Category: 科学易学

  • 判读者的尺子到底会漂多少:那条「0.167 个标准差就撞线」的临界值,与真人实测差了两三倍

    Evidence level: C

    一个临界限是”大”还是”小”,不能靠直觉,只能拿实测抖动去比。比完之后,原来那句”小到没人当回事”变成了它自己的反证。

    正文

    上一篇(#1393)把结论压在一个数上:主档占比 m=0.75 时,两位判读者只要把尺子整体挪 0.167 个标准差,就能在打分标准差之比 κ_s 恒等于 1.000(尺度完全一致、判合格)的前提下,把两人对主档占比的判断差 δ 推到 5 个百分点——正好撞线。上一篇接着说:「这个错位量小到任何校准训练都不会当回事。」

    这句话是本篇要审的。0.167 没错,错的是”小”这个形容词——它没有分母。 要让”小”有意义,必须回答:真人重复判读时位置实际会漂多少?库里从没量过。

    一、两个”位置”必须拆开

    固定错位:甲读者天然比乙严,零点稳定分开一段距离,同一次判读里是常数。自身漂移:同一人隔段时间再判同一批,零点还会移动,是随机量。0.167 SD 说的是两者之差,而可实测的是第二种。

    在潜变量阈值模型里(受试者真值 SD 记为 1),一名读者的重测信度与他严厉度漂移标准差有直接关系:σ_drift = √(1−ICC);两人彼此独立时,差的标准差 σ_pair = σ_drift×√2。

    二、拿实测 ICC 换算,得到 0.4 左右

    判读文献里 ICC 报得最勤。取两组实测:脊柱肌萎缩 RHS 量表、21 名资深物理治疗师、间隔两周看同一批录像,读者内 ICC 0.922–1.0(PMID 36542615);压力痛阈与时间总和、24 人同会话重复三次,ICC 0.80–0.97(PMID 33434233)。

    实测 ICC σ_drift σ_pair 是 0.167 的几倍
    0.995 0.071 0.100 0.60
    0.980 0.141 0.200 1.20
    0.950 0.224 0.316 1.89
    0.920 0.283 0.400 2.39
    0.900 0.316 0.447 2.68
    0.850 0.387 0.548 3.28

    文献区间(ICC 0.80–0.97)对应 σ_pair 0.245–0.632,是 0.167 的 1.5 到 3.8 倍。反解更直白:要 σ_pair 不超过 0.167,需要 ICC ≥ 0.986,而多数判读任务在 0.80–0.95。

    另一条独立路径交叉验证:教育测量报告评分者严厉度实测散布(以能力 SD 为单位)为 0.27–0.37,随机配对两人之差即 0.38–0.52,同样落在 2.3–3.1 倍区间。两条路径同量级。

    三、”撞线”不是偶发,是常态

    δ 的抽样分布近似正态,标准差 = φ(c)×σ_pair×100(m=0.75 时 φ=0.318)。代入:

    σ_pair δ 的 SE 超过 5pp 的概率(m=0.75) 同(m=0.85)
    0.100 3.2 pp 1.4% 0.3%
    0.167 5.3 pp 31.8% 14.2%
    0.245 7.8 pp 49.5% 26.0%
    0.400 12.7 pp 67.6% 46.6%
    0.548 17.4 pp 76.0% 57.0%

    按文献中心值 σ_pair≈0.44(ICC 约 0.90),m=0.75 时撞线概率约 67%,δ 的期望绝对值约 11 pp。这不是”会不会撞线”,而是”绝大多数时候本来就在线外”。上一篇那句「κ_s 合格在实践中够用」必须撤回:κ_s 只保证尺子一样长,不保证在同一高度;而尺子不在同一高度是常态。

    四、这个量还测不准,且加病例救不了

    漂移的估计精度只由重复次数决定。个体漂移用 r 次重复估计时自由度是 r−1,相对标准误约 1/√(2(r−1)):3 次 50%、5 次 35%、10 次 24%、20 次 16%。要压到 ±20% 以内,得把多名读者池化(3 次×7 人 或 5 次×4 人)。是靠加读者,不是加病例——这与 #1403 的教训是同一件事的两面:位置参数的自由度来自读者数。

    五、最要紧的一层:判据线正好压在噪声底上

    线性近似下,δ 到 5pp 所需错位 b≈5/(100·φ)。而 δ̂ 的标准误是 SE=100·φ·σ_pair。把 σ_pair 代成 b,SE 精确等于 5 pp。各 m 验证:m=0.60 时 5.10、0.75 时 5.31、0.85 时 5.72、0.90 时 6.38——都在 5 附近。

    判据线恰好落在 δ̂ 的噪声底上,信噪比约 1。 这不是巧合,是一条恒等式:判定线的位置和测量它所需精度的门槛由同一个 φ·b* 决定。凡是”δ 危险到能撞线”的研究,恰恰是”δ 测不准”的同一批。

    由此得到与样本量无关的下限。σ_pair=0.167 时(已是”刚好撞线”配置),n 从 60 加到 3840:合计 SE 依次 9.52 / 7.71 / 6.62 / 6.00 / 5.66 / 5.40 pp,n 趋于无穷也停在 5.31 pp。漂移是系统性的、不可平均——N 个受试者共享同一对读者,两个零点偏移产生的效果同向、不随 n 稀释。这是”增加样本量”这条最常用补救在这里完全失效的原因。若要 SE 压到 1.79 pp(80% 功效检出 5pp),需 σ_pair≤0.040,即 ICC≥0.9984,判读研究达不到。

    六、改法:报告项从 δ 换成 σ_Δ

    既然 δ 的绝对值在现实信度下测不准,判据就不该建在 δ 上。改成直接报漂移:σ_Δ 作为主报告项,δ 只作派生量。成本极低——池化估到 ±20% 只需 3 次×7 名读者,比任何一次一致性研究都小。

    σ_Δ 区间 δ̂ 的 SE 该做什么
    <0.040 <1.8 pp δ 可测、5pp 门槛有效
    0.040–0.118 1.8–5.3 pp 门槛可判,须报 CI
    0.118–0.167 5.3–7.5 pp 门槛失效,只报漂移
    >0.167 >7.5 pp κ_s 合格这个结论不承载信息

    一个设计坑:想用”同一批病例多时点重复”分离漂移与病例难度,理论可行(难度是常数、漂移是时点效应),但一旦难度与时点有交互——判读者记住上次答案、或连续判读疲劳——两者就不分离,退化成 #1406 的记忆效应问题。这条路线必须同时报重复间隔。

    证据与局限

    已验证:漂移与 ICC 的换算 σ_drift=√(1−ICC)、σ_pair=σ_drift·√2 是潜变量模型的直接推论;文献实测 ICC(0.80–0.97,PMID 33434233;0.922–1.0,PMID 36542615)与严厉度散布(0.31–0.42 logits / 能力 SD 1.14)两条独立路径互校后同量级;漂移估计的自由度只来自重复次数(r−1);SE(σ_pair=b*)≈5pp 的恒等式在 m=0.60–0.90 全部成立(近似误差 +0.10 至 +1.38pp)。

    部分验证:文献 ICC 来自量表类判读(物理功能评分、痛阈),与”气功外气变化幅度”这类综合印象判断的相似度只是类比——后者通常信度更低、漂移更大,但无直接数据。另有教育测量场景”115 名评分者仅 1 人跨轮次显著变化”的实测,反解出的漂移/SE 比约 0.65,与 ICC 路径的 0.8 同量级,但属单一场景、n 小,只能作旁证。

    待验证:气功外气判读任务的 ICC 从未测过,所有换算都是跨领域外推,须先做一次最小规模实测(≥20 名判读者 × ≥60 段材料 × 各 3 次)才能定标;漂移是否有共同模会显著改变结论(ρ=0.9 时撞线率从 70% 降到 9%,各漂各的则前述全成立),该相关系数无实测;漂移是随时间单调累积还是随机游走,决定”定期重校准”是否有效(随机游走步长 0.05 SD 时约 11 轮漂回 0.167,步长 0.15 SD 时一轮半就回来)。

    L1 网络状态:L1 边表中无「判读者一致性 / 位置漂移 / 信度」对应边。检索仅命中 CORRESPONDENCE_NETWORK.md 第 192 行「| 生克一致性 | 相生/相克方向能否复现 | 未测 |」(指五行生克方向,与本主题无关),以及 EDGE_TABLE 中「六十甲子 | 13 篇 | A 级 2 篇」的干支纪日可靠性条目(亦无关)。该主题尚未进入 L1 对应网络。

    相关阅读

    证据等级: C
    参考文献: PMID 36542615, PMID 33434233, PMID 36284065, PMID 37283589, PMID 8501467, PMID 35670017

    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。

    How far the reader’s ruler actually floats: the critical threshold of \”0.167 standard deviations hitting the line\” differs from real human measurements by two to three times

    Evidence level: C

    A borderline threshold is “large” or “small” not by intuition, but only by comparing it to the measured jitter. After comparison, the original claim that “it is so small nobody cares” becomes its own counter-evidence.

    Body

    The previous post (#1393) pressed the conclusion onto a single number: when the main archive proportion $m=0.75$, two judges can shift their ruler by 0.167 standard deviations and still hit the line, provided the scoring standard deviation ratio $\kappa_s$ remains exactly 1.000 (scales perfectly aligned, deemed qualified). The previous post then stated: “This misalignment is so small that any calibration training would not take it seriously.”

    This sentence is what this post must examine. 0.167 is correct; the error lies in the adjective “small”—it lacks a denominator. To make “small” meaningful, one must answer: how much does the position actually drift when real people repeat the judgment? This has never been measured in the literature.

    I. Two “positions” must be separated

    Fixed offset: Reader A is naturally stricter than Reader B, with a stable zero-point separated by a fixed distance; within the same judgment session this is a constant. Self-drift: The same person, after some time, re-judges the same batch; the zero-point shifts again, representing a random quantity. The 0.167 SD refers to the difference between these two, whereas what can be measured is the second type.

    In latent-variable threshold models (where the true subject SD is denoted as 1), a judge’s test-retest reliability is directly related to the standard deviation of their severity drift: $\sigma_{\text{drift}} = \sqrt{1-\text{ICC}}$; when two judges are independent, the standard deviation of their difference is $\sigma_{\text{pair}} = \sigma_{\text{drift}} \times \sqrt{2}$.

    II. Converting via measured ICC yields approximately 0.4

    Judgment literature reports ICC most frequently. Take two sets of measurements: spinal muscular atrophy RHS scale, 21 senior physical therapists, viewing the same video clips with a two-week interval, yielding intra-judge ICC 0.922–1.0 (PMID 36542615); pressure pain threshold and time sum, 24 people repeating three times in the same session, ICC 0.80–0.97 (PMID 33434233).

    Measured ICC $\sigma_{\text{drift}}$ $\sigma_{\text{pair}}$ Multiple of 0.167
    0.995 0.071 0.100 0.60
    0.980 0.141 0.200 1.20
    0.950 0.224 0.316 1.89
    0.920 0.283 0.400 2.39
    0.900 0.316 0.447 2.68
    0.850 0.387 0.548 3.28

    The literature interval (ICC 0.80–0.97) corresponds to $\sigma_{\text{pair}}$ 0.245–0.632, which is 1.5 to 3.8 times 0.167. Inverting this is even clearer: to keep $\sigma_{\text{pair}}$ below 0.167, one needs ICC ≥ 0.986, whereas most judgment tasks fall between 0.80–0.95.

    An independent cross-validation path: educational measurement reports the measured spread of rater severity (in units of ability SD) as 0.27–0.37; the difference between randomly paired judges is 0.38–0.52, also falling within the 2.3–3.1 times interval. Both paths are of the same order of magnitude.

    III. “Hitting the line” is not accidental; it is the norm

    The sampling distribution of $\delta$ approximates normality, with standard error $= \phi(c) \times \sigma_{\text{pair}} \times 100$ (when $m=0.75$, $\phi=0.318$). Substituting:

    $\sigma_{\text{pair}}$ SE of $\delta$ Probability exceeding 5pp ($m=0.75$) Same for $m=0.85$
    0.100 3.2 pp 1.4% 0.3%
    0.167 5.3 pp 31.8% 14.2%
    0.245 7.8 pp 49.5% 26.0%
    0.400 12.7 pp 67.6% 46.6%
    0.548 17.4 pp 76.0% 57.0%

    Using the literature central value $\sigma_{\text{pair}} \approx 0.44$ (ICC around 0.90), the probability of hitting the line when $m=0.75$ is approximately 67%, and the expected absolute value of $\delta$ is about 11 pp. This is not “whether it will hit the line,” but rather “in most cases it already lies outside the line.” The previous claim that “$\kappa_s$ being qualified is practically sufficient” must be withdrawn: $\kappa_s$ only may support the rulers are the same length, not that they are at the same height; and rulers not being at the same height is the norm.

    IV. This quantity is also hard to measure precisely, and adding cases cannot save it

    The precision of drift estimation depends solely on the number of repetitions. When estimating individual drift using $r$ repetitions, the degrees of freedom are $r-1$, and the relative standard error is approximately $1/\sqrt{2(r-1)}$: 3 repetitions yield 50%, 5 repetitions yield 35%, 10 repetitions yield 24%, and 20 repetitions yield 16%. To press this within ±20%, one must pool multiple readers (3 times × 7 people or 5 times × 4 people). It relies on adding readers, not adding cases—this mirrors the lesson of #1403: the degrees of freedom for the location parameter come from the number of readers.

