没动」。而且这不是样本量不够——把 n 从 100 加到 1600,z 只从 0.05 涨到 0.21,永远够不到显著线。因为间距变化对偏度的真效应本身就只有 0.012,比 n=400 时的抽样噪声小一个数量级。

    两种机制在矩上的指纹对照:

    真实机制 Δ均值 Δ偏度 Δ峰度
    只动 b −1.467 +0.901 −0.547
    只动切点 +0.013 +0.012 +0.342
    两者都动 −1.340 +0.718 −0.266

    判据选的「偏度」恰好是区分力最差的量:它对位置极度敏感(0.901),对切点几乎瞎(0.012)。判据把最灵敏的通道接到了错误的因上,把最钝的通道当成了主判据。

    四、那到底哪里不可识别,哪里可以识别

    这里要把两件事分开,混在一起会得出错误结论。

    不可识别是真的。 自由切点模型有 5 个参数(位置 b 加 4 个切点),单轮五档只给 4 个独立比例,自由度 −1。多起点拟合同一组观测,能给出 b = 0.0、0.5、1.0、−0.5 等无限多组解,残差全在 1e−16 量级。这是严格不可识别,#1421 的判断没错。

    但不可识别的来源不是「位置与间距分不开」,而是「切点形状完全未知」。 上面那四组解里,切点间距的绝对值随 b 变,但相对形状(间距比 0.964 : 0.964 : 1.071)每一组都完全一致。也就是说:形状是可从单轮识别的,只有绝对尺度不可识别。

    一旦把切点形状固定下来——比如假定等距,只留位置 b 和一个间距 s——参数降到 2 个,单轮 4 个比例给出自由度 2,模型变成可识别。多起点拟合五组不同的真值 (−0.8,0.7)、(0,1.0)、(0.9,1.3)、(−1.2,1.8)、(0.4,0.5),全部唯一收敛到真值,残差 1e−21 到 1e−31。

    所以正确的表述是:训练有没有改变切点形状,这件事本身是可检验的,但前提是先接受一个切点形状的约束模型(如等距),再用模型比较去检验这个约束。 不能靠无约束的矩去绕开它。

    五、换成似然比检验,判据才站得住

    正确的做法是嵌套模型比较:零模型让两轮共享同一个间距 s(3 参数),备择模型让两轮各自有 s(4 参数),用似然比统计量做卡方检验(自由度 1)。同一套数据,功效是这样的:

    每读者病例数 检出「切点形状漂移」的把握
    60 0%
    120 1%
    240 8%
    480 36%
    960 91%

    而真只动 b 时的误报率,n=120 和 n=480 都是 0%(名义水平 5%)。也就是说这个检验是「保守但正确」——不会冤枉只动 b 的读者,只是在样本小时抓不到真的漂移。

    关键结论:n=120 例/读者的常规判读研究,对切点形状漂移的检验力只有 1%,等同于没有检验。 20 名读者各 120 例合并起来,把握仍然是 0%——合并读者不能救回单读者层面的形状漂移,因为漂移是读者特异的方向,合并会互相抵消。

    要做这件事,单读者需要 900 例以上。这个数字直接决定了「切点有没有被训练动过」这个问题在现有数据集上根本没有答案,而不是「答案是不显著」。

    六、这条结论的边界

    第一,这里的计算基于正态潜变量加固定切点的分档模型。真实判读数据的潜变量未必正态,厚尾或双峰会改变矩的具体数值,但不会改变方向性:离散化之后位置移动必然改形状,这条与分布族无关(只要格子数有限、格子宽度有限)。

    第二,等距切点是一个强假定,本文用它只是为了让模型可识别,不是说真实切点等距。更稳的做法是把切点形状参数化(比如用两个参数控制不对称性),再做嵌套比较。

    第三,本文只处理了「形状漂移」这一种机制。训练还可能同时改变评分者使用的档级数(category collapse),那会让某些格子概率趋零,此时似然比的渐近卡方近似会失效,需要用精确检验或参数自助法。

    证据与局限

    已验证:
    – 离散分级下固定切点只移动位置,偏度必然改变:本文网格计算,位置从 −1.0 到 +1.0 偏度跨 1.27,峰度跨 0.67
    – 多面 Rasch 模型中评分者位置参数与时间特异严重度漂移是已建模对象,且时间独立性假设被明确批评为低效(PMID 36284065)
    – 评分者宽严(leniency/stringency)是可估计、可跨时间复现的真实效应(PMID 16919156:1259 名考官 142,030 条评分,12% 主效应方差来自考官宽严;PMID 8322872 早期证实评分者严重度显著)
    – 小样本下 kappa 类系数的置信区间不可靠、需要足够测试例数(PMID 20473138)

    部分验证:
    – 多面 Rasch 的五类评分者效应(含 central tendency 与 leniency)在模型内可分离(PMID 14523257),但这是连续潜变量层面,不是离散档级分布层面
    – 训练确实能改变判读行为(PMID 21899969 专用培训改变观察者间变异;PMID 37127220 国际间乳腺密度判读差异),但这两项都报告系数的变化,未报告档级分布形状的变化

    待验证:
    – 本文的偏度真效应量(0.012)依赖正态潜变量假定,尚未在真实五档判读数据上用半参数方法复核
    – 「单读者 900 例才能检出形状漂移」这个数字是在间距从 1.0 拉到 1.3(效应量偏大)条件下算的,真实效应更小时所需样本还要更大
    – 档级塌缩(category collapse)下的检验方法未做功效评估

    相关阅读

    证据等级: C
    参考文献: PMID 36284065, PMID 16919156, PMID 8322872, PMID 14523257, PMID 20473138, PMID 21899969, PMID 37127220

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md(1881 行)全文检索「切点」仅 1 处(E-060 条目内 AC2 定切点的实操说明,与本主题无关)、「分档」1 处(KB-496 整数分箱风险,属 GDV 参数分箱,不同层)、「偏度」0 处、「峰度」0 处、「平移」0 处、「位置参数」0 处、「可识别」0 处、「severity」0 处;EDGE_TABLE_science-yixue.md(63 行、59 条概念边)无任何一致性系数或分档分布层概念。故本主题无对应边,证据等级封顶 C。

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Using skewness and kurtosis to determine whether the cut point has been trained, this criterion is inverted: moving the position on the discrete binning inherently alters the sh…

    Evidence level: C

    The premise assumes that “moving only the location while keeping cutpoints fixed preserves the graded distribution’s shape, so skewness and kurtosis can verify the shape”—a hypothesis valid for continuous distributions but flawed when applied to five discrete bins. Shifting the location reassigns bin boundaries, inevitably altering the shape. Consequently, the criterion reverses the two mechanisms it aims to distinguish.

    I. What the Problem Aims to Solve and Its Claim

    Prior work (#1421) established a constraint: in the Multidimensional Rasch model, the rater location parameter $b$ is perfectly collinear with a uniform shift of the entire cutpoint group. Under a six-point profile scan, $\Delta\text{NLL}$ remains zero; estimation requires the exogenous assumption that “two rounds share the same cutpoints.” #1458 seeks to make this exogenous assumption testable: if training shifts only $b$, the frequency distribution of grades for the same rater across two rounds should represent “the same shape sliding along the latent axis,” leaving skewness and kurtosis unchanged; if training alters cutpoint spacing (making one grade finer), the shape itself changes, and skewness captures this. The criterion states clearly: a significant first-moment difference with insignificant second through fourth moment differences implies “only $b$ moved.”

    This logic holds in textbook continuous normal models. The problem lies in the fact that raters do not provide continuous values but rather fall into five discrete bins (1 to 5).

    II. Hard Calculation: Moving Location Alone Alters Shape

    Using an uneven five-level cutpoint scheme (spacing ratios approximately 0.9 : 0.9 : 1.0, reflecting real-world grading habits), we fix the spacing and shift the location from −1.0 to +1.0, calculating five-grade probabilities and the first four moments at each step:

    Location Mean Variance Skewness Excess Kurtosis
    −1.0 4.079 0.813 −0.757 +0.016
    −0.5 3.633 1.018 −0.408 −0.408
    0.0 3.138 1.128 −0.126 −0.592
    +0.5 2.629 1.118 +0.159 −0.658
    +1.0 2.144 0.977 +0.512 −0.486

    With cutpoint intervals unchanged, skewness shifts from −0.757 to +0.512, a span of 1.27. Kurtosis moves from +0.016 to −0.658, a span of 0.67. The reason is straightforward: skewness and kurtosis describe the shape of continuous distributions, assuming the support set does not shift with location. Once bins are fixed and the distribution slides over them, large numbers of samples drop from bin 5 to bin 4 or rise from bin 2 to bin 3. The bins slice the continuous distribution into different-shaped polynomials—the shape necessarily changes.

    Thus, the premise of the criterion fails. “Second through fourth moments unchanged” is not a property to serve as the null hypothesis; it is itself violated.

    III. The Criterion Not Only Fails but Operates in Reverse

    More troubling is the direction of failure. Consider a true “only $b$ moves” scenario (location shifts from −0.9 to +0.6, spacing fixed at 1.0 throughout), calculating skewness difference:

    True·only b moves: Δskewness = +0.901, sampling SD (n=400) = 0.122, z = 7.37
    

    The criterion stipulates “significant skewness difference ⇒ cutpoints moved.” In this scenario where cutpoints did not move at all, the criterion declares with $z=7.37$ that “cutpoints were altered.” Larger samples increase certainty: at $n=1600$, $z=14.7$—increasing sample size only strengthens the misjudgment.