    V. The most critical layer: the criterion line sits exactly on the noise floor

    Under linear approximation, the misalignment $b^$ required for $\delta$ to reach 5pp is approximately $5/(100 \cdot \phi)$. Meanwhile, the standard error of $\hat{\delta}$ is $SE = 100 \cdot \phi \cdot \sigma_{\text{pair}}$. Substituting $\sigma_{\text{pair}}$ with $b^$, the SE precisely equals 5 pp. Verification across various $m$: $m=0.60$ yields 5.10, $m=0.75$ yields 5.31, $m=0.85$ yields 5.72, $m=0.90$ yields 6.38—all hovering near 5.

    The criterion line falls exactly on the noise floor of $\hat{\delta}$, with a signal-to-noise ratio of about 1. This is not a coincidence; it is an identity: the position of the decision line and the precision threshold required to measure it are both determined by the same $\phi \cdot b^*$. Any study where “$\delta$ is dangerous enough to hit the line” is precisely the same set of studies where “$\delta$ cannot be measured accurately.”

    This yields a lower bound independent of sample size. When $\sigma_{\text{pair}}=0.167$ (already the “just-hitting-the-line” configuration), increasing $n$ from 60 to 3840 results in SEs of 9.52 / 7.71 / 6.62 / 6.00 / 5.66 / 5.40 pp respectively, stopping at 5.31 pp even as $n$ approaches infinity. Drift is systematic and cannot be averaged out—$N$ subjects share the same pair of judges, so the two zero-point offsets act in the same direction and do not dilute with $N$. This is why “increasing sample size,” the most common remedy, completely fails here. To press SE down to 1.79 pp (80% power to detect 5pp), one needs $\sigma_{\text{pair}} \leq 0.040$, i.e., ICC ≥ 0.9984, which judgment studies cannot achieve.

    VI. The fix: replace reporting $\delta$ with reporting $\sigma_\Delta$

    Since the absolute value of $\delta$ cannot be accurately measured under realistic reliability conditions, the criterion should not be built on $\delta$. Instead, report drift directly: let $\sigma_\Delta$ be the primary reported quantity, with $\delta$ serving only as a derived measure. The cost is extremely low—pooling estimation to ±20% requires only 3 times × 7 readers, far smaller than any single consistency study.

    $\sigma_\Delta$ interval SE of $\hat{\delta}$ What to do
    <0.040 <1.8 pp $\delta$ is measurable; the 5pp threshold is valid
    0.040–0.118 1.8–5.3 pp Threshold is assessable; confidence intervals must be reported
    0.118–0.167 5.3–7.5 pp Threshold fails; report only drift
    >0.167 >7.5 pp The conclusion “$\kappa_s$ is qualified” carries no information

    One design pitfall: theoretically, repeating the same cases at multiple time points could separate drift from case difficulty (since difficulty is constant while drift is a time-point effect). However, once there is interaction between difficulty and time point—judges remembering previous answers or experiencing fatigue during consecutive judgments—the two cannot be separated, degenerating into the memory effect problem of #1406. This route must simultaneously report the repetition interval.

    Evidence and Limitations

    Verified: The conversion $\sigma_{\text{drift}}=\sqrt{1-\text{ICC}}$, $\sigma_{\text{pair}}=\sigma_{\text{drift}} \cdot \sqrt{2}$ is a direct consequence of latent-variable models; literature-measured ICC (0.80–0.97, PMID 33434233; 0.922–1.0, PMID 36542615) and severity spread (0.31–0.42 logits / ability SD 1.14) from two independent paths cross-validate to the same order; degrees of freedom for drift estimation come solely from repetition count ($r-1$); the identity $SE(\sigma_{\text{pair}}=b^*) \approx 5\text{pp}$ holds for all $m=0.60–0.90$ (approximation error +0.10 to +1.38 pp).

    Partially verified: Literature ICC comes from scale-based judgments (physical function scores, pain thresholds), whereas similarity to “Qi Gong external Qi change amplitude” such holistic impression judgments is only an analogy—the latter typically has lower reliability and greater drift, but no direct data. Additionally, the educational measurement scenario “115 raters with only 1 showing significant change across rounds” yields an inverted drift/SE ratio of about 0.65, comparable to the ICC-path’s 0.8, but this is a single scenario with small $n$, serving only as corroborating evidence.

    To be verified: ICC for Qi Gong external Qi judgment tasks has never been measured; all conversions are cross-domain extrapolations requiring a minimum-scale real measurement (≥20 judges × ≥60 materials × 3 repetitions each) before calibration; whether drift shares a common mode significantly alters conclusions (when $\rho=0.9$, the hitting-line rate drops from 70% to 9%; if each drifts independently, all prior results hold); this correlation coefficient lacks measurement; whether drift accumulates monotonically over time or performs random walk determines if “periodic re-calibration” is effective (with step size 0.05 SD, about 11 cycles return to 0.167; with step size 0.15 SD, one and a half cycles suffice).

    L1 Network Status: No corresponding edge exists in the L1 edge table for “judge consistency / position drift / reliability.” Search only hits CORRESPONDENCE_NETWORK.md line 192 “| Generating/Overcoming Consistency | Whether generating/overcoming directions can be reproduced | Not measured |” (referring to Five Elements generating/overcoming directions, unrelated to this topic), and the EDGE_TABLE entry “Sixty Heavenly Stems | 13 Chapters | A-level 2 chapters” regarding the reliability of Gan-Zhi date marking (also unrelated). This topic has not yet entered the L1 correspondence network.

    Related Reading

    Evidence Level: C
    References: PMID 36542615, PMID 33434233, PMID 36284065, PMID 37283589, PMID 8501467, PMID 35670017

    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.

  • 天花板本来就是一个减法:0.0078 不是「几乎够用」,是那个参数族压根没让另一个数动起来

    Evidence level: C

    判读一致性的天花板 κmax 只由两个数决定——总变差距离 TVD 和期望一致率 p_e,且是精确的减法关系。上一轮报的「残差只有 0.0078」不是代理量够用,是那个参数族根本没能让 p_e 动起来。

    正文

    前一篇(#1387)把一致性系数的最大可达值 κmax 的残余自由度算成 0.0015–0.0078,结论是「报 (m̄, TVD) 这个二元组差不多够用,残差只占 0.05 判据带宽的 3%–16%」。这个结论方向没错,但理由是错的,而且临界区间刚好落在最不该出错的地方。

    先看一个代数事实。三档判读里若两位判读者的边际分布是 p 与 q,则最大一致率等于 1 减去总变差距离,即 p_o^max = 1 − TVD,TVD = max|Δpᵢ|(二十万组随机三类边际上实测偏差 2.2×10⁻¹⁶,是恒等式不是近似)。把它代入 κmax 的定义,得到一个不含任何参数化假设的闭式:

    κmax = (1 − TVD − p_e) / (1 − p_e),其中 p_e = Σpᵢqᵢ

    三十万次抽样下该式与逐表计算值的最大偏差为 0。也就是说,天花板不是一个「受很多因素影响的量」,它是一个二元函数:给它 TVD 和 p_e,它就没有任何自由度了。

    那 0.0078 是怎么来的? 出在那个参数族上。上一篇的族写作 A 的第二档占比 0.5+ε、B 的第三档占比 0.5−ε,此时 p_e = m_A m_B + 2s_A s_B(0.25 − ε²)。注意后一项带着小量 s_A s_B——在 m_A=0.92、m_B=0.88 处 s_A s_B = 0.0096,于是 ε 从 0 扫到 0.5 整段,p_e 只动了 0.0048。κmax 对 p_e 的偏导是 −TVD/(1−p_e)²,量级 0.58,所以那 0.0048 的 p_e 摆幅只翻成 0.0028 的 κmax 摆幅。0.0078 是「这个族把 p_e 钉住了」的读数,不是「二元组充分」的证据。 换一个族,同一个 (m̄,TVD) 下残余立刻放大。

    放大到什么程度? 约束 m̄ 与 TVD 精确固定、让 δ(主档占比差)与次档方向自由扫遍全可行域:m̄=0.84 时 κmax 残余跨度 0.016;m̄=0.90 时 0.021;m̄=0.94 时 0.034;m̄=0.98 时 0.043。穿越 0.05 带宽大约在 m̄≈0.96。也就是说这个二元组在低 m̄ 区确实够用,在高 m̄ 区不够——而高 m̄ 恰恰是天花板压得最低、最需要精确知道它有多低的场合。

    补位参数应当是 p_e,不是 m̄。 把报告项从 (m̄, TVD) 换成 (TVD, p_e),残余跨度在八十六万点网格上恒为 0。m̄ 之所以看起来能用,是因为它与 p_e 在真实数据里高度共变(网格相关 0.9737)——这是样本内的巧合,不是恒等式。两个判读者可以 m̄ 相同而 p_e 不同(次档分布一偏一直),此时按 m̄ 查表就错。

    真正的麻烦在有限样本。 κmax 对 TVD 的偏导是 −1/(1−p_e):p_e=0.815 时放大 5.4 倍,0.90 时 10 倍,0.95 时 20 倍,0.99 时 100 倍。而 TVD̂ 是有偏的——它是三个绝对差取最大,E[max] ≥ max[E],n=71 时实测高估 +0.0186。两个误差不抵消,同向叠加:E[κ̂max] − κmax = −0.105,95% 置信区间宽度 0.70,是 0.05 判据带宽的 14 倍。要把这个区间压到半个带宽以内,m̄=0.75 需要 n≈3840,0.90 需要 7680,0.95 需要 15360;而一致性研究的样本量中位数只有 71(PMID 36123642)。

    这还意味着 δ 本身没那么重要。 前几篇一直把主档占比差 δ 当成核心量,但在这个恒等式里 δ 只是 TVD 的一个下界成分(δ ≤ TVD 恒成立:δ=TVD 当两个次档同向偏移,δ<TVD 当次档反向)。固定 m̄ 与 TVD 后让 δ 从 0 扫到 TVD,κmax 的摆动只有 0.0004–0.0178——在 m̄=0.90、TVD=0.10 处,δ=0 与 δ=0.10 两个极端给出的 κmax 分别是 0.4737 与 0.4721,差 0.0016。所以「δ 大不大」这件事对天花板几乎无影响,「TVD 大不大」才是全部。 前几篇把注意力压在 δ 上,等于盯住了一个分量而漏掉了那个真正的自变量;这也解释了为什么 #1386 会得出「尺度差对 δ 的杠杆是阈值平移的十倍」——那句话对 δ 成立,但对 κmax 不承重。

    顺带推翻前一篇的一个担心:最大值切换(两条边际差并列最大时 TVD̂ 的方差被低估)并没有真的发生。稳定最大与并列最大两种构造下,n=120 起覆盖率分别为 94.3% 与 92.6%,n=960 为 92.4% 与 92.8%,都在名义 95% 附近。偏倚是单向的、且主要是小样本效应,不是双峰分布。坏消息不是「方差被低估」,而是「误差进 κmax 时被乘了 5 到 100 倍」。

    还有一个反直觉的发现:κmax 本身并不比 κ 更难估。同一份数据上,两者的 95% 区间宽度之比在 0.65–1.19 之间,量级相当;而当两位判读者边际完全相同时(TVD=0),κmax 恒等于 1,天花板整个消失。所以问题不在「多报一个量」,而在「报了这个量却不能用于它本来要服务的那个比较」——κ/κmax 这个比值在 n=71 时的 95% 区间宽达 0.57,完全没有判别力。

    实用结论有三条:一,κmax 应当按 (TVD, p_e) 报,不要报 m̄ 代理;二,任何跨研究的天花板比较必须先报 n,因为 n=71 时这个量根本不可比;三,若研究想回答的是「一致率够不够」,直接检视 TVD 与 p_e 比检视 κmax 更稳——前者是输入,后者是把误差放大后的输出。

    证据与局限

    已验证:κmax = (1 − TVD − p_e)/(1 − p_e) 是恒等式(三十万次抽样偏差 0);p_o^max = 1 − TVD(二十万组偏差 2.2×10⁻¹⁶);(TVD, p_e) 是精确充分统计量(残余跨度恒为 0);TVD̂ 的正偏倚(n=71 时 +0.0186)与 κ̂max 的负偏倚(−0.105)。

    部分验证:残余跨度随 m̄ 上升的曲线、穿越 0.05 的位置(m̄≈0.96)来自 0.002 步长的网格扫描,是数值结论而非解析式;所需样本量(3840/7680/15360)由 4000 次模拟的区间半宽反解,依赖所假设的边际构型。

    待验证:本组从未实测过真实三档裁决数据的 p_e 与 TVD 联合分布,所有数值扫描都建立在参数化边际上;κmax 的置信区间用的是 bootstrap 近似,未与解析标准误(PMID 41230191 给出的方法)逐点对照;「直接报 (TVD, p_e)」这条建议在审稿实践中的可接受度未知。

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 2348207, PMID 8621993, PMID 25168681, PMID 36123642, PMID 15733050, PMID 26095449, PMID 24338600, PMID 41230191

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 三域节点表仅登记单点信号(物理侧「相位 / 节律 | — | 周期性耦合的关键」;生理侧「气感报告 | 结构化自评」「体质分型 | 量表」),全表检索「判读 / 信度 / 一致率 / 边际 / 偏倚 / 带宽 / 充分性 / 样本量」全部零命中;边表止于 E-050(纳音↔频率,L1、open,现代证据栏原文「尚无任何可测来源」);网络对齐指标行原文「L2 以上边数 | 0」;EDGE_TABLE_science-yixue.md 47 个概念无「判读一致性 / kappa / 观察者间信度」任何条目。故本篇证据等级封顶 C。


    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 ceiling itself is a subtraction: 0.0078 is not \”almost enough,\” it’s that the parameter family never moved the other number.

    Evidence level: C

    The ceiling of consistency κmax is determined solely by two numbers—the Total Variation Distance TVD and the expected agreement rate p_e—and follows an exact subtraction relationship. The previously reported “residual of only 0.0078” does not indicate that the proxy quantity is sufficient; rather, it indicates that the parameter family itself failed to move p_e.

    Body

    The previous article (#1387) calculated the residual degrees of freedom for the maximum attainable value of the consistency coefficient κmax as 0.0015–0.0078, concluding that “the pair (m̄, TVD) is roughly sufficient, with the residual occupying only 3%–16% of the 0.05 criterion bandwidth.” This conclusion is directionally correct but based on flawed reasoning, and the critical interval falls precisely where errors are most unacceptable.

    First, consider an algebraic fact. If two raters have marginal distributions p and q across three categories, the maximum agreement rate equals one minus the Total Variation Distance: p_o^max = 1 − TVD, where TVD = max|Δpᵢ| (measured across 200,000 random three-category marginals, the deviation is 2.2×10⁻¹⁶; this is an identity, not an approximation). Substituting this into the definition of κmax yields a closed-form expression containing no parameterization assumptions:

    κmax = (1 − TVD − p_e) / (1 − p_e), where p_e = Σpᵢqᵢ

    Across 300,000 simulations, the maximum deviation between this formula and table-by-table calculations is exactly 0. Thus, the ceiling is not “a quantity affected by many factors”; it is a binary function: given TVD and p_e, it has no degrees of freedom.

    So where does 0.0078 come from? It arises from the parameter family itself. In the previous article’s family, Category A has a second-tier proportion of 0.5+ε and Category B has a third-tier proportion of 0.5−ε, yielding p_e = m_A m_B + 2s_A s_B(0.25 − ε²). Note the second term contains the small quantity s_A s_B; at m_A=0.92 and m_B=0.88, s_A s_B = 0.0096. As ε sweeps from 0 to 0.5, p_e changes only by 0.0048. Since the partial derivative of κmax with respect to p_e is −TVD/(1−p_e)² (magnitude 0.58), this 0.0048 swing in p_e translates into only a 0.0028 swing in κmax. The value 0.0078 is the reading that “this family pinned p_e,” not evidence that “the pair is sufficient.” Under a different family, the residual immediately expands even with the same (m̄, TVD).

    How large does it expand? When m̄ and TVD are fixed precisely while δ (the difference in primary category proportions) and the direction of secondary categories freely sweep across the feasible region: at m̄=0.84, the κmax residual spans 0.016; at m̄=0.90, it spans 0.021; at m̄=0.94, it spans 0.034; at m̄=0.98, it spans 0.043. Crossing the 0.05 bandwidth occurs around m̄≈0.96. Thus, this binary pair is indeed sufficient in the low m̄ region but insufficient in the high m̄ region—precisely where the ceiling is most compressed and knowing its exact magnitude is most critical.

    The missing parameter should be p_e, not m̄. Rewriting the reported item from (m̄, TVD) to (TVD, p_e) yields a residual span of exactly 0 across an 860,000-point grid. m̄ appears usable only because it is highly co-varying with p_e in real data (grid correlation 0.9737)—a coincidence within the sample, not an identity. Two raters can share the same m̄ yet have different p_e (if secondary distributions diverge in one direction and converge in another); consulting tables by m̄ then leads to error.

    The real difficulty lies in finite samples. The partial derivative of κmax with respect to TVD is −1/(1−p_e): at p_e=0.815 it amplifies 5.4-fold, at p_e=0.90 it amplifies 10-fold, at p_e=0.95 it amplifies 20-fold, and at p_e=0.99 it amplifies 100-fold. Meanwhile, TVD̂ is biased upward—it equals the maximum of three absolute differences, and E[max] ≥ max[E]; with n=71, the measured overestimation is +0.0186. These two errors do not cancel; they compound in the same direction: E[κ̂max] − κmax = −0.105, with a 95% confidence interval width of 0.70, which is 14 times the 0.05 criterion bandwidth. To compress this interval to half the bandwidth, m̄=0.75 requires n≈3840, m̄=0.90 requires n=7680, and m̄=0.95 requires n=15360; yet the median sample size in consistency studies is only 71 (PMID 36123642).

    This also implies δ itself is not so important. Previous articles have treated the primary category difference δ as central, but in this identity δ is merely a lower-bound component of TVD (δ ≤ TVD always holds: δ=TVD when secondary categories shift in the same direction, δ<TVD when they shift oppositely). Fixing m̄ and TVD while letting δ sweep from 0 to TVD, the κmax fluctuation ranges only from 0.0004 to 0.0178—at m̄=0.90 and TVD=0.10, the extremes δ=0 and δ=0.10 yield κmax values of 0.4737 and 0.4721, differing by just 0.0016. Thus, whether δ is large or small has almost no impact on the ceiling; what matters entirely is TVD. Previous articles focused on δ, effectively盯着 one component while missing the true independent variable; this also explains why #1386 concluded “the leverage of scale difference on δ is ten times the threshold shift”—that statement holds for δ but does not bear weight on κmax.

    Incidentally, this overturns a concern from the previous article: maximum switching (when two marginal differences tie for maximum, causing TVD̂ variance to be underestimated) does not actually occur. Under both stable maximum and tied-maximum constructions, coverage rates for n≥120 are 94.3% and 92.6%, respectively, and for n=960 are 92.4% and 92.8%, all near the nominal 95%. Bias is unidirectional and primarily due to small-sample effects, not a bimodal distribution. The bad news is not “variance is underestimated,” but rather “error entering κmax gets multiplied by 5 to 100 times.”

    Another counterintuitive finding: κmax itself is no harder to estimate than κ. On the same dataset, the ratio of their 95% interval widths ranges from 0.65 to 1.19, of comparable magnitude; and when two raters’ marginals are identical (TVD=0), κmax is constantly 1, causing the ceiling to vanish entirely. Thus, the problem is not “reporting an extra quantity,” but rather “reporting this quantity yet failing to use it for the comparison it was meant to serve”—the ratio κ/κmax has a 95% interval width of 0.57 at n=71, rendering it completely non-discriminatory.

    Practical conclusions are threefold: first, κmax should be reported as (TVD, p_e), not via m̄ as a proxy; second, any cross-study comparison of ceilings must first report n, because at n=71 this quantity is incomparable; third, if a study aims to answer “is the agreement rate sufficient,” directly examining TVD and p_e is more stable than examining κmax—the former is the input, while the latter amplifies error.

    Evidence and Limitations

    Verified: κmax = (1 − TVD − p_e)/(1 − p_e) is an identity (300,000 simulations show zero deviation); p_o^max = 1 − TVD (200,000 deviations measured at 2.2×10⁻¹⁶); (TVD, p_e) is an exact sufficient statistic (residual span always zero); TVD̂’s positive bias (+0.0186 at n=71) and κ̂max’s negative bias (−0.105).

    Partially verified: The curve of residual span increasing with m̄ and the crossing point at 0.05 (m̄≈0.96) stem from a 0.002-step grid scan, representing numerical rather than analytical conclusions; required sample sizes (3840/7680/15360) were derived by inverting the half-width of 4,000 simulations, depending on assumed marginal configurations.

    To be verified: This group has never empirically measured the joint distribution of p_e and TVD in real three-category adjudication data; all numerical sweeps rely on parameterized marginals; κmax’s confidence intervals use bootstrap approximation, not yet pointwise compared against the analytical standard error method (PMID 41230191); the feasibility of “directly reporting (TVD, p_e)” in review practice remains unknown.

    Related Reading

    Evidence Level: C
    References: PMID 8501467, PMID 2348207, PMID 8621993, PMID 25168681, PMID 36123642, PMID 15733050, PMID 26095449, PMID 24338600, PMID 41230191

    L1 Network Status: This topic has not yet entered the L1 corresponding network. The CORRESPONDENCE_NETWORK.md three-domain node table registers only single-point signals (physical side “phase/rhythm | — | periodic coupling key”; physiological side “qi-sensation report | structured self-assessment”, “constitution typology | scale”), with full-table retrieval of “adjudication / reliability / agreement rate / marginal / bias / bandwidth / sufficiency / sample size” yielding zero hits; edge table stops at E-050 (Na Yin ↔ Frequency, L1, open, original text of modern evidence column states “no measurable source yet”); network alignment indicator row reads “L2 and above edges | 0”; EDGE_TABLE_science-yixue.md has 47 concepts but no entries for “adjudication consistency / kappa / inter-rater reliability.” Hence, this article’s evidence level is capped at 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.

  • ε 到底是这人的脾气还是他手抖:一个做不出来的分解,和一个做得出的替身

    Evidence level: C

    一个量测不出来,不代表它的后果算不出来。上一轮想把 ε 拆成「稳定的个人风格」与「同一人的试次摇摆」,并打算用「谁占六成」来决定死区判据要不要用。三刀下去,这个分解不但做不出来,而且拆开之后的方向跟原计划完全相反。

    正文

    上一篇(kappa-marginal-tvd-deadband-closed-form)把整条一致率天花板的推导收成一个式子:ε* = δ/(4(1−m̄))。m̄ 是主档占比,δ 是两位读者主档占比的差。ε 一旦越过 ε*,总变差距离 TVD 就开始超过 δ,天花板被压得更低。但 ε 从头到尾没人实测过——上一篇只能写「若 ε≤0.05 则安全」,说不出真实研究里这个数长什么样。

    于是有了一个看起来非常自然的计划:找一批 ≥120 例、≥2 位读者、每人重复判读 ≥2 次的三档数据,把 ε 的方差拆成两层——「人×稳定」(这人一贯偏乐观)和「人×病例×试次」(同一批片子看两遍自己就不一致)。按原设想:若稳定成分 >60%,死区判据可用;若随机成分 >60%,死区警告可降级为误报。

    第一刀:这个分解的自由度只有读者数减一。

    稳定成分的方差 σ²_a 是靠「各读者的长期均值之间有多大差异」估出来的。k 位读者,就只有 k 个均值,df = k−1。这个自由度跟你收多少病例、每人判多少次,完全无关。k=2 时 df=1,σ²_a 的相对标准差是 141%,w_s(稳定成分占比)的 90% 置信区间半宽 ±52.9 个百分点——这个区间宽到能把任何一个真值同时判成「稳定主导」和「随机主导」。要到 ±10pp 需要 k≥30 位读者,±5pp 需要 k≥114 位。

    气功外气这类研究,判读者通常就是两三位老师。在这个配置下,w_s 是一个原理上估不准的数,不是「样本量不够再补点」。

    文献侧也印证了这一点。Kreiter 与 Wilson(PMID 26925540)明确指出:在嵌套设计下,系统性的评分者效应与「评分者×被评者」交互效应会被合并成一个混杂的方差成分,无法分开估计;要分开只能靠完全交叉模型。这不是统计技巧问题,是数据结构本身不给这个信息。而 Svensson 与 Holm(PMID 7701145)早在 1994 年就给出了一套把有序分类配对变异分为系统(由边际决定)与随机(相对秩方差)的方法——方法存在,但对两位读者的情形,系统部分的估计精度同样被 df 卡死。Sturman 等(PMID 28587597)追踪考官严厉度发现,即便没有任何反馈干预,严厉度在几个月内也会明显漂移;他们自己都说「这种固有的不稳定性使考官训练策略的研究变得复杂」。也就是说,「稳定的个人风格」这个前提本身就未必成立——它可能只是一个移动得足够慢、被误当成常数的变量。

    第二刀:就算测出来了,判决方向是反的。

    这是本轮最要紧的发现。原计划的推理是:随机成分期望为零、不系统性地把 TVD 推高,所以随机占主导时死区可以不管。

    但这个推理漏了一个符号问题:TVD 对 ε 是偶函数。三档下 TVD = max{|p₁−q₁|, |p₂−q₂|, |p₃−p₃|},而我们的参数化让次档差恰好正比于 |ε_A − ε_B|。也就是说,TVD 只吃两位读者偏离的差值绝对值,完全不吃这个差值的正负号。一位偏乐观、一位偏保守(符号相反,差值最大)会推高 TVD;两位都偏乐观、但程度不同(符号相同,差值非零)同样推高 TVD。所以「期望为零」在这里一点保护作用都没有。

    量化一下:同方差 s 下,稳定成分(两点分布 ±s)的 E|Δ| = s;随机成分(高斯)的 Δ 服从 N(0, 2s²),E|Δ| = √(4/π)·s ≈ 1.128s。随机成分的抬升能力反倒高出 12.84%。

    更要紧的是硬上界:如果两位读者的稳定风格都落在 |ε| ≤ ε* 区间内,那么 |Δ| ≤ 2ε*,TVD 恒等于 δ,永不越线。在 m̄=0.75/0.85/0.90 三档上把 |ε|≤ε* 的全网格扫了一遍,max(TVD) 精确等于 δ,误差在 10⁻¹² 量级。而随机成分服从连续分布,尾部永远能越过任意阈值:m̄=0.85、s=ε*=8.33% 时,纯稳定 |ε|=ε* 的越线率是 0.00%,纯随机 sd=ε* 的越线率是 16.50%(5 万次模拟)。

    结论反过来了:稳定成分是安全的,随机成分才是危险的。原计划「随机占主导 ⇒ 死区警告降级为误报」这一条,方向错了。

    第三刀:w_s 是尺度无关的,而天花板负担不是。

    这可能是最根本的漏洞。w_s = σ²_a/(σ²_a+σ²_r) 是个比值,两组数据可以 w_s 完全相同而总方差差一个量级。实测(m̄=0.85,每组 2 万次):

    σ²_a σ²_r w_s E[TVD] − δ
    0.65 0.35 0.65 +11.44pp
    0.13 0.07 0.65 +4.02pp
    0.0325 0.0175 0.65 +1.21pp

    w_s 一模一样,天花板负担差了 9.5 倍。用 w_s 做判决,等于用一个无量纲比值去管一个有量纲的后果——量纲就不对。

    还有一个自相矛盾:原计划说「重复判读越多,随机成分被平均掉,稳定成分占比越高,死区判据越该收紧」。但重复判读恰恰是降负担的手段。r 从 1 增到 10,w_s 从 35.0% 升到 84.3%(按 σ²_a=0.35、σ²_r=0.65 算),可同一过程中纯随机的越线率从 16.34% 掉到 0.00%。也就是说,「w_s 升高」这件事本身正是「风险下降」的副产品,把它读成「该更警惕」,逻辑反了。