    Now consider a true “cutpoints moved” scenario (location fixed, spacing stretched from 1.0 to 1.3):

    True·cutpoints moved: Δskewness = +0.012, sampling SD (n=400) = 0.123, z = 0.10
    

    The criterion stipulates “insignificant skewness difference ⇒ only $b$ moved.” Thus, where cutpoints indeed moved, the criterion declares with $z=0.10$ that “they did not move.” This is not due to insufficient sample size—raising $n$ from 100 to 1600 increases $z$ only from 0.05 to 0.21, never reaching significance. The spacing change exerts a true effect on skewness of only 0.012, an order of magnitude smaller than sampling noise at $n=400$.

    Fingerprints of the two mechanisms on moments:

    True mechanism ΔMean ΔSkewness ΔKurtosis
    Only $b$ moves −1.467 +0.901 −0.547
    Only cutpoints move +0.013 +0.012 +0.342
    Both move −1.340 +0.718 −0.266

    Skewness—the metric chosen by the criterion—is precisely the least discriminative quantity: highly sensitive to location (0.901) but nearly blind to cutpoints (0.012). The criterion connects the most sensitive channel to the wrong cause and treats the least sensitive channel as the primary criterion.

    IV. What Is Unidentifiable and What Is Identifiable

    We must separate these two issues; conflating them yields erroneous conclusions.

    Unidentifiability is genuine. The free cutpoint model has 5 parameters (location $b$ plus 4 cutpoints). A single round with five grades provides only 4 independent proportions, yielding degrees of freedom −1. Fitting multiple starting points to the same observations yields infinitely many solutions ($b = 0.0, 0.5, 1.0, −0.5$, etc.), with residuals on the order of $10^{-16}$. This is strict unidentifiability; #1421’s judgment is correct.

    However, the source of unidentifiability is not “location and spacing inseparable,” but rather “cutpoint shape completely unknown.” Among the four solution sets above, the absolute values of cutpoint spacing vary with $b$, yet the relative shape (spacing ratios 0.964 : 0.964 : 1.071) remains identical across all solutions. Thus: shape is identifiable from a single round; only absolute scale is unidentifiable.

    Once cutpoint shape is fixed—for example, assuming equal spacing and retaining only location $b$ and one spacing $s$—parameters drop to 2. A single round with four proportions yields degrees of freedom 2, rendering the model identifiable. Fitting five distinct true values $(−0.8,0.7), (0,1.0), (0.9,1.3), (−1.2,1.8), (0.4,0.5)$ converges uniquely to the true values, with residuals ranging from $10^{-21}$ to $10^{-31}$.

    Therefore, the correct statement is: whether training altered cutpoint shape is itself testable, but only after accepting a constrained cutpoint shape model (e.g., equal spacing) and then using model comparison to test that constraint. One cannot bypass this by relying on unconstrained moments.

    V. Likelihood Ratio Testing Makes the Criterion Viable

    The proper approach is nested model comparison: the null model assumes both rounds share the same spacing $s$ (3 parameters), while the alternative model allows each round its own $s$ (4 parameters). A chi-square test using the likelihood ratio statistic (degrees of freedom 1) yields the following power for the same dataset:

    Number of cases per rater Power to detect “cutpoint shape drift”
    60 0%
    120 1%
    240 8%
    480 36%
    960 91%

    Meanwhile, the false positive rate when only $b$ moves is 0% for both $n=120$ and $n=480$ (nominal level 5%). Thus, this test is “conservative yet correct”—it never falsely accuses a rater of only $b$ moving, though it fails to detect true drift with small samples.

    Key conclusion: the conventional criterion of 120 cases per rater has only 1% power to detect cutpoint shape drift, equivalent to no test at all. Combining 20 raters with 120 cases each still yields 0% power—merging raters cannot recover shape drift at the individual level because drift is rater-specific in direction, and merging causes cancellation.

    To achieve this, a single rater requires more than 900 cases. This number directly determines that the question “whether cutpoints were altered by training” has no answer on existing datasets, rather than “the answer is insignificant.”

    VI. Boundaries of This Conclusion

    First, these calculations assume normal latent variables with fixed cutpoints. True grading data may not be normally distributed; heavy tails or bimodality alter specific moment values but do not change directionality: after discretization, location shifts inevitably alter shape, regardless of the distribution family (provided finite bins and finite bin widths).

    Second, equal-spacing cutpoints is a strong assumption; we use it solely to render the model identifiable, not because true cutpoints are equally spaced. A more robust approach parameterizes cutpoint shape (e.g., using two parameters controlling asymmetry) and then performs nested comparison.

    Third, this paper addresses only “shape drift.” Training may also alter the number of grades used by raters (category collapse), causing certain bin probabilities to approach zero. In such cases, the asymptotic chi-square approximation of the likelihood ratio fails, requiring exact tests or parametric bootstrapping.

    Evidence and Limitations

    Verified:
    – Under fixed cutpoints in discrete grading, moving only location necessarily changes skewness: our grid calculation shows skewness spanning 1.27 and kurtosis spanning 0.67 as location shifts from −1.0 to +1.0
    – In the Multidimensional Rasch model, rater location parameters and time-specific severity drift are modeled objects, with the assumption of time independence explicitly criticized as inefficient (PMID 36284065)
    – Rater leniency/stringency is an estimable, reproducible effect across time (PMID 16919156: 1,259 examiners graded 142,030 items, with 12% of main effect variance attributable to examiner leniency; PMID 8322872 early confirmed significant rater severity)
    – Under small samples, confidence intervals for kappa-type coefficients are unreliable and require sufficient test cases (PMID 20473138)

    Partially Verified:
    – The five-grade rater effects in Multidimensional Rasch (including central tendency and leniency) are separable within the model (PMID 14523257), but this pertains to continuous latent variables, not discrete grade distributions
    – Training indeed alters grading behavior (PMID 21899969: specialized training reduces inter-observer variability; PMID 37127220: international differences in breast density grading), but both report changes in coefficients without reporting changes in grade distribution shape

    Pending Verification:
    – Our true skewness effect size (0.012) depends on the normal latent variable assumption and has not been re-verified using semi-parametric methods on real five-grade grading data
    – The figure “900 cases per rater needed to detect shape drift” was calculated under spacing stretched from 1.0 to 1.3 (large effect size); true effects smaller would require even larger samples
    – Testing methods under category collapse have not undergone power analysis

    Related Reading

    Evidence Level: C
    References: PMID 36284065, PMID 16919156, PMID 8322872, PMID 14523257, PMID 20473138, PMID 21899969, PMID 37127220

    L1 Network Status: This topic has not entered the L1-corresponding network. A full-text search of CORRESPONDENCE_NETWORK.md (1,881 lines) for “cutpoints” yields only one occurrence (AC2 operational instructions for setting cutpoints within entry E-060, unrelated to this topic), “grading” one occurrence (KB-496 integer binning risk, belonging to GDV parameter binning at a different layer), “skewness” zero occurrences, “kurtosis” zero occurrences, “shift” zero occurrences, “location parameter” zero occurrences, “identifiable” zero occurrences, “severity” zero occurrences; EDGE_TABLE_science-yixue.md (63 lines, 59 concept edges) contains no consistency coefficients or grade distribution layer concepts. Hence, this topic has no corresponding edge, and evidence level is capped at C.

    Evidence Level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • δ=5pp 的警报线是相对的:阳性率越低,这条线反而越难看清——φ(c) 在分子分母里互相抵消

    Evidence level: C

    上一轮说「位置漂移 0.126 个标准差就撞 5pp 警报线」。这一轮要问:这条线在真实判读数据里到底看不看得见?答案是——阳性率越低,这条线越难看见,而换来的「需要的漂移量更大」一点忙都帮不上。

    正文

    上一轮给了条漂亮的换算:把阳性率 5 个百分点(5pp)的判读改变,换算成潜变量刻度上的位置漂移量,结论是「c=50% 时只要 0.126 个标准差(σ),c=20% 时 0.167σ,c=10% 时 0.245σ,c=5% 时 0.363σ」。这个换算本身没错,但它有一个没被点破的漏洞:它只回答了「需要多大漂移」这个分子,从没回答「这个漂移量能不能被数据看见」这个分母。把两个答案拼在一起,结论会反转。

    先讲清数学。假设判读背后有一个连续潜变量 X(例如”异常程度”),读者用切点把 X 切成阳性/阴性两档。阳性率 c 就是切点右上方的高斯面积:c = 1−Φ(τ)。把切点移动 b 个标准差,阳性率的变化近似等于密度 φ(τ) 乘 b——φ(τ) 就是换算表里那个”斜率”。所以 b(5pp) ≈ 0.05/φ(c),这就是 0.126σ 到 0.363σ 的来源。

    但检出必须同时看分母。n 例判读里阳性例数是二项变量,比例的标准误是 SE = √[c(1−c)/n]。检验”阳性率变了 5pp”的 z 值是 5pp 除以 SE,φ(c) 根本不出现。也就是说:5pp 可不可检出,只由 c 和 n 决定,与 φ(c) 无关。0.126σ 那句”警报”,是拿”恰好产生 5pp 的漂移”反代回去凑的——它必然等于”5pp 本身的 z 值”,φ(c) 在分子分母里对消了。