    替身:不拆,直接测那个差值。

    既然 TVD 只吃 |Δ|=|ε_A−ε_B|,那要判的本来就是一个数。判据直接写:|Δ| ≤ 2ε* = δ/(2(1−m̄))。m̄=0.85 时这条线是 16.67%,m̄=0.75 时是 10.00%,m̄=0.90 时是 25.00%。

    这是标准的非劣性检验,所需病例数(单侧 α=.05,次档基准比例 p=0.25):

    m̄ 判据线 2ε* n(r=1) n(r=2) n(r=3)
    0.70 8.33% 147 74 49
    0.75 10.00% 102 51 34
    0.80 12.50% 65 33 22
    0.85 16.67% 37 19 13
    0.90 25.00% 17 9 6

    对照一下:分解 w_s 的路子要 ≥40 位读者才有 80% 功效,而直接测 |Δ| 只需要两位读者、几十到一百多例。同一笔研究预算,一个做得出来,一个做不出来。

    证据与局限

    已验证:在三档对称参数族 m_A=m̄±δ/2 下,TVD = max(三条边际差绝对值),且次档差正比于 |ε_A−ε_B|——这是纯代数,闭式在 ε=ε* 处精确给出 TVD=δ(三档 m̄ 逐点核验,误差 <10⁻¹²)。|ε|≤ε* 全网格上 max(TVD)=δ,稳定成分的硬上界成立。同方差下随机成分 E|Δ| 比稳定成分高 12.84%(√(4/π)),5 万次模拟越线率 16.50% vs 0.00%。w_s 相同而负担差 9.5 倍,实测三组一致。所需样本量按非劣性公式反解,为解析值。

    部分验证:σ²_a 的 df=k−1 与相应的置信区间半宽是标准方差成分理论(正态近似),未做小样本精确分布的校正;w_s 的区间宽度用 delta 法一阶近似,在 k 很小时偏差会更大。文献侧只印证了「系统成分与交互成分在嵌套数据下不可分」「考官严厉度本身会漂移」这两大类结论,没有任何文献做过本文这个具体的「稳定 vs 随机对 TVD 抬升能力」的比较。重复判读只压随机成分这一设定是理想化的——实际重复判读可能引入记忆效应,把第二次判读变成第一次的复制,此时 r 的实际收益低于 √r。

    待验证:真实三档裁决数据里 |Δ| 的实测分布;稳定成分是否真的「稳定」(Sturman 等的结果暗示可能不是,需要 ≥4 个时间点的追踪才能区分慢漂移与常数);记忆效应使重复判读收益折损多少;m̄ 与 δ 本身带抽样误差时 2ε* 这条线会怎么移动(#1405 专线)。

    证据等级: C
    参考文献: PMID 26925540, PMID 7701145, PMID 28587597, PMID 31169540

    L1 网络状态:CORRESPONDENCE_NETWORK.md 边表止于 E-050(纳音↔频率,L1 | open,现代证据栏原文「尚无任何可测来源」);网络对齐指标行原文「L2 以上边数 | 0」。三域节点表无「判读 / 评分者 / 方差成分 / 一致率」任何登记位(全表检索「判读 / 裁决 / 一致率 / 边际 / 偏倚 / 方差 / 噪声 / 测量」零命中),物理侧「相位 / 节律」语义为「两振荡锁相」而非边际分布比较。故本文证据等级封顶 C,且该主题尚未进入 L1 对应网络(已登记至 l1-gaps)。

    相关阅读


    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 ε the person’s temper or his shaky hand: an impossible decomposition and a possible stand-in

    Evidence level: C

    Measuring something does not imply its consequences cannot be calculated. In the previous iteration, we attempted to decompose ε into “stable personal style” and “trial-to-trial fluctuation of the same person,” intending to use “who accounts for 60%” to decide whether to apply the deadband criterion. After three cuts, this decomposition not only failed but reversed the direction of the split compared to the original plan.

    Main Text

    The previous article (kappa-marginal-tvd-deadband-closed-form) condensed the derivation of the consistency ceiling into a single formula: ε = δ/(4(1−m̄)). Here, m̄ is the proportion of the main category, and δ is the difference in main-category proportions between two readers. Once ε exceeds ε, the Total Variation Distance (TVD) surpasses δ, and the ceiling is pushed lower. However, ε has never been empirically measured; the previous article could only state “if ε ≤ 0.05, it is safe,” without describing what this number looks like in real research.

    Thus, a seemingly natural plan emerged: collect three-category data from ≥120 cases, ≥2 readers, with each case rated ≥2 times, and decompose the variance of ε into two layers—”person × stable” (this person is consistently optimistic) and “person × case × trial” (the same set of images yields inconsistent results when viewed twice). According to the original hypothesis: if the stable component accounts for >60%, the deadband criterion applies; if the random component accounts for >60%, the deadband warning can be downgraded to a false alarm.

    Cut 1: The degrees of freedom for this decomposition equal the number of readers minus one.

    The variance of the stable component, σ²_a, is estimated by “how large the differences are between the long-term means of each reader.” With k readers, there are only k means, so df = k−1. This degree of freedom is completely independent of how many cases or repetitions per person are collected. When k=2, df=1, and the relative standard deviation of σ²_a is 141%; the 90% confidence interval half-width for w_s (the proportion of the stable component) is ±52.9 percentage points—so wide that any true value could be classified as either “stable-dominated” or “random-dominated.” To achieve ±10pp precision, k≥30 readers are needed; to reach ±5pp, k≥114 readers.

    Researches on phenomena like Qi Gong external energy typically involve only two or three raters. Under this configuration, w_s is a number that is fundamentally impossible to estimate accurately, not merely “insufficient sample size.”

    Literature corroborates this point. Kreiter and Wilson (PMID 26925540) explicitly state: under nested designs, systematic rater effects and “rater × subject” interaction effects are merged into a single confounded variance component, making them inseparable; separation is only possible via a fully crossed model. This is not a statistical technique issue but a structural limitation of the data itself. Svensson and Holm (PMID 7701145) provided a method in 1994 to decompose ordered categorical paired variation into systematic (determined by marginal totals) and random (relative rank variance) components—though the method exists, for two readers, the estimation precision of the systematic part is still capped by df. Sturman et al. (PMID 28587597) tracked examiner severity and found that even without feedback intervention, severity drifted noticeably over months; they themselves noted “this inherent instability makes the study of examiner training strategies complex.” In other words, the premise of “stable personal style” may not hold—it might simply be a variable drifting slowly enough to be mistaken for a constant.

    Cut 2: Even if measured, the judgment direction is reversed.

    This is the most critical finding of this round. The original reasoning assumed that the random component has zero expectation and does not systematically inflate TVD, so when randomness dominates, the deadband can be ignored.

    This reasoning overlooks a sign issue: TVD is an even function of ε. Under three categories, TVD = max{|p₁−q₁|, |p₂−q₂|, |p₃−p₃|}, while our parameterization makes the subcategory difference exactly proportional to |ε_A − ε_B|. Thus, TVD depends only on the absolute difference of the two readers’ deviations, not the sign. One reader optimistic and one conservative (opposite signs, maximum difference) inflates TVD; both optimistic but to different degrees (same sign, non-zero difference) also inflate TVD. Therefore, “zero expectation” offers no protection here.

    Quantitatively: under equal variance s, the stable component (two-point distribution ±s) has E|Δ| = s; the random component (Gaussian) has Δ ~ N(0, 2s²), so E|Δ| = √(4/π)·s ≈ 1.128s. The random component’s inflation capacity is actually 12.84% higher.

    More critically, the hard upper bound: if both readers’ stable styles fall within |ε| ≤ ε, then |Δ| ≤ 2ε, and TVD always equals δ, never crossing the line. Scanning the full grid where |ε|≤ε for m̄=0.75/0.85/0.90 yields max(TVD) exactly equal to δ, with error on the order of 10⁻¹². Meanwhile, the random component follows a continuous distribution, so its tails always exceed any threshold: when m̄=0.85 and s=ε=8.33%, the crossing rate for pure stable |ε|=ε is 0.00%, while for pure random sd=ε it is 16.50% (from 50,000 simulations).

    The conclusion is reversed: the stable component is safe; the random component is dangerous. The original plan’s claim that “random dominance ⇒ deadband warning downgraded to false alarm” is directionally wrong.

    Cut 3: w_s is scale-invariant, but the ceiling burden is not.

    This may be the most fundamental flaw. w_s = σ²_a/(σ²_a+σ²_r) is a ratio, so two datasets can have identical w_s yet differ in total variance by an order of magnitude. Empirical tests (m̄=0.85, 20,000 per group):

    σ²_a σ²_r w_s E[TVD] − δ
    0.65 0.35 0.65 +11.44pp
    0.13 0.07 0.65 +4.02pp
    0.0325 0.0175 0.65 +1.21pp

    w_s is identical, yet the ceiling burden differs by 9.5 times. Using w_s for judgment means applying a dimensionless ratio to a dimensional consequence—dimensionally inconsistent.

    Furthermore, there is an internal contradiction: the original plan claimed “more repetitions average out random components, increasing w_s, so the deadband criterion should tighten.” But more repetitions actually reduce the burden. As r increases from 1 to 10, w_s rises from 35.0% to 84.3% (calculated with σ²_a=0.35, σ²_r=0.65), yet the pure random crossing rate drops from 16.34% to 0.00%. Thus, “w_s increasing” is itself a byproduct of “risk decreasing”; interpreting it as “should be more cautious” reverses the logic.

    Proxy: Do not decompose; directly measure that difference.

    Since TVD only consumes |Δ|=|ε_A−ε_B|, what needs judgment is essentially one number. The criterion should directly state: |Δ| ≤ 2ε* = δ/(2(1−m̄)). For m̄=0.85, this line is 16.67%; for m̄=0.75, it is 10.00%; for m̄=0.90, it is 25.00%.

    This is a standard non-inferiority test, requiring sample size (one-sided α=.05, subcategory baseline proportion p=0.25):

    m̄ Criterion line 2ε* n (r=1) n (r=2) n (r=3)
    0.70 8.33% 147 74 49
    0.75 10.00% 102 51 34
    0.80 12.50% 65 33 22
    0.85 16.67% 37 19 13
    0.90 25.00% 17 9 6

    Comparison: decomposing w_s requires ≥40 readers to achieve 80% power, whereas directly measuring |Δ| needs only two readers and dozens to over a hundred cases. With the same research budget, one approach is feasible; the other is not.

    Evidence and Limitations

    Verified: Under three-category symmetric parameter families m_A=m̄±δ/2, TVD = max(three marginal differences absolute values), and the subcategory difference is proportional to |ε_A−ε_B|—this is pure algebra, with the closed-form solution exactly giving TVD=δ at ε=ε (verified pointwise across all three m̄ values, error <10⁻¹²). On the full grid where |ε|≤ε, max(TVD)=δ, confirming the hard upper bound for the stable component. Under equal variance, the random component’s E|Δ| exceeds the stable component by 12.84% (√(4/π)); simulation of 50,000 runs yields crossing rates of 16.50% vs 0.00%. When w_s is identical, the burden differs by 9.5 times; empirical tests confirm all three groups consistently. Required sample sizes were solved analytically via the non-inferiority formula.

    Partially Verified: The degrees of freedom df=k−1 and corresponding confidence interval half-width for σ²_a follow standard variance component theory (normal approximation), without correction for small-sample exact distributions; w_s’s interval width uses the delta method’s first-order approximation, which incurs greater bias when k is very small. Literature only confirms the two general conclusions that “systematic and interaction components are inseparable under nested data” and “examiner severity itself drifts”; no study has performed the specific comparison of “stable vs. random components’ ability to inflate TVD.” The assumption that repeated ratings only suppress the random component is idealized—actual repetition may introduce memory effects, turning the second rating into a copy of the first, so the actual benefit of r is less than √r.

    To Be Verified: The empirical distribution of |Δ| in real three-category adjudication data; whether the stable component is truly “stable” (Sturman et al.’s results suggest it may not be, requiring tracking across ≥4 time points to distinguish slow drift from constant); how much memory effects reduce the benefit of repeated ratings; how the line 2ε* shifts when m̄ and δ themselves carry sampling error (see #1405 dedicated thread).

    Evidence Level: C
    References: PMID 26925540, PMID 7701145, PMID 28587597, PMID 31169540

    L1 Network Status: The CORRESPONDENCE_NETWORK.md edge table ends at E-050 (Na Yin ↔ Frequency, L1 | open); the original text of the modern evidence column states “no measurable source yet“; the network alignment indicator row states “0 edges beyond L2.” The three-domain node table contains no registration slots for “rating / rater / variance component / consistency” (full-table search for “rating / adjudication / consistency / marginal / bias / variance / noise / measurement” yields zero hits); the physical side’s “phase / rhythm” semantics refer to “two oscillators phase-locked,” not marginal distribution comparison. Therefore, this article’s evidence level is capped at C, and the topic has not yet entered the corresponding L1 network (registered to l1-gaps).

    Further Reading


    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.

  • 2.772 这个系数没错,错的是你以为它需要修正:相依折价就是样本量虚报

    Evidence level: C

    上一篇担心「同一批受试者上的两个 kappa 不独立,2.772 会偏保守」。#1388 说这个修正能把门槛砍掉一半。实际算下来:修正量确实存在,但它不是一个新的统计量,而是「你的 n 被虚报了 1/(1−ρ) 倍」这件旧事的另一种写法——而真正高 ρ 的场合,恰好是真差为零、最不该给折扣的场合。

    正文

    一、先把「相依折价」翻译成人话

    1385 给出可辨差下限:

    D_min = B + 1.96√2 · SE

    其中 1.96√2 = 2.772 来自「两条独立 95% 区间不重叠」。若两个 kappa 算在同一批受试者上、共享同一份抽样误差,则差值标准误应为 1.96√(2−2ρ)·SE。#1388 据此提出一条折扣表,并预告「ρ=0.5 时门槛降 41%」。

    那条预告是错的,而且不止一处。但更值得先说清的是:整件事根本不需要新公式。

    二、这个修正是一个恒等式,不是新发现

    把 SE = σ/√n 代进去:

    1.96√(2−2ρ)·σ/√n = 1.96√2 · σ/√n_eff,其中 n_eff = n/(1−ρ)

    我用 5 组 ρ(0.15 / 0.30 / 0.50 / 0.70 / 0.85)逐点验算,两种算法的 D_min 差最大 5.55×10⁻¹⁷,其余四组差恰好 0。

    结论:所谓「相依折价」在代数上完全等价于「你以为自己有 60 人,其实只有 60/(1−ρ) 人的独立信息」。 ρ=0.5 时 60 人只值 120 人的价——注意方向,n_eff 比 n 大,因为相依让两个 kappa 差得更近、差值的方差更小。

    这一条改写把问题从「要不要修系数」变成「你的样本量报得虚不虚」,后者可以查、可以报、可以被审稿人一句话问穿。而前者只会变成一个没人核的修正因子。

    顺带把 #1388 的算术也纠了:ρ=0.5 时 √(2−2ρ)/√2 = √0.5 = 0.707,SE 项降 29.3% 而不是 41%;再算上不折扣的 B 项,D_min 实际只降 28.0%(m=0.90、n=200 时从 0.337 到 0.243)。「近半要改判」这个说法需要 ρ>0.70 才站得住。

    三、ρ 到底有多大:仿真给出的答案和原来的假设相反

    1388 的理由是「共享同一批受试者的边际与个体难判程度,抽样误差高度正相关」。这个理由听着顺,但没测过。我用潜变量阈值模型做了分层仿真(三档、m=0.60/0.90、n=60/240、判读噪声 σ_e 0.35–0.75、读者风格偏移 σ_τ 0/0.25/0.5,每组 600 次重复):

    共享结构 实测 ρ 范围
    不同受试者 + 不同读者对(真独立) −0.079 ~ 0.123
    同批受试者 + 换一组读者对 0.015 ~ 0.308
    同批受试者 + 同一读者对 + 换场合 0.073 ~ 0.974

    第二行是 #1388 描述的情形,ρ 中位数 0.111,上限 0.308,36 个格子里有 5 个实测为负值、10 个低于 0.05。也就是说:共享受试者几乎不产生相关——共享的只是「这批人难不难判」这一个共同成分,而它在 kappa 的总方差里占比很小。

    真正把 ρ 推高的是共享读者阈值(同一对人马重复判读):σ_τ=0.25 时 ρ 到 0.5–0.86,σ_τ=0.5 时到 0.90+。ρ 的主来源是「读者风格固定」,不是「受试者同一批」。 #1388 把归因放错了地方,于是「ρ 可能大于 0.5」这个预判只在一种场景下成立:同一组判读者相隔一段时间再判一遍。

    四、最危险的一格:折扣在最该保守的地方最狠

    ρ 高意味着两个 kappa 高度同步——而它们最同步的时刻,是真实差值本来为零的时刻。上面第三行那个场景(同一读者对、换场合)真差恒为 0,我却实测到:

    配置 实测 ρ D_min(不折扣) D_min(折扣后) 判「可辨」的假阳性率
    m=0.90, n=60, σ_τ=0 0.256 0.273 0.238 2.10% → 4.40%
    m=0.90, n=60, σ_τ=0.5 0.793 0.477 0.225 0.00% → 3.83%
    m=0.60, n=60, σ_τ=0.5 0.908 0.607 0.205 0.00% → 2.37%

    不折扣时假阳性率是 0–2.5%,套上 ρ 折扣后升到 2–4.4%。折扣把「同一批人在两个场合被同一组人判了两遍」这种最该判为无可比性的设计,放进了「可辨」的篮子。 这不是数值瑕疵:ρ 高的原因恰恰是两条读数出自同一条流水线,此时「差」这个量本身就没承载新信息。

    所以折扣不能无条件套。它的适用前提是「两个 kappa 的比较本身有意义」,而 ρ→1 时这个前提自我取消(差恒为 0,任何门槛都失去对象)。ρ 应当同时被当作「折扣因子」和「可比性红牌」读。

    五、折不掉的那部分:m 的惩罚

    1388 最想解决的其实是这个:#1385 发现 m 越大越没资格比(m=0.90 时 D_min 0.337,m=0.60 时只有 0.143,差 2.36 倍)。折扣能补多少?

    要让 m=0.90 的 D_min 降到 m=0.60 的水平,需要 ρ ≥ 0.84–0.86(n=200 时 0.842、n=60 时 0.859)。这个值只在「同一读者对 + 强风格偏移」的窄带里达到,而在最常见的「换一组读者对」场景(ρ≈0.11 中位数)下,降幅只有 5.4%。

    即使把 ρ 拉到 0.5 并且真的想守住 0.05 那条线,n 仍需约 1084 人——ρ=0.5 只是把原来的 16,963 减到 1084,量级没变。

    m 的惩罚不是相依性造的,所以相依性折不掉它。 这一点值得单独讲:m 大时 Σp²→1,分母 (1−Σp²)² 把 SE 炸开,这是 kappa 这个系数的定义域问题,不是抽样设计问题。想靠「反正我两次判读是同一批人」把它绕过去,方向就错了。

    六、ρ 怎么估:三种结构三种估法

    这一条比折扣表本身更有用。ρ 不能靠假设,得按共享结构分三种估法:

    • 换读者对、同受试者:ρ̂ 可以把同一批受试者随机劈成两半,两半各自算 kappa,用两半的相关系数当 ρ̂(半样本法),再按 Spearman-Brown 校正,虽然这一步本身又会引入新误差,但量级上够用。
    • 同读者对、换场合:直接拿两个场合的 kappa 算 Pearson 相关即可,这是最干净的一种,但要注意它有偏——两个场合共享的不只是读者风格,还有「这批片子本来就难」这个共同成分。
    • 完全独立样本:ρ 应为 0,报 0 即可,但若实测出显著的 ρ,说明两组样本有重叠或来自同一中心,那是数据管理问题,不是统计问题。

    这三种估法给出的 ρ̂ 量级差半个数量级(换读者对中位数 0.11 vs 同读者对中位数 0.50),混用会让折扣表失去意义。所以报告规范该要求的不只是「报 ρ」,而是「报 ρ 且报它是在哪种共享结构下估的」。

    七、该改的是什么

    不是 2.772,是报告口径。任何跨样本比两个 kappa 的研究,把「n=60」换成「n_eff = 60/(1−ρ̂)」写出来,并同报三样:

    • ρ̂ 的估计方式(同一读者对重复判读 / 换读者对 / 独立样本)
    • 若 ρ̂>0.5,同时声明「两次判读共享读者与受试者,本比较的可比性本身存疑」
    • B、SE、D_min 三个数照旧(#1385 口径),D_min 用 √(2−2ρ̂)

    证据与局限

    已验证:ρ 折扣与 n_eff 折算的代数等价(5 组 ρ 逐点验算,最大偏差 5.55×10⁻¹⁷);ρ 的三层结构(独立 / 换读者对 / 同读者对)在 27 组参数网格上的分层结果;折扣后假阳性率升高(每组 3000 次重复);#1388 的 41% 降幅应为 29.3%(算术复核)。

    部分验证:ρ 对 σ_τ 的单调依赖来自潜变量正态模型,真实判读者阈值分布未必正态,也未含「学习效应」(第二遍判得更一致,会把 ρ 推到比本文更高);三档设定固定,五档有序分类下 ρ 的绝对数值会变,但「共享读者主导、共享受试者边缘」这个定性结构应当保持。

    待验证:① 真实三档裁决研究中 ρ 的实测分布从未被报过,本文给的是模型值不是实测值;② 本文的 D_min 沿用 #1385 的 B 闭式,而 B 本身在边际不等(#1384)与 TVD 偏离 δ(#1387)下要重算,ρ 与这两条修正的交互未算;③ 「ρ>0.5 即可比性存疑」这条红线的阈值是拍的,没有依据。

    相关阅读

    证据等级: C
    参考文献: PMID 18430244, PMID 25321842, PMID 24488927, PMID 36123642
    L1 网络状态:CORRESPONDENCE_NETWORK.md 三域节点表仅登记单点信号(物理侧「相位 / 节律 | — | 周期性耦合的关键」,语义为两振荡锁相;生理侧「呼吸率 / 呼吸模式 | 可穿戴」「气感报告 | 结构化自评」),全表检索「判读 / 信度 / 一致率 / 样本量 / 相依 / 有效样本量」全部零命中;边表止于 E-050(纳音↔频率,L1、open,现代证据栏原文「尚无任何可测来源」);网络对齐指标行原文「L2 以上边数 | 0」。该主题尚未进入 L1 对应网络。


    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。

    2.772 is correct; the error lies in assuming it requires adjustment: dependency discount equals inflated sample size.

    Evidence level: C

    In the previous post, concern was raised that two independent kappas calculated on the same batch of subjects would not be truly independent, potentially making 2.772 overly conservative. #1388 suggested this correction could halve the threshold. Actual calculation reveals: while a correction factor does exist, it is not a new statistic but merely another way of stating the old fact that “your reported n is inflated by 1/(1−ρ) times.” Moreover, in cases of truly high ρ, the true difference is zero—the exact scenario where discounting should be avoided.

    Main Body

    I. First, translate “dependency discounting” into plain English

    1385 provides the lower bound for the discernible difference:

    D_min = B + 1.96√2 · SE

    where 1.96√2 = 2.772 originates from “two independent 95% intervals not overlapping.” If two kappas are calculated on the same batch of subjects, sharing the same sampling error, then the standard error of the difference should be 1.96√(2−2ρ)·SE. #1388 accordingly proposes a discount table and previews that “when ρ=0.5, the threshold drops by 41%.”

    This preview is incorrect, and not just in one place. But more importantly, what needs to be clarified first is: the entire matter fundamentally does not require a new formula.

    II. This correction is an identity, not a discovery

    Substituting SE = σ/√n into the equation:

    1.96√(2−2ρ)·σ/√n = 1.96√2 · σ/√n_eff, where n_eff = n/(1−ρ)

    I verified this point-by-point using five sets of ρ (0.15 / 0.30 / 0.50 / 0.70 / 0.85). The maximum difference in D_min between the two methods is 5.55×10⁻¹⁷; for the other four sets, the difference is exactly zero.

    Conclusion: so-called “dependency discounting” is algebraically completely equivalent to “you thought you had 60 people, but actually only have 60/(1−ρ) independent pieces of information.” When ρ=0.5, 60 people are worth the price of 120—note the direction: n_eff is larger than n because dependency makes the two kappa differences closer together, reducing the variance of the difference.

    This reformulation shifts the problem from “whether to correct the coefficient” to “whether your reported sample size is inflated.” The latter can be checked, reported, and exposed by a single sentence from reviewers. The former merely becomes an unverified correction factor.

    Incidentally, #1388’s arithmetic is also corrected: when ρ=0.5, √(2−2ρ)/√2 = √0.5 = 0.707, meaning the SE term drops by 29.3%, not 41%; including the undiscounted B term, D_min actually drops only by 28.0% (from 0.337 to 0.243 when m=0.90, n=200). The claim “nearly half must be reclassified” holds only when ρ>0.70.

    III. How large is ρ? Simulation yields the opposite of the original assumption

    1388’s reasoning is that “sharing the same batch of subjects makes marginal and individual difficulty hard to distinguish, leading to high positive correlation in sampling error.” This sounds plausible but was never tested. Using a latent variable threshold model, I performed stratified simulations (three tiers: m=0.60/0.90, n=60/240, reading noise σ_e 0.35–0.75, reader style shift σ_τ 0/0.25/0.5, with 600 repetitions per group):

    Shared Structure Measured ρ Range
    Different subjects + different reader pairs (truly independent) −0.079 ~ 0.123
    Same batch of subjects + different reader pairs 0.015 ~ 0.308
    Same batch of subjects + same reader pair + different occasions 0.073 ~ 0.974

    The second row corresponds to #1388’s described scenario, where the median ρ is 0.111, the upper limit is 0.308, and among 36 grids, 5 had negative measured values and 10 were below 0.05. This means: sharing subjects almost does not produce correlation—the shared component is merely “how hard this batch of people is to judge,” which accounts for a very small portion of the total variance in kappa.

    What truly drives ρ upward is shared reader thresholds (the same pair of readers repeatedly judging): when σ_τ=0.25, ρ reaches 0.5–0.86; when σ_τ=0.5, it exceeds 0.90. The main source of ρ is “fixed reader style,” not “same batch of subjects.” #1388 misattributed the cause, so the prediction “ρ may exceed 0.5” holds only in one scenario: when the same group of readers judges again after some time interval.