    那真实判读研究里这条线看得见吗?以最常见的 n=120 为例(Gwet AC1 分层设计文献里的常出现量级,PMID 40298274):c=50% 时 SE=0.0645,5pp 的 z 只有 0.77——连一个标准误都不到,够不着 1.96;c=20% 时 z=0.97;c=10% 时 z=1.29。在 120 例的量级上,5pp 的警报线在绝大多数阳性率下根本检不出来。要让 5pp 达到 z=1.96,需 n 至少 1560(c=50%)到 8400(c=2%),远超判读信度研究的常规样本量。

    更麻烦的是低阳性率下有两个反向力量在打架。分子这头:阳性率越低,产生 5pp 所需的漂移越大(0.126σ → 0.363σ → 0.58σ → 0.77σ,跨度 6.1 倍)。分母那头:阳性率越低,SE 越小(0.0645 → 0.0128,跨度 5.0 倍),小到把”更大的漂移”抵消掉。精确算下来:「所需漂移」的跨度 6.1x 与「噪声地板 MDDb」的跨度 2.9x 是同向的——φ(c) 在分子分母里同时出现、方向相反,一进一出互相抵消。这就是”警报线是相对的”的精确含义:阳性率越低,这条线越难看清,而换来的更大漂移量一点忙都帮不上。

    第三个坑是方向不对称。c≤5% 时”阳性率再降 5pp”物理上不存在(降到 0 要把切点移到无穷远)。低阳性率场景(罕见征象、外气显著变化的判定)只有”升向”可行。所以那个换算表在 c<5% 时只能回答”升 5pp”,”降 5pp”那一半是虚的——对称性本身不成立。

    最后,φ 线性化在低阳性率下也失真。精确解是 b 等于两个分位点之差。c=50% 时线性化误差只有 −0.3%;c=10% 时 +16%;c=5% 时 +33%;c=2% 时 +79%;c=1% 时 +143%。换算表在低阳性率区段整体高估了所需漂移——把本已够吓人的结论又往大里说了近一倍。

    合起来,方法学结论很清楚:“0.126σ 就是警报”只在基线对半、且 n≥1500 的判读任务里成立。真实多读者判读研究(120 例量级、阳性率常偏一头)里,5pp 的判读改变既产生不了可检出的 z 值,也产生不了站得住的换算表——两者在低阳性率下一起失效。判读信度文献早警告过同源问题:Cohen κ 对基率敏感、低基率下被严重惩罚,应改用 G 统计量并分开报阳性/阴性一致率(PMID 25090041);κ 值必须配基率与偏倚指标一起报,否则跨研究比较会误导(PMID 8501467)。2023 年一项国际多读者乳腺密度判读研究(PMID 37127220)是少见范例——同时报告四档边际分布与 Stuart-Maxwell 检验,补上了基率缺位,正是回溯研究可以抄的模板。

    对气功外气意味着什么?外气显著变化的判定是典型的低阳性率任务(异常变化本就少见)。在这个场景里,0.126σ~0.36σ 的位置漂移对判读结果的影响大概率淹没在比例噪声里——“位置漂移几乎不构成威胁”在罕见征象判读中反而是更可能成立的描述,而原来的警报把方向说反了。真正该盯的是切点(档位门槛)的形状变化——那是上一篇证明过、且低阳性率下更隐蔽的漂移来源。

    证据与局限

    已验证:
    – 5pp 在 n=120 下、c=50% 时 z 仅 0.77,远不显著;c=10% 时 z=1.29,仍不显著。
    – φ(c) 在分子(所需漂移)和分母(噪声地板)中同向出现、方向相反,一进一出对消;”所需漂移的跨度 6.1x”与”噪声地板的跨度 2.9x”是同向的。
    – 方向不对称:c≤5% 时”降 5pp”物理上不存在,只有升向可行。
    – φ 线性化在 c=2% 时误差 +79%、c=1% 时 +143%,#1459 换算表低阳性率区段整体高估。

    部分验证:
    – 判读信度文献对基率敏感的警告(PMID 25090041、8501467)与本文的数值推演一致;但文献讨论的是 κ 系数的惩罚,本文讨论的是比例检验的检出能力,两者是同一个”基率缺位”问题的两个侧面,直接换算关系未建立。
    – PMID 37127220 作为”同时报告边际分布 + 对称性检验”的方法模板,其阳性率区间(乳腺密度四档)与本系列关注的低阳性率判读场景不完全一致,迁移需谨慎。

    待验证:
    – 对 ≥30 项多读者判读研究做 φ(c) 回溯:提取报告的分类别阳性率(或由敏感度/特异度与病例构成反算),建立阳性率的经验分布;再按该分布对”5pp 警报线”的可检出性做加权,给出领域平均阈值而非单点。核心判据:若阳性率中位数 ≥30%,0.126σ 可作全局警戒值;若落在 10-30%,须按阳性率分三档报(≥30% 用 0.126 / 10-30% 用 0.167-0.245 / <10% 用 0.36 以上);若中位数 <10%,警报须整体撤下,改写为”罕见征象判读中位置漂移几乎不构成威胁”,检测重点从 b 转向切点。
    – 用精确二项/Fisher 精确检验替代 φ 线性化,重算低阳性率区段的 b 换算表,确认 +79%/+143% 的误差量级。
    – 检验”方向不对称”是否影响双尾判读研究的设计:当 c≤5% 时,双尾检验实际上只剩单尾可检验,功效计算需按单尾重算。

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 25090041, PMID 37127220, PMID 40298274
    L1 网络状态: E-060 标签前提(KB-334):”两类评定者的阳性率不同时,一致性系数与患病率必须分开报告,否则得出「标签不可靠」的假结论”;KB-317(conflict):”所有加权系数在「中低一致性 + 大样本 + 类别数多」时精度显著下降”。本文是该方法学冲突在”位置漂移检出”上的数值化延伸;该主题尚未进入 L1 对应网络(缺”判读-基率-漂移可检出性”边)。


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    The alert line for δ=5pp is relative: the lower the positivity rate, the harder it becomes to see this line—φ(c) cancels out in both numerator and denominator.

    Evidence level: C

    In the previous round, we stated that “a position drift of 0.126 standard deviations (σ) hits the 5pp alert line.” This round asks: Is this line actually visible in real diagnostic data? The answer is—the lower the positive rate, the harder this line becomes to see, and the larger drift required “helps not at all.”

    Main Text

    The previous round provided a neat conversion: translating a diagnostic shift of 5 percentage points (5pp) into the magnitude of position drift on the latent variable scale. The conclusion was “at c=50%, only 0.126σ is needed; at c=20%, 0.167σ; at c=10%, 0.245σ; at c=5%, 0.363σ.” This conversion itself is correct, but it has an unspoken flaw: it answers only the numerator—”how much drift is needed”—but never the denominator—”can this drift magnitude be seen by the data?”. Combining both answers reverses the conclusion.

    Let us clarify the mathematics. Suppose behind the diagnostic judgment lies a continuous latent variable X (e.g., “degree of abnormality”), and readers use a cut-off point to divide X into positive/negative categories. The positive rate c is the Gaussian area above the cut-off: c = 1−Φ(τ). Moving the cut-off by b standard deviations changes the positive rate approximately by density φ(τ) multiplied by b—where φ(τ) is that “slope” in the conversion table. Thus, b(5pp) ≈ 0.05/φ(c), which yields the 0.126σ to 0.363σ range.

    But detection must also consider the denominator. In a sample of n cases, the number of positives is a binomial variable, and the standard error of the proportion is SE = √[c(1−c)/n]. The z-value for testing “a 5pp change in positive rate” is 5pp divided by SE, where φ(c) does not appear. Therefore: whether 5pp is detectable depends only on c and n, not on φ(c). The “0.126σ alert” is derived by substituting back the drift that exactly produces 5pp—it necessarily equals “the z-value of 5pp itself,” so φ(c) cancels out in numerator and denominator.

    So, is this line visible in real diagnostic studies? Taking the common case of n=120 (a typical scale found in Gwet AC1 stratified design literature, PMID 40298274): at c=50%, SE=0.0645, so the z-value for 5pp is only 0.77—less than one standard error, far below 1.96; at c=20%, z=0.97; at c=10%, z=1.29. At the scale of 120 cases, the 5pp alert line is essentially invisible across most positive rates. To make 5pp reach z=1.96, n must be at least 1560 (for c=50%) up to 8400 (for c=2%), far exceeding conventional sample sizes in diagnostic reliability studies.

    More problematically, at low positive rates two opposing forces clash. On the numerator side: the lower the positive rate, the larger the drift needed to produce 5pp (0.126σ → 0.363σ → 0.58σ → 0.77σ, spanning 6.1 times). On the denominator side: the lower the positive rate, the smaller the SE (0.0645 → 0.0128, spanning 5.0 times), which cancels out the “larger drift.” Precisely calculated: the span of “required drift” (6.1x) and the span of “noise floor MDDb” (2.9x) move in the same direction—φ(c) appears in both numerator and denominator but with opposite directions, so they cancel each other out. This is the precise meaning of “the alert line is relative”: the lower the positive rate, the harder this line becomes to see, and the larger drift required helps not at all.

    A third pitfall is directional asymmetry. When c≤5%, “a further drop of 5pp in positive rate” is physically impossible (dropping to zero would require moving the cut-off to infinity). Low-positive-rate scenarios (rare signs, judgments of significant external Qi changes) allow only “rising” shifts. Thus, the conversion table can answer “rise by 5pp” for c<5%, but “drop by 5pp” is purely hypothetical—the symmetry itself fails.

    Finally, linearizing φ becomes distorted at low positive rates. The exact solution is b equaling the difference between two quantiles. At c=50%, linearization error is only −0.3%; at c=10%, +16%; at c=5%, +33%; at c=2%, +79%; at c=1%, +143%. The conversion table overestimates required drift across the low-positive-rate region—doubling an already alarming conclusion.

    Putting it together, the methodological conclusion is clear: “0.126σ is the alert” holds only when baseline is 50% and n≥1500. In real multi-reader diagnostic studies (scale of 120 cases, positive rates often skewed), the 5pp diagnostic shift neither generates a detectable z-value nor produces a robust conversion table—both fail at low positive rates. Diagnostic reliability literature has long warned of this base-rate problem: Cohen κ is sensitive to base rate and severely penalized at low base rates; one should use G statistics and report positive/negative agreement rates separately (PMID 25090041); κ values must be reported alongside base rate and bias indices, otherwise cross-study comparisons will mislead (PMID 8501467). A 2023 international multi-reader breast density diagnostic study (PMID 37127220) is a rare example—reporting all four marginal distributions and the Stuart-Maxwell test, filling the base-rate gap, serving as a template for retrospective studies.

    What does this mean for Qi Gong external Qi? Judging significant external Qi changes is a classic low-positive-rate task (abnormal changes are rare). In this scenario, position drifts of 0.126σ~0.36σ likely get drowned in proportion noise—“position drift hardly constitutes a threat” is actually the more accurate description for rare-sign detection, while the original alert gets the direction wrong. What truly matters is the shape change of the cut-off (threshold)—which was proven in the previous article and is even more hidden at low positive rates.

    Evidence and Limitations

    Verified:
    – At n=120, 5pp yields z=0.77 at c=50% (not significant); z=1.29 at c=10% (still not significant).
    – φ(c) appears in both numerator (required drift) and denominator (noise floor) but with opposite directions, so they cancel; the span of “required drift” (6.1x) and “noise floor MDDb” (2.9x) move together.
    – Directional asymmetry: at c≤5%, “drop 5pp” is physically impossible; only rising shifts are feasible.
    – φ linearization errors: +79% at c=2%, +143% at c=1%; the #1459 conversion table overestimates drift in low-positive-rate regions.

    Partially Verified:
    – Warnings from diagnostic reliability literature about base-rate sensitivity (PMID 25090041, 8501467) align with this numerical derivation; however, those papers discuss κ coefficient penalties, whereas this paper discusses proportion-test detectability—both reflect the same “missing base rate” issue from different angles, without direct conversion established.
    – PMID 37127220 serves as a methodological template for “reporting marginal distributions + symmetry tests,” but its positive-rate intervals (four breast density categories) do not fully match the low-positive-rate diagnostic scenarios of this series; transfer requires caution.

    To Be Verified:
    – Back-calculate φ(c) across ≥30 multi-reader diagnostic studies: extract reported categorical positive rates (or derive from sensitivity/specificity and case mix), build an empirical distribution of positive rates; then weight the “5pp alert line” detectability against this distribution to yield domain-average thresholds rather than single points. Core criterion: if median positive rate ≥30%, 0.126σ serves as global alert; if falling in 10-30%, report three tiers based on positive rate (≥30% uses 0.126; 10-30% uses 0.167–0.245; <10% uses 0.36 or higher); if median <10%, remove the alert entirely and rewrite as “in rare-sign detection, position drift hardly constitutes a threat,” shifting focus from b to the cut-off.
    – Replace φ linearization with exact binomial/Fisher exact tests, recalculate the b conversion table for low-positive-rate regions to confirm the +79%/+143% error magnitude.
    – Test whether “directional asymmetry” affects two-tailed diagnostic study design: when c≤5%, the two-tailed test effectively becomes one-tailed; power calculations must be recomputed accordingly.

    Further Reading

    Evidence Level: C
    References: PMID 8501467, PMID 25090041, PMID 37127220, PMID 40298274
    L1 Network Status: E-060 label prerequisite (KB-334): “When two raters have different positive rates, agreement coefficients and prevalence must be reported separately; otherwise false conclusions of ‘unreliable labels’ arise”; KB-317 (conflict): “all weighted coefficients suffer significant precision loss under ‘medium-low agreement + large sample + many categories’.” This paper extends that methodological conflict numerically to “position drift detection”; this topic has not yet entered the corresponding L1 network (missing the “diagnosis-base rate-drift detectability” edge).


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 那个「一致率越低通道越大」的发现,量的是角落里的一根线:δ 通道恒等于平方律,跟 P_o 的关联是假的

    Evidence level: C

    δ 通道对 κ 的贡献有闭式解,只与 δ 的平方成正比;它随 P_o 变大,是因为取的 δ 上界本身随 P_o 变大,不是有哪个机制在放大。

    正文

    一道算术,先确认它是对的

    上一轮留下一个数:在 P_o=0.85、边际均值 0.43 下,把评分者之间的系统偏置 δ 从可行上界一路降到零,Cohen κ 最多只能涨 0.0069。若把 P_o 挪到 0.75,同一算式给出的上限是 0.0306,是前者的 4.4 倍;挪到 0.70 则是 0.0515,7.5 倍。

    这串数字先要复算。二分类 2×2 表的期望一致率有闭式表达:p_e = 1 − 2m̄ + 2m̄² − δ²/2,其中 m̄ 是两个评分者阳性率的平均,δ 是二者之差。κ =(P_o − p_e)/(1 − p_e)。把 P_o 从 0.90 扫到 0.65、δ 一律取到上界 1 − P_o,得到的增幅依次是 0.0021、0.0069、0.0160、0.0306、0.0515、0.0793。原题列的五个点逐位吻合。算术没错。

    错的是对这条曲线的解释。

    那条「上界」其实是退化的

    题目的推导把 δ 的可行上界定为 1 − P_o:一致率越低,剩下的不一致就越多,δ 可用的「空间」就越大。听起来像个资源约束。但把它代进四格表就露了底:当 δ 恰好等于 1 − P_o 时,表里 p₀₁ 那一格恒等于零。二十组参数、四个 P_o 水平、三个边际水平,p₀₁ 全是 ±5.6e−17 量级的浮点零。

    原因是一行恒等式:p₀₁ + p₁₀ 恒等于 1 − P_o。总不一致率就是这个数,跑不掉。而 δ 是 p₀₁ 与 p₁₀ 之差的绝对值,由三角不等式,δ ≤ p₀₁ + p₁₀ = 1 − P_o。所以「δ 的上界等于 1 − P_o」不是哪个区间打开了,而是「把所有不一致全部压到一侧」这个奇点。δ 取到上界,等于说其中一个评分者在某一个方向上从不犯错。这是退化,不是自由度。

    拿这个角落去代表 δ 通道的量级,等于用一个人为构造的极端点给整条曲线定价。

    真正的机制是平方律,不是 P_o

    把 δ 固定住,再扫 P_o,曲线立刻变得温顺。δ=0.05 时,P_o 从 0.92 降到 0.65,增幅只从 0.00042 走到 0.00182,全程没越过 0.002;δ=0.10 时从 0.00165 走到 0.00721。换句话说,前面那条「两个数量级」的陡坡,几乎全部来自 δ 自己在动,而不是 P_o 在放大什么。

    闭式解能把这件事说死。做一次代数化简:

    δκ = (δ²/2)(1 − P_o) / [v(v + δ²/2)],其中 v = 2m̄(1 − m̄)。

    在 P_o 从 0.60 到 0.95、m̄ 从 0.05 到 0.95、δ 取遍可行域的网格上回代,最大绝对误差 2.9e−15。这条公式说明了三件事。

    第一,δ 通道是二次的。δ 减半,通道降到四分之一。δ 从 0.05 走到 0.10(翻倍),通道从 0.0013 走到 0.0051(约四倍)。这不是近似,是恒等。所以「δ 通道能不能忽略」这个问题不该问 P_o,该问 δ 有多大。

    第二,符号无关。δ=+0.05 与 δ=−0.05 给出完全相同的 κ(0.491301)。谁比谁严,κ 一个字都不告诉。这个量测的是「偏置有多大」,不是「往哪偏」。

    第三,P_o 确实在公式里,但它是一阶的乘子,不是放大器;真正能让通道变大的分母是 v = 2m̄(1 − m̄)。m̄ 越靠近 0 或 1,v 越小,同样的 δ 打出的通道越大。m̄=0.43 时 v=0.49;m̄=0.88 时 v=0.21。所以低一致率若伴随极端边际,通道会被抬高——但那是 v 干的,不是 P_o 干的。