    IV. The most dangerous case: the discount is strongest exactly where conservatism is needed

    High ρ means the two kappas are highly synchronized—and they are most synchronized precisely when the true difference is zero. In the third-row scenario (same reader pair, different occasions), the true difference remains constant at zero, yet I measured:

    Configuration Measured ρ D_min (undiscounted) D_min (discounted) False positive rate for “discernible”
    m=0.90, n=60, σ_τ=0 0.256 0.273 0.238 2.10% → 4.40%
    m=0.90, n=60, σ_τ=0.5 0.793 0.477 0.225 0.00% → 3.83%
    m=0.60, n=60, σ_τ=0.5 0.908 0.607 0.205 0.00% → 2.37%

    Without discounting, the false positive rate is 0–2.5%; applying the ρ discount raises it to 2–4.4%. The discount puts designs where “the same batch of people was judged twice by the same readers”—which should be deemed incomparable—into the “discernible” basket. This is not a numerical flaw: the reason ρ is high is precisely that both readings come from the same pipeline, at which point the quantity “difference” itself carries no new information.

    Therefore, the discount cannot be applied unconditionally. Its applicability presupposes “the comparison between two kappas itself is meaningful,” which self-cancels when ρ→1 (since the difference is constantly zero, any threshold loses its target). ρ should be read simultaneously as a “discount factor” and a “comparability red flag.”

    V. The part that cannot be discounted: the penalty of m

    1388 most wanted to address this: #1385 found that larger m makes comparison less qualified (when m=0.90, D_min is 0.337; when m=0.60, it is only 0.143—a difference of 2.36 times). How much can the discount compensate?

    To bring the D_min of m=0.90 down to the level of m=0.60, ρ must be ≥ 0.84–0.86 (0.842 when n=200, 0.859 when n=60). This value is achieved only in the narrow band of “same reader pair + strong style shift,” whereas under the most common scenario of “different reader pairs” (median ρ≈0.11), the drop is only 5.4%.

    Even if ρ is pushed to 0.5 and we truly want to hold the 0.05 line, n still requires approximately 1084 people—ρ=0.5 merely reduces the original 16,963 to 1084, leaving the order of magnitude unchanged.

    The penalty of m is not caused by dependency, so dependency cannot discount it. This point deserves separate emphasis: when m is large, Σp²→1, causing the denominator (1−Σp²)² to explode the SE; this is a domain issue of the kappa coefficient itself, not a sampling design problem. Trying to bypass it by “since both readings are on the same batch of people” goes in the wrong direction.

    VI. How to estimate ρ: three structures, three estimation methods

    This section is more useful than the discount table itself. ρ cannot rely on assumptions; it must be estimated according to shared structure via three methods:

    • Different reader pairs, same subjects: ρ̂ can randomly split the same batch of subjects into two halves, calculate kappa separately for each half, and use the correlation coefficient between the halves as ρ̂ (half-sample method), then apply Spearman-Brown correction. Although this step itself introduces new error, the order of magnitude is sufficient.
    • Same reader pair, different occasions: directly compute the Pearson correlation between the two occasions’ kappas. This is the cleanest method, but note it is biased—what is shared between the two occasions is not only reader style but also “this batch of images was inherently difficult,” a common component.
    • Completely independent samples: ρ should be 0; report 0 directly. If a significant ρ is measured, it indicates overlap between the two samples or that they originate from the same center, which is a data management issue, not a statistical one.

    The three estimation methods yield ρ̂ differing by half an order of magnitude (median 0.11 for different reader pairs vs. median 0.50 for same reader pairs). Mixing them renders the discount table meaningless. Therefore, reporting standards should require not just “report ρ,” but “report ρ and specify under which shared structure it was estimated.”

    VII. What should be changed

    Not 2.772, but the reporting convention. Any cross-sample comparison of two kappas should replace “n=60” with “n_eff = 60/(1−ρ̂)” and report three items simultaneously:

    • The estimation method for ρ̂ (same reader pair repeated judgment / different reader pairs / independent samples)
    • If ρ̂>0.5, also declare “the two judgments share both readers and subjects; the comparability of this comparison itself is questionable”
    • B, SE, D_min remain as per #1385’s convention, with D_min using √(2−2ρ̂)

    Evidence and Limitations

    Verified: algebraic equivalence between ρ discounting and n_eff conversion (verified point-by-point across five ρ sets, maximum deviation 5.55×10⁻¹⁷); stratified results of ρ’s three structures (independent / different reader pairs / same reader pair) across 27 parameter grids; increased false positive rate after discounting (3000 repetitions per group); #1388’s 41% drop should be 29.3% (arithmetic verification).

    Partially verified: ρ’s monotonic dependence on σ_τ stems from the latent variable normal model; true reader threshold distributions may not be normal, and “learning effects” (second judgment more consistent, pushing ρ higher than reported here) are not included; the three-tier setting is fixed, so under five-tier ordered classification the absolute value of ρ will change, but the qualitative structure “reader sharing dominates, subject sharing is marginal” should remain.

    Pending verification: ① the empirical distribution of ρ in real three-tier adjudication studies has never been reported; this paper provides model values, not empirical ones; ② this paper’s D_min uses #1385’s closed-form B, which itself must be recalculated under marginal inequality (#1384) and TVD deviation δ (#1387), and the interaction between ρ and these two corrections is not computed; ③ the threshold “ρ>0.5 implies questionable comparability” is arbitrarily set, lacking basis.

    Further Reading

    Evidence Level: C
    References: PMID 18430244, PMID 25321842, PMID 24488927, PMID 36123642
    L1 Network Status: CORRESPONDENCE_NETWORK.md three-domain node table registers only single-point signals (physical side “phase / rhythm | — | periodic coupling key”, semantically two oscillators phase-locked; physiological side “breathing rate / breathing pattern | wearable”, “qi sensation report | structured self-assessment”), full table search for “judgment / reliability / consistency rate / sample size / dependency / effective sample size” yields zero hits; edge table stops at E-050 (Na Yin ↔ Frequency, L1, open, original text of modern evidence column states “no measurable source yet”); network alignment indicator row original text “L2 and above edges | 0“. This topic has not yet entered the corresponding L1 network.


    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.

  • 两条修正不是相加也不是相乘:一条能用样本量买回来,另一条买不回来

    Evidence level: C

    边际偏倚 δ 和相依 ρ 作用在可辨差 D_min 的两个不同项上,交叉项实测中位只占主效应 0.054%。但真正要紧的不是「可不可加」,而是 δ 那一项里根本没有样本量、ρ 那一项与样本量成反比——前者买不回来。

    正文

    上一篇(#1388)用 √(2−2ρ) 给可比性门槛打了折,但全程假定带宽 B 不变;而更早的 #1384、#1387 已经证明边际一有偏倚、天花板就从 1 掉下来。两条修正凑在一起,合成式该写成:

    D_min = C(δ) + 1.96√(2−2ρ)·SE(δ, ρ, n),其中 C(δ) = δ/(1−p_e)

    于是留下一个从未检查的结构问题:交叉项有多大?「只改 δ」「只改 ρ」「两个都改」三种算法,差的是不是就是两者的和?

    先给答案的一半。把四种组合直接代数展开,交叉项不多不少等于

    cross = [K(ρ) − K(0)]·[SE(δ) − SE(0)]

    两个小量的乘积。为什么是两个小量?因为 p_e 有精确闭式 p_e = m̄² + (1−m̄)²/2 − (3/8)δ²(全网格实测最大偏差 1.1×10⁻¹⁶)——δ 只以平方的形式进入 p_e。m̄=0.90、δ=5pp 时 p_e 只降 0.00094,相对 1−p_e=0.185 是 0.5%,SE 因此只动 0.25%。乘上 ΔK 的最大值 1.53,交叉项落在 10⁻⁴ 量级。45 格网格实测:占主效应中位 0.054%、最大 1.00%(最极端那格是 m̄=0.98、δ=0.10、ρ=0.8)。

    按判据「交叉项 <10% 即可独立套用」,这一支成立。

    为什么会这么小?因为它决定了这个结论能推多远。δ 要影响 SE,必须先穿过 p_e 这道门,而 p_e 对 δ 的敏感度是 −(3/4)δ——δ 越小,敏感度越小,是二次关系。典型的 δ 在 1–5pp,此时敏感度只有 0.0075–0.0375,比 δ 本身还小一个量级。也就是说,边际偏倚在「天花板」这条路上是全额的,在「噪声」这条路上被平方压成二级效应。两条路一个走一次、一个走两次,交叉项自然只剩残渣。反过来读:交叉项可忽略不是因为两条修正弱耦合,而是因为其中一条本身就弱。

    但这个 0.054% 是「对称参数族」送的礼物。 换成真实研究里常见的非对称——一位读者主档 0.93、另一位 0.83——交叉项立刻放大:不对称度 β∈[0.2, 0.8] 时中位升到 1.78%、最大 6.61%;β∈[0.1, 0.9] 时 2.43% / 8.97%。10% 的判据线已经贴边。上一篇刚因为「参数族把某个量钉死、读数被当成结构性质」栽过一次,同一个坑还在同一个位置。

    真正的答案不在交叉项上。 它在两项的符号与可补偿性上。

    δ 项与 ρ 项方向相反:δ 抬 D_min,ρ 压 D_min。「两个都改」不等于「两个都重要」——m̄=0.90、n=71 时,δ=1pp 的 δ 项贡献 +0.054,而 ρ=0.5 的 ρ 项贡献 −0.151:中等相依直接淹没小偏倚;δ=5pp 时 δ 项 +0.268,反过来淹没相依。分界落在 C(δ) = ΔK·SE(0) 那条线上。

    更要紧的是一条不对称:C(δ) 里没有 n,而 K(ρ)·SE 里 SE ∝ 1/√n。 实测 m̄=0.90:n=1120 时,ρ 从 0 走到 0.9,要维持 D_min=0.30 所需 δ 从 3.16pp 抬到 4.82pp(+1.66pp);n 加到 10⁵,同样的 ρ 跨度只让所需 δ 从 5.33pp 抬到 5.50pp(+0.17pp)——差额缩了十倍。反过来,δ=5pp、ρ=0、n 无穷大时,C(0.05)=0.269,天花板照样停在 0.731。

    所以「同时发生时该按谁算」有一个不需要看具体参数的答案:先看 δ。δ 决定的那部分永远买不回来,ρ 决定的那部分只要肯加样本就能买到零。

    由此引出等值线问题。「D_min 相同即可比性相同」在原理上不成立。m̄=0.90、n=560、D_min 同为 0.30:ρ=0 的研究只需 δ=2.16pp,持久天花板 0.883;ρ=0.9 的研究需要 δ=4.50pp,天花板 0.758。同一个门槛数,两种研究的真实天花板差 0.125,等于 2.5 个 0.05 带宽。判据第三支成立:报 (δ, ρ) 二元组,不报 D_min 单值。

    这个「同数不同病」的后果很具体。假设两份研究都写着 D_min=0.30、都被判「差值 0.32 可比」。ρ=0.9 那份,δ 是 4.5pp,它的报告值 κ 本来就被天花板压在 0.758 以下,测出来的 0.75 已经是「几乎顶格」;ρ=0 那份,δ 只有 2.2pp,天花板 0.883,同样的 0.75 只是「中上」。两篇文章都在讲一个 0.75,但前者说明这套判读方法已到极限、再练也上不去,后者说明还有余量。并排放进一篇综述做「一致性水平比较」,比的是两个含义不同的 0.75。 这不是统计精度问题,是语义问题。

    还有一个反直觉的副产品。让 δ 撑满结构上界 1−m̄,天花板损失收敛到 C_max = 1/(1+m̄−(1−m̄)/8):m̄→1 时 →0.500,m̄=0.50 时 0.696(闭式与数值差 ≤1.2×10⁻¹⁴)。也就是说,无论边际多集中,只要两位读者的边际差撞到上界,天花板一定掉到 0.47–0.50。 这修正了「高 m̄ 区更危险」的说法:高 m̄ 区只是「很小的 δ 就能造成大损失」(放大倍数 (1−p_e)⁻¹ 到 25 倍),而最大可能的损失反而更小(0.50 对 0.696)。两个方向刚好相反:低 m̄ 区损失大但难以触发,高 m̄ 区损失小但一触即发。 判风险要看「触发难度 × 损失幅度」的乘积,而不是只看其中一头——这也解释了为什么前面几篇一会儿说高 m̄ 危险、一会儿说不要比,读起来像自相矛盾:它们各看了一半。

    最后一条能直接写进报告规范的经验法则:令 C(δ*)=0.05,则 δ*/(1−m̄) 从 7.0%(m̄=0.60)单调升到 9.9%(m̄=0.99),几乎是常数。主档占比差一旦超过结构上界的约 10%,单靠这个差就能吃掉整个 0.05 带宽——此后加多少样本都无效。

    这条法则的实际用法只要两个数。第一步,从报告里读两位读者的主档占比,取差得那个 δ;第二步,算 δ 占 (1−m̄) 的百分比。落在 10% 以下,这篇研究还有「加样本能救」的空间;一旦过 10%,无论报告里写了多大的 n、多漂亮的置信区间,它的一致性上限已经被自己的边际差锁死,救不回来。举个具体数字:m̄=0.95 的研究,(1−m̄)=5pp,10% 就是 δ 超过 0.5pp——两位读者只要在主档占比上差半个百分点,天花板就已经吃掉 0.053。 半个百分点是 60 个人里多判一个人进主档,随机误差就能做到。

    所以「两条修正该按谁算」的最终操作口径是这样三步,顺序不能换:① 先算 δ 占比,过 10% 就判定这篇研究的上限已被锁死、进入「不可补偿」区;② 没过,再算 δ 与 ρ 分界带,看当下是 δ 项还是 ρ 项在主导 D_min;③ 若主导的是 ρ,才轮到讨论「要不要加样本把它买回来」。 反过来做——先看 D_min 数字、再倒推两条修正各占多少——会掉进等值线陷阱:同一个数字背后可能是买得回来的病,也可能是买不回来的病。