    三个约束,不止一个

    还有一层:δ ≤ 1 − P_o 只是三个约束之一。四格非负还要求 δ ≤ 2m̄ 与 δ ≤ 2(1 − m̄)。实际可用上界是三者的最小值。m̄=0.10、P_o=0.70 时,绑住的是 2m̄=0.20,不是 1 − P_o=0.30;m̄=0.88、P_o=0.75 时,绑住的是 2(1 − m̄)=0.24,也不是 0.25。原题只用了 1 − P_o,等于默认 m̄ 落在中间那一段,而这恰恰是需要论证而不是默认的。

    真实的 δ 有多大

    这一步决定了上面所有数字有没有现实意义。若 δ 只是 0.05——两个评分者阳性率差五个百分点,在判读研究里已经算明显——那么即使 P_o 压到 0.70,通道也只有 0.00156,是原题 0.007 那条线的两成。要让通道越过 0.007,δ 至少要 0.10 以上,且在 P_o=0.80 时要达到可行域的 65%、P_o=0.75 时达到 47%。

    真实的 δ 分布没人系统测过,这仍是缺口。但可以从已发表数据里读出一两个锚点。一项胸部病理判读研究(PMID 38283651)报告肺腺癌五种组织学型的判读:acinar 型的患病率 88%、观察一致率 88%、κ=0.43、AC1=0.85、阳性一致率 93% 对阴性一致率 50%——这是「κ 悖论」的教科书案例。把 P_o=0.88、κ=0.43 代进公式反解,唯一解是 m̄=0.88、δ=0.00。回代验算:κ 的 p_e=0.7888,得 κ=0.432;AC1 的 p_e=0.2112,得 AC1=0.848。两个都对上。也就是说,这个被反复引用的悖论案例里,δ 通道的贡献精确为零,κ 被压低完全来自 v=0.21 这个分母。

    按评分者敏感度/特异度设定做的小规模模拟也指向同一方向:两个评分者能力相同时 δ 约 0.001–0.000;只有刻意让一人偏松一人偏严(敏感度 0.85 对 0.75、特异度 0.85 对 0.90),δ 才到 0.075,占不一致预算的 28%,通道 0.0031。

    那 0.007 到底怎么用

    原题的结论「δ 通道可忽略」在 P_o≥0.85 时成立,这个方向性判断没错,但支撑它的理由要换掉:不是因为高一致率压住了通道,而是因为真实 δ 本来就小。反过来,原题提的「若领域 P_o 中位数低于 0.78 则结论须反转」这条也不成立——P_o 低并不自动让 δ 变大,它只是把 δ 的上界抬高,而抬到那个上界意味着表已经退化。

    真正该做的前置检查只有一条:报 δ,或者报足以算出 δ 的两组边际。这正是三十年前就被提出的要求(PMID 8501467):比较一致性研究时只报 κ 会误导,应同时给出偏置与患病率的定量指标。也有作者主张 P_o 的依赖根本不构成反对 κ 的理由(PMID 15939215),但那是从潜变量患病率角度讲的,与这里「必须报边际才能反解 δ」并不冲突。GRRAS 报告规范(PMID 21130355)也把边际与一致率列入必备项。

    回到那条更早的结论:训练把 δ 降到零,κ 最多涨 0.0069。这个上限现在可以更精确地说:它是「δ 从可行域角落走到零」的上限,是个天花板上的天花板。真实训练若把 δ 从 0.05 压到 0.02,在 P_o=0.75、m̄=0.43 下只值 0.0011;角落给的 0.0306 是它的 28 倍,P_o=0.70 时是 39 倍。

    证据与局限

    已验证: δ 通道的闭式解 δκ = (δ²/2)(1 − P_o)/[v(v + δ²/2)] 在 35×91×可行域网格上回代最大误差 2.9e−15,属恒等式级别。p₀₁ + p₁₀ = 1 − P_o 与 δ ≤ 1 − P_o 的三角不等式来源亦为精确推导。PMID 38283651 的 acinar 案例反解出 δ=0.000,κ 与 AC1 两个值同时复现。

    部分验证: 真实判读研究中 δ 的分布没有系统测量,本文的 0.02–0.08 区间来自评分者敏感度/特异度模拟与一项研究的反解,属估计而非普查。κ 在极小样本下的抽样波动参考 PMID 15733050 与 PMID 21969984 的区间估计框架,本文的 SE 数值来自 6000 次自助模拟,未与解析公式逐点对表。

    待验证: ① 对 ≥30 项判读训练研究做边际回溯,算出 δ 的实测中位数与四分位,才能判定 0.007 这个量级是普遍成立还是只在特定子域成立;② 三评分者及多档情形下,δ 通道是否仍守平方律、v 是否换成广义形式;③ 当 P_o 与 m̄ 同时被报告时,(P_o, κ, m̄) 三元组能否唯一定出 δ——本文只验证了二分类两点情形,一般情形需要可辨识性证明。

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 15939215, PMID 38283651, PMID 19159474, PMID 21130355, PMID 25090041

    L1 网络状态:该主题尚未进入 L1 对应网络 —— 边表 63 行无「一致性系数 / 评分者间信度 / 判读训练」相关边;网络 1881 行中「一致率」仅 1 处、「判读」仅 1 处、「κ」3 处,全部落在 E-060 标签前提(conflict 未裁决,涉及 Gwet AC1/AC2 与 Brennan-Prediger 的精度区间错配),与本条讨论的 δ 通道量级无对应关系。

    证据等级说明:A=有 PubMed 直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索。本文属方法学推演,核心结论为本文自行推导的闭式解与模拟,PubMed 文献仅提供报告规范与实例锚点,故按 C 处理。


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    The finding that \”the lower the consistency rate, the larger the channel\” measures only a single line in the corner: the δ channel is constantly equal to the square law, and its…

    Evidence level: C

    The closed-form solution for the δ-channel’s contribution to κ scales strictly with δ²; it increases with P_o solely because the chosen upper bound for δ itself grows with P_o, not because any mechanism amplifies it.

    Body

    One Arithmetic Step to Confirm It Is Correct

    A number left over from the previous round: with P_o = 0.85 and marginal means of 0.43, driving the systematic bias δ between raters down from its feasible upper bound to zero can increase Cohen’s κ by at most 0.0069. If P_o is shifted to 0.75, the same formula yields an upper limit of 0.0306, which is 4.4 times the former; at 0.70 it becomes 0.0515, or 7.5 times.

    This sequence of numbers must first be re-calculated. The expected agreement for a 2×2 contingency table has a closed-form expression: p_e = 1 − 2m̄ + 2m̄² − δ²/2, where m̄ is the average of the two raters’ positive rates and δ is their difference. κ = (P_o − p_e)/(1 − p_e). Sweeping P_o from 0.90 down to 0.65 while setting δ always to its upper bound of 1 − P_o yields increments of 0.0021, 0.0069, 0.0160, 0.0306, 0.0515, and 0.0793 respectively. The five points listed in the original problem match digit-for-digit. The arithmetic is correct.

    The error lies in the interpretation of this curve.

    That “Upper Bound” Is Actually Degenerate

    The original derivation sets the feasible upper bound for δ at 1 − P_o: lower agreement means more remaining disagreement, so the available “space” for δ is larger. This sounds like a resource constraint. But substituting it into the four-cell table reveals the bottom: when δ exactly equals 1 − P_o, the cell p₀₁ in the table is constantly zero. Across twenty parameter sets and four P_o levels, p₀₁ is a floating-point zero of ±5.6e−17 magnitude.

    The reason is an identity: p₀₁ + p₁₀ always equals 1 − P_o. Total disagreement is this number, which cannot escape. Since δ is the absolute difference between p₀₁ and p₁₀, the triangle inequality gives δ ≤ p₀₁ + p₁₀ = 1 − P_o. Thus “δ’s upper bound equals 1 − P_o” is not an interval opening up, but the singularity where “all disagreement is pushed to one side.” When δ reaches its upper bound, it means one rater never makes a mistake in one direction. This is degeneracy, not degrees of freedom.

    Using this corner to represent the scale of the δ-channel is like pricing the entire curve using an artificially constructed extreme point.

    The True Mechanism Is a Square Law, Not P_o

    Fixing δ and sweeping P_o immediately makes the curve gentle. At δ = 0.05, lowering P_o from 0.92 to 0.65 increases the increment only from 0.00042 to 0.00182, never exceeding 0.002; at δ = 0.10 it goes from 0.00165 to 0.00721. In other words, the steep slope of “two orders of magnitude” previously attributed mostly comes from δ moving itself, not P_o amplifying anything.

    The closed-form solution makes this definitive. After algebraic simplification:

    δκ = (δ²/2)(1 − P_o) / [v(v + δ²/2)], where v = 2m̄(1 − m̄).

    Back-substituting across the grid of P_o from 0.60 to 0.95, m̄ from 0.05 to 0.95, and δ over its feasible domain yields a maximum absolute error of 2.9e−15. This formula establishes three facts.

    First, the δ-channel is quadratic. Halving δ reduces the channel to one-quarter; doubling δ from 0.05 to 0.10 increases the channel from 0.0013 to 0.0051 (roughly four times). This is not an approximation, but an identity. Therefore the question “Can the δ-channel be ignored?” should not ask about P_o, but about how large δ is.

    Second, the sign is irrelevant. δ = +0.05 and δ = −0.05 yield identical κ values (0.491301). Who is stricter tells κ nothing. This measurement captures “how much bias exists,” not “which direction it biases toward.”

    Third, while P_o does appear in the formula, it is a first-order multiplier, not an amplifier; the denominator v = 2m̄(1 − m̄) is what truly allows the channel to grow. As m̄ approaches 0 or 1, v shrinks, so the same δ produces a larger channel. When m̄ = 0.43, v = 0.49; when m̄ = 0.88, v = 0.21. Thus low agreement accompanied by extreme margins elevates the channel—but that is due to v, not P_o.

    Three Constraints, Not Just One

    There is another layer: δ ≤ 1 − P_o is only one of three constraints. Non-negativity in the four-cell table also requires δ ≤ 2m̄ and δ ≤ 2(1 − m̄). The actual usable upper bound is the minimum of these three. When m̄ = 0.10 and P_o = 0.70, the binding constraint is 2m̄ = 0.20, not 1 − P_o = 0.30; when m̄ = 0.88 and P_o = 0.75, it is 2(1 − m̄) = 0.24, not 0.25. The original problem used only 1 − P_o, effectively assuming m̄ falls in the middle segment, which is precisely what needs argumentation rather than default assumption.

    How Large Is the Real δ

    This step determines whether all the above numbers have practical meaning. If δ is merely 0.05—meaning the difference between two raters’ positive rates is five percentage points, which is already noticeable in diagnostic studies—then even with P_o pressed down to 0.70, the channel is only 0.00156, just twenty percent of the original 0.007 line. To push the channel beyond 0.007, δ must be at least 0.10, reaching 65% of the feasible domain when P_o = 0.80 and 47% when P_o = 0.75.

    No systematic measurement of the true δ distribution exists; this remains a gap. However, one or two anchors can be read from published data. A chest pathology diagnosis study (PMID 38283651) reports the diagnosis of five histological types of lung adenocarcinoma: acinar type has a prevalence of 88%, observed agreement of 88%, κ = 0.43, AC1 = 0.85, positive agreement of 93% versus negative agreement of 50%—this is the textbook case of the “κ paradox.” Substituting P_o = 0.88 and κ = 0.43 into the formula yields a unique solution: m̄ = 0.88 and δ = 0.00. Back-substitution verifies: κ’s p_e = 0.7888, giving κ = 0.432; AC1’s p_e = 0.2112, giving AC1 = 0.848. Both match perfectly. Thus, in this repeatedly cited paradox case, the δ-channel’s contribution is precisely zero; κ is suppressed entirely due to the denominator v = 0.21.

    Small-scale simulations based on rater sensitivity/specificity settings also point in the same direction: when both raters have equal ability, δ is approximately 0.001–0.000; only when one is deliberately set loose and the other tight (sensitivity 0.85 vs 0.75, specificity 0.85 vs 0.90) does δ reach 0.075, accounting for 28% of the disagreement budget, yielding a channel of 0.0031.

    So How Is That 0.007 Actually Used

    The original conclusion that “the δ-channel is negligible” holds when P_o ≥ 0.85, and this directional judgment is correct, but the supporting reason must be replaced: it is not because high agreement suppresses the channel, but because the true δ is inherently small. Conversely, the claim in the original that “if the domain’s P_o median falls below 0.78, the conclusion must reverse” is also invalid—low P_o does not automatically make δ larger; it merely raises δ’s upper bound, and reaching that upper bound implies the table has degenerated.

    The only prerequisite check that truly matters is this: report δ, or report the two marginal values sufficient to calculate δ. This requirement was proposed thirty years ago (PMID 8501467): comparing consistency studies reporting only κ is misleading; quantitative measures of bias and prevalence must also be provided. Some authors argue that dependence on P_o fundamentally does not oppose κ (PMID 15939215), but this is stated from the perspective of latent prevalence and does not conflict with “marginals must be reported to solve for δ.” The GRRAS reporting guidelines (PMID 21130355) also list marginals and agreement as mandatory items.

    Returning to the earlier conclusion: training δ down to zero can increase κ by at most 0.0069. This upper limit can now be stated more precisely: it is the ceiling for “δ moving from the feasible corner to zero,” a ceiling-on-a-ceiling. Real-world training reducing δ from 0.05 to 0.02, with P_o = 0.75 and m̄ = 0.43, yields only 0.0011; the corner-derived 0.0306 is 28 times larger, and at P_o = 0.70 it is 39 times larger.

    Evidence and Limitations

    Verified: The closed-form solution for the δ-channel, δκ = (δ²/2)(1 − P_o)/[v(v + δ²/2)], was back-substituted across a 35×91×feasible-domain grid with a maximum absolute error of 2.9e−15, establishing it as an identity-level result. The derivations of p₀₁ + p₁₀ = 1 − P_o and δ ≤ 1 − P_o via the triangle inequality are also exact. The acinar case from PMID 38283651 was solved backwards to yield δ = 0.000, successfully reproducing both κ and AC1 values.

    Partially Verified: The distribution of δ in real diagnostic studies has not been systematically measured; the 0.02–0.08 interval in this paper stems from rater sensitivity/specificity simulations and the back-solution of one study, representing estimation rather than census. Sampling variability of κ under extremely small samples references the interval estimation frameworks of PMID 15733050 and PMID 21969984; the SE values in this paper derive from 6,000 bootstrap simulations and were not matched point-by-point against analytical formulas.

    Pending Verification: ① Conduct marginal back-calculation on ≥30 diagnostic training studies to compute the empirical median and quartiles of δ, thereby determining whether the 0.007 magnitude is universally valid or restricted to specific subdomains; ② In multi-rater and multi-level scenarios, does the δ-channel still obey the square law, and does v adopt a generalized form; ③ When both P_o and m̄ are reported, can the triplet (P_o, κ, m̄) uniquely determine δ—this paper verified only the binary two-point case, while general cases require identifiability proofs.

    Related Reading

    Evidence Level: C
    References: PMID 8501467, PMID 15939215, PMID 38283651, PMID 19159474, PMID 21130355, PMID 25090041

    L1 Network Status: This topic has not yet entered the corresponding L1 network—the edge table row 63 lacks edges related to “consistency coefficient / inter-rater reliability / diagnostic training”; among 1,881 rows in the network, “agreement rate” appears only once, “diagnosis” once, and “κ” three times, all falling under the E-060 label premise (conflict unresolved, involving mismatched precision intervals between Gwet’s AC1/AC2 and Brennan-Prediger), which bears no correspondence to the δ-channel magnitude discussed here.

    Evidence Level Explanation: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration. This paper belongs to methodological derivation; core conclusions are the closed-form solution and simulations derived herein, while PubMed literature provides only reporting guidelines and example anchors, hence classified as C.


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

  • 锚和读者共用一副眼睛时,误差不是被放大而是被抵消:那个「除以根号 1 减 ρ 平方」的修正方向反了

    Evidence level: C

    上一篇担心「资深者锚和普通读者共用一个脑袋,误差同向,所以锚的杀伤力会被放大」。把协方差正式写进模型之后,结论正好相反:同向误差在校正时被减掉两次,锚反而变得更有用。真正的危险不是锚和读者同向,而是锚和读者不同向。

    正文

    上一篇(#1422)把精度账算到了「校准人数才是主项」这一层。但它留了一个未检验的假定:锚的误差与读者的误差互相独立。现在这条假定被拿出来单算,因为「以资深者按皮温曲线判断变化幅度作锚」这个做法里,资深者和普通读者看的是同一条曲线、判的是同一个「变化够不够大」,误差很难不同向。

    原本的担心很自然:两项误差若是同向的,它们会叠加,于是提出把锚项按 s 除以根号下 1 减 ρ 平方来放大——ρ 到 0.7 时放大 1.4 倍,到 0.9 时放大 2.29 倍。这个式子看着眼熟,因为它是「两个独立误差相减」的标准结果。问题恰恰在这里:位置估计里锚是被减掉的,不是被加上的。

    一、锚进入模型的方式是减法,所以同向误差自己会抵消

    读者 i 在病例 j 上的读数是 z = x + b + e,锚是 y = x + h。有锚之后,读者位置估计是 b̂ = z̄ − Ȳ。把误差项单独拿出来看,b̂ 的误差就是 ē − h̄。

    两个误差相减,方差是 Var(e) 加 Var(h) 再减去两倍协方差。所以 ρ 越大,方差越小,不是越大。写成相对量:以无锚时的读者项 A = se²/(mn) 为单位,位置估计的相对方差是

    V/A = (1 − ρ²) + (ρ√m − τ)²

    其中 m 是每个读者对同一病例的重复判读次数,τ = s√m/se 是「锚噪声相对于读者噪声」的尺度。这个式子在 n=240、m=2、se=0.5 下与 4 万次蒙特卡洛模拟逐格对照,最大偏差 3.5e-4。

    ρ=0 时它退化成 #1422 的形式 1+τ²,与上一篇完全对上。ρ 大于 0 之后整条曲线往下压:τ=1.414(锚噪声等于读者噪声,即上一篇的平衡点)时,ρ=0 给 3.00,ρ=0.4 给 1.56,ρ=0.6 给 0.96,ρ=0.7 给 0.69。ρ=0.7 这一格,锚不但没有变成「用更差的尺子标定」,反而把精度提升了三成。