    证据与局限

    已验证(本文内部可复核的代数与数值):
    – 交叉项恒等式 cross = [K(ρ)−K(0)]·[SE(δ)−SE(0)],代数展开,无近似。
    – p_e = m̄² + (1−m̄)²/2 − (3/8)δ²,8×6 网格实测最大偏差 1.1×10⁻¹⁶。
    – 天花板绝对下界闭式 C_max = 1/(1+m̄−(1−m̄)/8),全域数值差 ≤1.2×10⁻¹⁴;m̄→1 极限 0.500。
    – 「ρ 买得回来、δ 买不回来」:SE ∝ n^(−1/2) 与 C(δ) 无 n,属定义层面的性质,不依赖参数选择。

    部分验证:
    – 交叉项占比的数值(中位 0.054%)只在对称参数族上成立;非对称族升到 1.78–2.43%、最大 8.97%,已贴近 10% 判据线。故「可独立套用」的结论须附「对称或弱非对称」这一前提。
    – 「分界带」的位置依赖 k₀=0.5 这个标称系数值;真实研究报告的 κ 若系统性偏离 0.5,SE 的绝对量级会变,分界带随之移动(方向未变,位置未标定)。

    待验证:
    – 全部结论建立在解析形式上,未与任何一组真实三档判读数据对照。真实数据里 (δ, ρ) 的联合分布未知,因此「常见配置落在分界带哪一侧」无法回答。
    – δ 与 ρ 被当作独立输入,但 #1388 已证 ρ 的主要来源是「共享读者阈值」——而读者阈值差异本身就可能同时制造 δ。若二者同源,本文的可加分解在因果解读上仍成立、在归因上不成立。

    L1 网络状态:该主题尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK.md 三域节点表(易学/中医侧、物理侧、生理/体验侧)无任何以「判读信度 / 一致性系数 / 边际偏倚 / 相依结构 / 误差可补偿性」为源或目标的节点,边表止于 E-050「六十甲子纳音 | 对应 | 60 个频率」(open,现代证据栏原文「尚无任何可测来源」),全文件 grep「判读 / 信度 / 一致率 / 边际 / 偏倚 / 带宽 / 充分性 / 样本量 / 相依 / kappa / 可辨 / 门槛 / 阈值」仅命中第 79 行层级定义与第 159 行「L2 以上边数 | 0」;EDGE_TABLE_science-yixue.md 仅统计 47 个概念的文章计数与最高证据级(频率 407 / HRV 214 / 脉象 73),无「可补偿性 / 交叉项 / 永久偏倚」条目。

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 2348207, PMID 8621993, PMID 25168681, PMID 41230191, PMID 36123642, PMID 24338600, PMID 18430244, PMID 25321842, PMID 24488927, PMID 10501650, PMID 37127220


    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。

    Two corrections are neither additive nor multiplicative: one can be bought back with sample size, the other cannot.

    Evidence level: C

    Marginal bias δ and dependence ρ act on two different terms of the discernible difference D_min. In cross-validation, the median of the cross-term accounts for only 0.054% of the main effect. What truly matters is not whether they are additive, but that the δ term lacks sample size entirely, while the ρ term is inversely proportional to sample size—the former cannot be bought back.

    Main Text

    The previous post (#1388) applied the discount factor √(2−2ρ) to the comparability threshold, but assumed bandwidth B remained constant; earlier posts #1384 and #1387 already demonstrated that any marginal bias causes the ceiling to drop below 1. Combining both corrections yields the composite formula:

    D_min = C(δ) + 1.96√(2−2ρ)·SE(δ, ρ, n), where C(δ) = δ/(1−p_e)

    This leaves an unresolved structural question: how large is the cross-term? Do the three approaches—adjusting only δ, adjusting only ρ, or adjusting both—differ simply by the sum of the two effects?

    First, half the answer. Algebraically expanding all four combinations yields the cross-term exactly as:

    cross = [K(ρ) − K(0)]·[SE(δ) − SE(0)]

    A product of two small quantities. Why two small terms? Because p_e has the exact closed-form expression p_e = m̄² + (1−m̄)²/2 − (3/8)δ² (verified across a full 8×6 grid with maximum deviation 1.1×10⁻¹⁶)—δ enters p_e only quadratically. When m̄=0.90 and δ=5pp, p_e decreases by only 0.00094; relative to 1−p_e=0.185, this is 0.5%, causing SE to shift by only 0.25%. Multiplied by the maximum ΔK of 1.53, the cross-term falls within the 10⁻⁴ order of magnitude. In a 45-cell grid validation, it accounts for median 0.054% and maximum 1.00% of the main effect (the most extreme case occurs at m̄=0.98, δ=0.10, ρ=0.8).

    According to the criterion “cross-term <10% allows independent application,” this branch holds.

    Why is it so small? Because it determines how far this conclusion can be extrapolated. For δ to affect SE, it must pass through the p_e gate, and p_e’s sensitivity to δ is −(3/4)δ—the smaller δ, the lower the sensitivity, following a quadratic relationship. Typical δ values range from 1–5pp, yielding sensitivities of only 0.0075–0.0375, an order of magnitude smaller than δ itself. Thus, marginal bias acts fully along the “ceiling” path but gets squared down to a secondary effect along the “noise” path. One path traverses once, the other twice, leaving the cross-term as mere residue. Reading this in reverse: the cross-term is negligible not because the two corrections are weakly coupled, but because one of them is inherently weak.

    However, this 0.054% is a gift from the “symmetric parameter family.” Switching to asymmetric configurations common in real research—where one reader’s main file proportion is 0.93 and another’s 0.83—the cross-term immediately expands: for asymmetry β∈[0.2, 0.8], the median rises to 1.78% and maximum to 6.61%; for β∈[0.1, 0.9], it becomes 2.43% / 8.97%. The 10% threshold is now borderline. The previous post fell into a pit once because “the parameter family fixed one quantity, causing the reading to be mistaken for a structural property”; the same pit remains in the same location.

    The true answer lies not in the cross-term, but in the signs and compensability of the two terms.

    The δ term and ρ term oppose each other: δ raises D_min, while ρ suppresses it. “Adjusting both” does not mean “both matter”—when m̄=0.90 and n=71, the δ term contributes +0.054 for δ=1pp, whereas the ρ term contributes −0.151 for ρ=0.5: moderate dependence completely swamps small bias; when δ=5pp, the δ term contributes +0.268, swamping dependence in reverse. The dividing line lies at C(δ) = ΔK·SE(0).

    More critically, there is an asymmetry: C(δ) contains no n, while K(ρ)·SE includes SE ∝ 1/√n. In validation with m̄=0.90: when n=1120, moving ρ from 0 to 0.9 requires δ to increase from 3.16pp to 4.82pp (+1.66pp); when n increases to 10⁵, the same ρ span only raises required δ from 5.33pp to 5.50pp (+0.17pp)—the gap shrinks tenfold. Conversely, with δ=5pp, ρ=0, and n→∞, C(0.05)=0.269, yet the ceiling still stops at 0.731.

    Thus, “which term to apply when both occur” has an answer requiring no specific parameters: first examine δ. The part determined by δ can never be bought back; the part determined by ρ can be bought back to zero simply by adding sample size.

    This leads to the isoline problem. “Same D_min implies same comparability” does not hold in principle. With m̄=0.90 and n=560, D_min=0.30 for both cases: when ρ=0, the study requires δ=2.16pp with a persistent ceiling of 0.883; when ρ=0.9, it needs δ=4.50pp with a ceiling of 0.758. The same threshold number masks a real ceiling difference of 0.125, equivalent to 2.5 bandwidth units of 0.05. The third criterion holds: report the (δ, ρ) pair, not the single D_min value.

    The consequence of “same number, different disease” is concrete. Suppose two studies both report D_min=0.30 and are judged “comparable with difference 0.32.” The ρ=0.9 study has δ=4.5pp; its reported κ is already suppressed below the ceiling of 0.758, so the measured 0.75 represents “near-maximum.” The ρ=0 study has δ=2.2pp with ceiling 0.883, so the same 0.75 represents “upper-middle.” Both papers discuss a 0.75, yet the former indicates the interpretation method has reached its limit and cannot improve further, while the latter suggests room remains. Placing them side-by-side in a review for “consistency level comparison” compares two 0.75s with different meanings. This is not a statistical precision issue but a semantic one.

    There is also an unintuitive side effect. Letting δ fill the structural upper bound 1−m̄ causes the ceiling loss to converge to C_max = 1/(1+m̄−(1−m̄)/8): as m̄→1, C_max→0.500; when m̄=0.50, C_max=0.696 (closed-form and numerical difference ≤1.2×10⁻¹⁴). Thus, regardless of how concentrated the marginal distribution is, if the difference between two readers’ margins hits the upper bound, the ceiling inevitably drops to 0.47–0.50. This corrects the notion that “high m̄ regions are more dangerous”: high m̄ regions merely mean “a very small δ can cause large loss” (amplification factor (1−p_e)⁻¹ up to 25×), while the maximum possible loss is actually smaller (0.50 vs. 0.696). The two directions are exactly opposite: low m̄ regions incur large losses but are hard to trigger; high m̄ regions incur small losses but are easily triggered. Assessing risk requires the product of “trigger difficulty × loss magnitude,” not just one factor—this explains why earlier posts alternately claimed high m̄ is dangerous and advised against comparison, appearing self-contradictory: each looked at only half the picture.

    The final empirical rule directly applicable to reporting standards: set C(δ)=0.05, then δ/(1−m̄) monotonically rises from 7.0% (m̄=0.60) to 9.9% (m̄=0.99), nearly constant. Once the difference in main file proportions exceeds approximately 10% of the structural upper bound, this difference alone consumes the entire 0.05 bandwidth—adding more sample size thereafter is futile.

    This rule requires only two numbers. Step one: read the two readers’ main file proportions from the report and take their difference as δ. Step two: compute δ’s percentage of (1−m̄). If it falls below 10%, the study still has room to “be saved by adding sample size.” Once it exceeds 10%, regardless of how large n or how beautiful the confidence interval reported, its consistency upper bound is locked by its own marginal difference and cannot be recovered. For example, in a study with m̄=0.95, (1−m̄)=5pp, so 10% corresponds to δ exceeding 0.5pp—if two readers differ by half a percentage point in main file proportion, the ceiling immediately consumes 0.053. Half a percentage point means one extra person classified into the main file among 60; random error can achieve this.

    Therefore, the final operational procedure for “which correction to apply” follows these three steps in strict order: ① first compute δ’s proportion; if it exceeds 10%, judge the study’s upper bound as locked and enter the “non-compensable” region; ② if not, compute the δ–ρ dividing band to determine whether the δ term or ρ term dominates D_min; ③ only if ρ dominates does the discussion of “adding sample size to buy it back” come into play. Doing the reverse—starting with D_min and working backward to estimate each correction’s share—falls into the isoline trap: behind the same number may lie either a disease that can be bought back or one that cannot.

    Evidence and Limitations

    Verified (algebraic and numerical checks within this paper):
    – Cross-term identity cross = [K(ρ)−K(0)]·[SE(δ)−SE(0)], derived algebraically without approximation.
    – p_e = m̄² + (1−m̄)²/2 − (3/8)δ², verified across an 8×6 grid with maximum deviation 1.1×10⁻¹⁶.
    – Absolute lower bound of the ceiling given by closed-form C_max = 1/(1+m̄−(1−m̄)/8), with numerical difference across the entire domain ≤1.2×10⁻¹⁴; limit as m̄→1 is 0.500.
    – “ρ can be bought back, δ cannot”: SE ∝ n^(−1/2) while C(δ) lacks n, representing a definitional property independent of parameter choice.

    Partially verified:
    – The numerical value of the cross-term proportion (median 0.054%) holds only for the symmetric parameter family; in asymmetric families it rises to 1.78–2.43%, with maximum 8.97%, approaching the 10% criterion threshold. Thus, the conclusion “can be applied independently” requires the additional premise of “symmetric or weakly asymmetric.”
    – The location of the “dividing band” depends on k₀=0.5, a nominal coefficient value; if the κ reported in real studies systematically deviates from 0.5, the absolute magnitude of SE will change, shifting the dividing band (its direction remains unchanged, but its position is not calibrated).

    Pending verification:
    – All conclusions rest on analytical forms and have not been compared against any set of real three-tier interpretation data. The joint distribution of (δ, ρ) in real data is unknown, so “which side of the dividing band common configurations fall on” cannot be answered.
    – δ and ρ are treated as independent inputs, yet #1388 proved that ρ’s primary source is “shared reader thresholds”—and differences in these thresholds themselves may simultaneously generate δ. If both originate from the same source, this paper’s additive decomposition holds causally but not attributively.

    L1 Network Status: This topic has not yet entered the L1 corresponding network—CORRESPONDENCE_NETWORK.md contains no nodes in its three-domain table (Yijing Studies, TCM side; Physics side; Physiology/Experience side) that originate from or target “interpretation reliability / consistency coefficient / marginal bias / dependence structure / error compensability.” The edge table stops at E-050 “Sixty Jiazi Na Yin | Correspondence | 60 frequencies” (open, original text of modern evidence column states “no measurable source yet”); grepping CORRESPONDENCE_NETWORK.md for terms like “interpretation / reliability / consistency / marginal / bias / bandwidth / sufficiency / sample size / dependence / kappa / discernible / threshold” hits only line 79 (level definition) and line 159 (“L2+ edge count | 0”). EDGE_TABLE_science-yixue.md merely counts article numbers and highest evidence levels for 47 concepts (frequency 407 / HRV 214 / pulse diagnosis 73), lacking entries for “compensability / cross-term / permanent bias.”

    Related Reading

    Evidence Level: C
    References: PMID 8501467, PMID 2348207, PMID 8621993, PMID 25168681, PMID 41230191, PMID 36127220, PMID 24338600, PMID 18430244, PMID 25321842, PMID 24488927, PMID 10501650, PMID 37127220


    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.