    二、真正该担心的是反方向

    把 τ 与 ρ 的关系整理成判据,加锚有净收益的窗口是

    τ 落在 ρ(√m − 1) 与 ρ(√m + 1) 之间

    窗口最宽处 τ = ρ√m,此处相对方差取到谷底 1 − ρ²。这条式子的意思很直白:ρ 越大,能被容忍的锚噪声就越大。* ρ=0.3 时只容忍 τ 在 0.12 到 0.72 之间(锚必须比读者准两倍以上),ρ=0.7 时容忍 0.29 到 1.69(锚可以比读者差将近七成还划算)。

    反过来说,ρ 为负才是灾难。锚和读者反向偏移意味着两者各自朝相反方向错——这不是「互相抵消」,因为在 b̂ = z̄ − Ȳ 里,一个正的读者误差减去一个负的锚误差等于相加。ρ=−0.5、τ=1.414 时相对方差是 5.0,是 ρ=0 时的 1.7 倍。所以这条线上的诊断问题应该改问:锚和读者会不会反向偏?而不是「它们会不会太像」。

    三、残差地板:加锚最多只能降到 1−ρ²

    谷底 1−ρ² 是个比放大倍数更硬的东西。ρ=0.7 时最好情况是把标准差降到无锚时的 0.714 倍,也就是最多改善 28.6%;ρ=0.5 时最多改善 13.4%;ρ=0.3 时只有 4.6%。这与 #1422 第五节那个共模地板是同一件事的两个说法:读者与锚共有的那部分误差,靠加锚这个动作动不了,因为它两边都有。

    更麻烦的是 ρ 还有一个由协方差矩阵正定性给出的上界。若 R 位读者彼此独立、每人各自与锚的误差相关系数为 ρ,那么三个人(读者一、读者二、锚)的协方差矩阵必须半正定,这要求 ρ ≤ 1/√R。两位读者时上限是 0.7071,四位 0.5,八位 0.354,十六位 0.25。也就是说,读者越多,锚与单个读者能共有的误差比例在数学上就越小,整套「加锚救精度」的收益上限随之收缩:两位读者最多降 29.3%,八位只剩 6.5%。

    四、那个「ρ 测不出来」的死结并不成立

    上一篇记下的难点是:没有第二个独立锚,就分不开「是锚错了」还是「读者错了」,因此 ρ 无法从已有数据估出来——要测 ρ 得先有独立锚,而独立锚正是要证的东西。

    这个论证在「只有一份判读数据」的前提下成立,但它忘了一件事:ρ 在这里不是一个自由参数,它改变的是同一批数据在不同设计下的方差。取同一批病例,人为往锚里注入两级已知噪声(例如先按原样读,再对锚施加一个可量化的扰动),得到两个相对方差 v₁、v₂。两式相减,ρ² 项直接消掉,只剩

    ρ = [(v₁−k₁) − (v₂−k₂)] ÷ [−2√m (τ₁−τ₂)]

    其中 k 是各级噪声对应的 τ²。模拟里真值 0.20 / 0.40 / 0.60 三点分别解出 0.195 / 0.364 / 0.596,s₀ 同时解出 0.42 / 0.44 / 0.40(真值 0.40)。不需要第二个独立锚,只需要第二次注入。代价是必须做一次可控的扰动实验,而这不是「已发表数据的二次分析」能替代的。

    五、还有一条更朴素的证据路径

    即便不做注入实验,ρ 的方向也能从差值通道读出来。两位读者的读数相减时,锚项被完整减掉两次,于是差值方差与 ρ 无关——模拟里 ρ 从 0 到 0.7,差值标准差在 0.046 到 0.033 之间小幅下降,这是两位读者共享了同一个共模源导致它们之间也变得更像,属二阶效应。真正有用的是绝对值通道与差值通道的比值:ρ=0 时两者相差一个固定的 √2 因子,ρ 越大比值越偏离。这条比值检验只需要一次常规多读者判读数据,不必扰动,但它只能给出 ρ 的量级,不能给出点估计。

    证据与局限

    已验证: 位置估计的误差是 ē − h̄,两者相减使同向误差抵消。相对方差式 (1−ρ²) + (ρ√m−τ)² 在 m=2、n=240 下与 4 万次模拟逐格相符(偏差 ≤3.5e-4),ρ=0 退化为上一篇的 1+τ²。加锚有净收益的 τ 窗口 [ρ(√m−1), ρ(√m+1)]、谷底 1−ρ²、以及 ρ ≤ 1/√R 的协方差上界,都是闭式可核的代数结论,不依赖任何特定领域参数。

    部分验证: 「ρ 可由两点注入反解」在模拟中给出 0.195/0.364/0.596(真值 0.20/0.40/0.60),s₀ 同时解出且误差在 0.04 内;这是合格量级的方法,但尚未在任何真实判读数据集上跑过,注入扰动的可实现方式(如何对一位资深者「定量地」加噪而他又察觉不到)在真实针灸/皮温判读场景里还没有先例可抄。相关方法学有外部支撑:Walter 等(PMID 22351623)证明参考标准的误差与待测误差若相关,敏感度/特异度的偏倚方向与大小取决于患病率与两类错误的协方差,可正可负,不能一律假定「正相关就高估」——与本篇「方向取决于符号而非大小」同构;Rifkin(PMID 7644859)给出相关误差下观测相关系数的分解式,并指出参考方法较差(与真值相关 <0.9)时误差相关可使斜率与相关系数显著偏离独立误差情形。

    待验证: ① 真实判读场景里 ρ 的符号尚无一例实测。若资深者与普通读者在同一屏幕上读同一条曲线,ρ 很可能为正;但资深者若因经验而系统性地「看得更细」(把微小变化读成显著),普通读者读得粗,两者的误差就可能反号,此时本篇结论整体翻转,加锚的代价从「收益打折」变成「精度倒退」。② 两点注入法的可实现性:注入噪声的方差必须先独立标定,若这一级标定本身带 20% 误差,解出的 ρ 误差有多大尚未算过。③ R 位读者这条上界假定读者彼此独立,真实的同一中心读者之间往往也有共模,此时上界放宽但地板抬高,两个效应谁大未算。

    L1 网络状态:CORRESPONDENCE_NETWORK.md 第 202 行 E-060「标签前提」边:KB-292(conflict):中医医师间诊断一致性在开放人群中「低(low)」,作者称为「令人担忧」(PMID 30383398);…但训练可改善(PMID 19538100)、量表化可行…→「用五脏标签做分组」这一步本身带噪声 —— 层级 冲突(未裁决),状态 conflict。第 261 行 E-060「KB-334(open):两类评定者的阳性率不同时,一致性系数与患病率必须分开报告」。本篇讨论的锚-读者误差相关性尚未作为独立边进入该网络,属「有边但边内缺一层」,故证据等级按硬规则封顶 C。

    相关阅读

    证据等级: C
    参考文献: PMID 22351623, PMID 7644859, PMID 9006315, PMID 25895962, PMID 28105775, PMID 40499027

    证据等级:A=有PubMed直接支撑,B=科学类比/间接证据,C=传统文化共识/初步探索


    免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

    When anchors and readers share the same eye, error is not amplified but canceled: the correction direction of \”dividing by the square root of 1 minus rho squared\” is reversed

    Evidence level: C

    In the previous post, we worried that “the senior anchor and the ordinary reader share one head, so their errors point in the same direction and the anchor’s lethality is amplified.” After formally writing covariance into the model, the conclusion is exactly opposite: co-directional errors are subtracted twice during correction, making the anchor actually more useful. The true danger is not that the anchor and reader share direction, but that they point in opposite directions.

    Body

    The previous post (#1422) calculated the precision account down to the layer where “the number of calibrated people is the main term.” However, it left an untested assumption: that the anchor’s error and the reader’s error are mutually independent. Now this assumption is taken out for separate calculation, because in the practice of “using a senior person to judge the magnitude of change based on the skin temperature curve as an anchor,” both the senior and the ordinary reader look at the same curve and judge the same “whether the change is large enough,” making it difficult for their errors to point in different directions.

    The original concern was natural: if the two errors are co-directional, they will add up. Thus, it was proposed to amplify the anchor term by dividing $s$ by the square root of $1$ minus $\rho$ squared. When $\rho$ reaches 0.7, this amplifies by 1.4 times; when $\rho$ reaches 0.9, it amplifies by 2.29 times. This formula looks familiar because it is the standard result for “subtracting two independent errors.” The problem lies precisely here: in position estimation, the anchor is subtracted, not added.

    I. Because the Anchor Enters the Model via Subtraction, Co-directional Errors Cancel Themselves Out

    Reader $i$’s reading on case $j$ is $z = x + b + e$, and the anchor is $y = x + h$. With the anchor, the reader’s position estimate becomes $\hat{b} = \bar{z} – \bar{Y}$. Isolating the error terms, the error of $\hat{b}$ is $\bar{e} – \bar{h}$.

    When subtracting two errors, the variance is $\text{Var}(e)$ plus $\text{Var}(h)$ minus twice the covariance. Therefore, the larger $\rho$ is, the smaller the variance becomes, not the larger. Expressed in relative terms: taking the reader term $A = s_e^2/(mn)$ without an anchor as the unit, the relative variance of position estimation is

    $$
    V/A = (1 – \rho^2) + (\rho\sqrt{m} – \tau)^2
    $$

    where $m$ is the number of repeated readings each reader makes on the same case, and $\tau = s\sqrt{m}/s_e$ is the scale of “anchor noise relative to reader noise.” This formula matches, cell-by-cell, 40,000 Monte Carlo simulations under $n=240$, $m=2$, and $s_e=0.5$, with a maximum deviation of $3.5 \times 10^{-4}$.

    When $\rho=0$, it degenerates into the form $1+\tau^2$ from #1422, perfectly matching the previous post. Once $\rho > 0$, the entire curve shifts downward: when $\tau=1.414$ (anchor noise equals reader noise, i.e., the balance point from the previous post), $\rho=0$ gives 3.00, $\rho=0.4$ gives 1.56, $\rho=0.6$ gives 0.96, and $\rho=0.7$ gives 0.69. At $\rho=0.7$, the anchor not only does not become “calibrating with a worse ruler,” but instead improves precision by 30%.

    II. What Should Truly Be Worried About Is the Opposite Direction

    Organizing the relationship between $\tau$ and $\rho$ into a criterion, the window where adding an anchor yields net gain is

    $$
    \tau \text{ falls between } \rho(\sqrt{m} – 1) \text{ and } \rho(\sqrt{m} + 1).
    $$

    The widest part of this window is $\tau^ = \rho\sqrt{m}$, where the relative variance reaches its valley at $1 – \rho^2$. This formula’s meaning is straightforward: the larger $\rho$ is, the larger the anchor noise that can be tolerated.* When $\rho=0.3$, only $\tau$ between 0.12 and 0.72 is tolerated (the anchor must be more than twice as accurate as the reader); when $\rho=0.7$, tolerance ranges from 0.29 to 1.69 (the anchor can be nearly 70% worse than the reader and still be worthwhile).

    Conversely, negative $\rho$ is disaster. If the anchor and reader shift in opposite directions, they each err in opposite ways—this is not “mutual cancellation,” because in $\hat{b} = \bar{z} – \bar{Y}$, a positive reader error minus a negative anchor error equals addition. When $\rho=-0.5$ and $\tau=1.414$, the relative variance is 5.0, which is 1.7 times that of $\rho=0$. Therefore, the diagnostic question on this line should be changed to: will the anchor and reader shift in opposite directions? rather than “will they be too similar.”

    III. Residual Floor: Adding an Anchor Can At Most Reduce to $1-\rho^2$

    The valley $1-\rho^2$ is a harder constraint than the amplification factor. When $\rho=0.7$, the best case is reducing the standard deviation to 0.714 times that without an anchor, meaning at most a 28.6% improvement; when $\rho=0.5$, at most 13.4%; and when $\rho=0.3$, only 4.6%. This is the same thing as the common-mode floor mentioned in Section V of #1422, just expressed differently: the part of error shared between reader and anchor cannot be moved by adding the anchor, because it exists on both sides.

    More troubling is that $\rho$ also has an upper bound given by the positive definiteness of the covariance matrix. If $R$ readers are mutually independent, and each has an error correlation coefficient $\rho$ with the anchor, then the covariance matrix of three people (reader one, reader two, and the anchor) must be semi-positive definite, requiring $\rho \leq 1/\sqrt{R}$. With two readers, the upper limit is 0.7071; with four, 0.5; with eight, 0.354; and with sixteen, 0.25. This means the more readers there are, the smaller the proportion of shared error between the anchor and a single reader becomes mathematically, causing the upper limit of the entire “adding an anchor to improve precision” benefit to contract: with two readers, at most 29.3% reduction; with eight readers, only 6.5%.

    IV. The Deadlock That “$\rho$ Cannot Be Measured” Does Not Hold

    The difficulty noted in the previous post was: without a second independent anchor, one cannot distinguish “is the anchor wrong” from “is the reader wrong,” so $\rho$ cannot be estimated from existing data—to measure $\rho$, one must first have an independent anchor, which is exactly what needs to be proven.

    This argument holds under the premise of “only one set of reading data,” but it forgets one thing: here $\rho$ is not a free parameter; it changes the variance of the same batch of data under different designs. Taking the same batch of cases, artificially inject two levels of known noise into the anchor (e.g., first read as is, then apply a quantifiable perturbation to the anchor), yielding two relative variances $v_1$ and $v_2$. Subtracting the two equations, the $\rho^2$ term cancels directly, leaving

    $$
    \rho = \frac{(v_1 – k_1) – (v_2 – k_2)}{-2\sqrt{m}(\tau_1 – \tau_2)}
    $$

    where $k$ is the $\tau^2$ corresponding to each level of noise. In simulation, true values at 0.20 / 0.40 / 0.60 yield solutions of 0.195 / 0.364 / 0.596, and $s_0$ simultaneously solves to 0.42 / 0.44 / 0.40 (true value 0.40). No second independent anchor is needed; only a second injection. The cost is that one must perform a controllable perturbation experiment, which cannot be replaced by “secondary analysis of published data.”

    V. There Is Also a Simpler Evidence Path

    Even without injection experiments, the direction of $\rho$ can be read from the difference channel. When subtracting readings of two readers, the anchor term is completely subtracted twice, so the difference variance is independent of $\rho$—in simulation, as $\rho$ goes from 0 to 0.7, the difference standard deviation drops slightly from 0.046 to 0.033, because two readers sharing the same common-mode source also become more similar to each other, a second-order effect. What is truly useful is the ratio of the absolute value channel to the difference channel: when $\rho=0$, they differ by a fixed $\sqrt{2}$ factor; the larger $\rho$ is, the more this ratio deviates. This ratio test requires only one conventional multi-reader reading dataset, no perturbation needed, but it can only give the order of magnitude of $\rho$, not a point estimate.

    Evidence and Limitations

    Verified: The error of position estimation is $\bar{e} – \bar{h}$, and subtracting them cancels co-directional errors. The relative variance formula $(1-\rho^2) + (\rho\sqrt{m}-\tau)^2$ matches 40,000 simulations cell-by-cell under $m=2$, $n=240$ (deviation $\leq 3.5 \times 10^{-4}$); when $\rho=0$, it degenerates into the previous post’s $1+\tau^2$. The windows for net gain when adding an anchor $[\rho(\sqrt{m}-1), \rho(\sqrt{m}+1)]$, the valley $1-\rho^2$, and the covariance upper bound $\rho \leq 1/\sqrt{R}$ are all closed-form algebraic conclusions, not relying on any specific domain parameters.

    Partially Verified: “$\rho$ can be solved from two-point injection” yields in simulation 0.195/0.364/0.596 (true values 0.20/0.40/0.60), and $s_0$ is simultaneously solved with errors within 0.04; this is a qualified-order method, but it has not yet run on any real reading dataset. The feasible way of injecting perturbation (how to “quantitatively” add noise to one senior person while they remain unaware) has no precedent in real acupuncture/skin-temperature reading scenarios. Relevant methodology has external support: Walter et al. (PMID 22351623) prove that if the error of the reference standard is correlated with the measured error, the bias direction and magnitude of sensitivity/specificity depend on prevalence and the covariance of two types of errors, which can be positive or negative, so one cannot always assume “positive correlation means overestimation”—this is isomorphic to this paper’s “direction depends on sign, not magnitude”; Rifkin (PMID 7644859) gives the decomposition formula of observed correlation coefficient under correlated errors and points out that when the reference method is poor (correlation with true value < 0.9), error correlation can cause slope and correlation coefficient to significantly deviate from the independent-error case.

    To Be Verified: ① In real reading scenarios, there is not yet a single measured case of the sign of $\rho$. If the senior and ordinary reader read the same curve on the same screen, $\rho$ is likely positive; but if the senior systematically “sees more finely” due to experience (reading tiny changes as significant), while the ordinary reader reads coarsely, their errors may have opposite signs, causing this paper’s conclusion to flip entirely, turning the cost of adding an anchor from “discounted gain” into “precision regression.” ② Feasibility of the two-point injection method: the variance of injected noise must first be independently calibrated; if this level of calibration itself carries 20% error, how much error the solved $\rho$ has is not yet calculated. ③ The upper bound assumption for $R$ readers presumes mutual independence among readers; in reality, readers at the same center often share common-mode effects, so while the upper bound relaxes, the floor rises, and which effect is larger has not been calculated.

    L1 Network Status: Line 202 of CORRESPONDENCE_NETWORK.md, E-060 “Label Premise” edge: KB-292 (conflict): Diagnostic consistency among TCM physicians is "low (low)" in the open population, authors call it "worrying" (PMID 30383398); ... but training can improve (PMID 19538100), scale-quantification is feasible... → "grouping by five-organ labels" itself carries noise — level conflict (unresolved), status conflict. Line 261, E-060 “KB-334 (open): When positive rates of two types of raters differ, consistency coefficient and prevalence must be reported separately.” The anchor-reader error correlation discussed in this paper has not yet entered that network as an independent edge, belonging to “having an edge but missing a layer within the edge,” so evidence level is capped at C by hard rules.

    Evidence Level: C
    References: PMID 22351623, PMID 7644859, PMID 9006315, PMID 25895962, PMID 28105775, PMID 40499027

    Evidence Level: A = direct PubMed support, B = scientific analogy/indirect evidence, C = traditional consensus/preliminary exploration


    Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.