  • 高阶受益更多,可能只是个算术:ΔICC 的阶次比较里混着多少天花板,三种归一化能不能判出来

    Evidence level: C

    「第 7 阶涨了 0.12、第 1 阶只涨了 0.02」这种说法,在原始 ICC 差得远的时候根本不成立——一个是 0.97 涨到 0.99,一个是 0.80 涨到 0.92,后者涨幅大不是因为高阶更受益,是因为它本来就有地方可涨。

    正文

    一、这个坑有多硬

    先把算术摆出来。ICC 的天花板是 1,所以一个指标的”可改善空间”就是 1 减去它的原始 ICC。第 1 阶原始 ICC 若是 0.97,它最多只能涨 0.03;第 10 阶若是 0.80,它有 0.20 的空间。这时你直接比较两者的 ΔICC,结论必然偏向高阶——这不是发现,这是减法。任何”高阶受益更多”的结论,只要没先处理这条上限,就都还站在算术里,没站在生理里。

    更麻烦的是,天花板不只是”封顶”这么简单,它会连带改变同一条量表不同位置的测量学属性。Berg 平衡量表 11 项研究 668 人的系统综述给了最干净的例证:相对信度极高(组内 0.98、组间 0.97),但绝对可靠性随量表位置而变——MDC95 在 2.8 到 6.6 分之间浮动,而且越靠近满分(56 分)绝对可靠性反而”更好”,原因正是天花板效应把变异压小了(PMID 23663794)。换句话说,同一个指标在量表高端和低端,压根不是同一把尺子。这条证据说明:把各阶 ICC 直接横向相减,等于把不同量程的尺子读数拿来比大小。

    二、三种归一化,三种假设

    既然不能直接减,就得归一化。但每一种归一化都偷偷带了一个假设。

    第一种,除以到上限的距离(ΔICC / (1 − ICC_原始))。它算的是”可改善空间兑现比例”。假设是:剩余空间等价值——一个指标从 0.97 到 1.0 的 0.03,和一个指标从 0.80 到 0.83 的 0.03,意义相同。这个假设在量表研究里已经被打过脸:顶端那 0.03 通常极难拿到,底端那 0.03 随手就有。Berg 的 MDC 在顶端反而更小,正是因为顶端挤不出变异(PMID 23663794)。

    第二种,MDC 比值(ΔICC 与各阶自身 MDC95 的比)。它没有上限问题,但受各阶噪声基线牵着走。而 MDC 本身位置依赖——同一个量表不同段落的 MDC 能差一倍以上(PMID 23663794)。用一个随位置变的尺子去除另一个随位置变的量,出来的比值不一定更干净。

    第三种,把 ICC 先做 Fisher z 变换再比。z 变换是为相关系数设计的方差稳定化手段,用在 ICC 上属于借用。借用有依据也有代价:非参数 bootstrap 给 ICC 建区间时,10 个簇的情形下覆盖严重不足,加到 50 个簇才接近名义 95%,而对 rho 先做方差稳定化变换再用 bootstrap-t,覆盖才回到正常(PMID 14673940)。这说明 ICC 的抽样分布确实不正态、确实需要变换;但同时也说明变换解决的是”区间准不准”,不是”该不该这样归一化”——它把统计推断修好了,却不会告诉你阶次分层是真是假。

    三者混在一起最危险的地方在于:它们可能给出一致的方向,却给出不一致的排序。而”哪几阶受益最多”这个结论,命全押在排序上。

    三、三角验证怎么做

    做法不复杂,但要求同一批数据。

    取 ≥120 人、跨季多点的数据(跨季是为了让个体间差异撑开,避免样本太齐把 ICC 分子压小),逐阶算原始 ICC 与校正后 ICC,得到 ΔICC 剖面;然后分别用三种归一化各算一条”阶次增益剖面”。三条剖面并排放,看峰值落在哪几阶。

    判据是排序一致性,用 Kendall 协和系数 W 量化:

    • W > 0.7:三条剖面峰值落在同一 2-3 个阶次内 → 阶次分层是稳健的,不是某种归一化的产物,可以进分级表
    • W < 0.4:三条峰值位置互不一致 → 当前样本量下”哪几阶受益最多”不可判定。此时任何单一归一化下的结论都不能发表,应当扩样到能同时压低三条剖面的置信区间
    • 只有 Fisher z 与其他两者分歧:问题不在生理分层,而在 ICC 的非正态抽样分布(PMID 14673940),出路是改 bootstrap 置信区间,而不是再找一个归一化

    还有一条前置检查很容易漏:比较两个 ICC 的差,本身需要区间。相依 ICC 之差的置信区间有成熟构造法,其覆盖率与尾误差在模拟中表现良好(PMID 19142855)。不做这一步,你连”第 7 阶的 0.12 和第 1 阶的 0.03 是不是真的不一样”都答不了——0.12 与 0.03 的差,可能整个躺在噪声里。

    四、为什么这个问题从来没被暴露

    因为文献里根本没人逐阶报。

    脉波谐波分析的标准协议研究,22 名健康受试者,前 11 阶的组内重测 ICC 极高、组间 0.83–0.96,Bland-Altman 图 90% 以上落在两倍标准差内(PMID 27170904);设备内部可靠性用人工脉波发生器验证,11 阶 ICC 全部 >0.95、CV 全部 <3%(PMID 25693606)。两篇都是很扎实的工作,但它们报的是整体区间,不是逐阶剖面。整体区间一报,上限问题就被平均掉了——你看到的是”这批谐波很可靠”,看不到”第 1 阶已经挤不出空间了”。

    再往上,高阶到底有没有独立的判别价值,现有数据也不乐观:112 名社区老人、40 个谐波指标做衰弱前期判别,常规机器学习 AUC 只到 0.63 以下,专门设计的脉波分布分析法到 0.70(PMID 41932586)。判别增益本就微弱,再叠一层统计假象,真假就更分不开。

    顺带一个方向相反的旁证:天花板低和反应性好经常同时出现。KOOS 与膝关节学会功能评分对比,KOOS 反应性更高、天花板更低,作者因此判 KOOS 更优(PMID 25416964);KOOS-12 与四个量表对比,以 >15% 满分比例判天花板存在,KOOS 的疼痛、ADL、QoL 子量表都超线,KOOS-12 没超(PMID 32300847)。这说明天花板比例是可测的、也有习惯阈值(15%)——”某阶顶到天花板”这件事并非不可查,只是没人查过。

    证据与局限

    已验证:① 天花板会改变同一条量表不同位置的绝对可靠性,MDC95 随位置从 2.8 分变到 6.6 分(PMID 23663794);② 天花板比例有通行判据,>15% 满分比例视为存在天花板(PMID 32300847);③ ICC 抽样分布不正态,10 簇时 bootstrap 区间覆盖严重不足,须方差稳定化变换(PMID 14673940);④ 相依 ICC 之差的置信区间有成熟构造法(PMID 19142855);⑤ 脉波 11 阶谐波整体重测 ICC 高、CV<3%,但文献只报整体区间,不报逐阶(PMID 27170904、PMID 25693606);⑥ 40 个谐波指标做判别,AUC 仅 0.63–0.70(PMID 41932586)。

    部分验证:三种归一化在概念上各有假设、各有偏,这一点成立;但”三条剖面并排能否判定阶次分层”这一步,尚无任何研究实际做过,也没有任何已发表数据可以反算 W 值。

    待验证:① 三种归一化在同一批脉象数据上是否给出相同排序,完全未知;② 判据 W > 0.7 / < 0.4 这两个门槛没有出处,是从”排序一致性高/低”的习惯切分里拍的,属于待标定的方便数;③ ≥120 人这个样本量未做功效推算,是从跨季多点重测的一般规模借来的;④ “剩余空间等价值”假设是否成立,只有量表侧的间接旁证,没有谐波侧的实测。

    这一篇的方法学性质必须说白:它不提供任何”高阶是状态层、低阶是身份层”的新证据,只提供一把尺子——用来判断过去那些”高阶受益更多”的结论里,有多少是算术、有多少是生理。在本篇的证据分级下,这个问题尚无一条直接实验数据,全部推演建立在量表测量学的外推上。

    相关阅读

    证据等级: C
    参考文献: PMID 23663794, PMID 32300847, PMID 25416964, PMID 14673940, PMID 19142855, PMID 27170904, PMID 25693606, PMID 41932586

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 全文检索「谐波 / ICC / 重测 / 信度 / 阶次」零命中;边表止于 E-050(纳音↔频率,L1、open,现代证据栏原文「尚无任何可测来源」);EDGE_TABLE_science-yixue.md 仅有「谐波 147 篇 / A 级 20」这类文章计数,无「重测信度 / 阶次分层 / 归一化」条目。故本篇证据等级上限为 C,全文不将任何阶次分层结论表述为已验证机制。


    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。

    Higher-order benefits prevail, possibly just arithmetic: How much ceiling is mixed into the order comparison of ΔICC, and can three normalizations discern it?

    Evidence level: C

    The claim that “the 7th harmonic increased by 0.12 while the 1st only increased by 0.02” does not hold when raw ICCs differ substantially—one rising from 0.97 to 0.99, the other from 0.80 to 0.92. The larger gain in the latter is not because higher orders benefit more; it is because they had more room to improve.

    Main Text

    I. How Hard Is This Pit?

    Let us lay out the arithmetic first. Since the ICC ceiling is 1, the “room for improvement” of any metric equals 1 minus its raw ICC. If the 1st harmonic’s raw ICC is 0.97, it can increase by at most 0.03; if the 10th harmonic’s raw ICC is 0.80, it has 0.20 of room. Comparing their ΔICC values directly will inevitably favor higher orders—this is not a discovery, it is subtraction. Any conclusion that “higher orders benefit more” remains purely arithmetic unless we first account for this ceiling; it does not reflect physiology.

    Compounding this, the ceiling also alters the psychometric properties of different positions on the same scale. A systematic review of 11 studies involving 668 participants using the Berg Balance Scale provides a clear example: relative reliability is extremely high (intra-rater 0.98, inter-rater 0.97), yet absolute reliability varies with scale position—the Minimal Detectable Change at 95% confidence (MDC95) ranges from 2.8 to 6.6 points, and paradoxically becomes smaller near the maximum score (56 points) because the ceiling effect suppresses variability (PMID 23663794). In other words, the same metric at the high end versus the low end of a scale is effectively measured with different rulers. This evidence shows that directly subtracting ICCs across harmonics compares readings from scales of different ranges.

    II. Three Normalizations, Three Assumptions

    Since direct subtraction is invalid, we must normalize. However, each normalization carries its own hidden assumption.

    First, divide by the distance to the ceiling (ΔICC / (1 − ICC_raw)). This calculates the “proportion of improvement space realized.” The assumption is equal residual value: a 0.03 increase from 0.97 to 1.00 is considered equivalent to a 0.03 increase from 0.80 to 0.83. This assumption has been challenged in scale research: the top 0.03 is usually extremely hard to achieve, while the bottom 0.03 comes easily. In the Berg scale, MDC is actually smaller at the top precisely because the ceiling suppresses variability (PMID 23663794).

    Second, the MDC ratio (ΔICC divided by each harmonic’s own MDC95). This avoids the ceiling problem but is dragged along by each harmonic’s noise baseline. Since MDC itself is position-dependent—the same scale can have MDC values differing by more than twofold across different segments (PMID 23663794)—dividing one position-dependent quantity by another yields a ratio that may not be cleaner.

    Third, apply Fisher’s z transformation to ICCs before comparison. The z-transformation is a variance-stabilizing technique designed for correlation coefficients; using it for ICCs is borrowed methodology. Borrowing has justification but also cost: when constructing confidence intervals via non-parametric bootstrap, coverage is severely inadequate for 10 clusters, approaching nominal 95% only at 50 clusters; applying variance stabilization to ρ before bootstrap-t restores proper coverage (PMID 14673940). This demonstrates that ICC sampling distributions are indeed non-normal and require transformation; however, the transformation fixes “how accurate the interval is,” not “whether this normalization is appropriate”—it repairs statistical inference but does not tell us whether harmonic stratification is real.

    The most dangerous aspect of mixing these three methods is that they may yield consistent directional trends but inconsistent rankings. Since the conclusion “which harmonics benefit most” hinges entirely on ranking, this inconsistency is critical.

    III. How to Perform Triangulation

    The procedure is straightforward but requires the same dataset across all methods.

    Collect data from ≥120 participants spanning multiple seasons (multi-season sampling ensures individual differences stretch out, preventing overly homogeneous samples from compressing the ICC numerator). Calculate raw ICCs and corrected ICCs for each harmonic to obtain ΔICC profiles; then apply each of the three normalizations separately to generate three “harmonic gain profiles.” Plot these three profiles side-by-side and observe where the peaks fall.

    The criterion is ranking consistency, quantified using Kendall’s coefficient of concordance W:

    • W > 0.7: Peaks of all three profiles fall within the same 2–3 harmonic orders → harmonic stratification is robust, not an artifact of any single normalization, and can proceed to graded tables
    • W < 0.4: Peak positions differ significantly across the three profiles → under current sample size, “which harmonics benefit most” is undeterminable. No conclusion based on any single normalization should be published; instead, expand the sample until confidence intervals for all three profiles can be simultaneously narrowed
    • Only Fisher z differs from the other two: The issue lies not in physiological stratification but in ICC’s non-normal sampling distribution (PMID 14673940). The solution is to adjust bootstrap confidence intervals rather than seek another normalization

    One preliminary check often overlooked: the difference between two ICCs itself requires a confidence interval. Confidence intervals for differences of dependent ICCs have established construction methods, with good coverage and tail error performance in simulations (PMID 19142855). Without this step, you cannot even answer “is the 0.12 gain of harmonic 7 truly different from the 0.03 gain of harmonic 1?”—the difference between 0.12 and 0.03 might lie entirely within noise.

    IV. Why Has This Problem Never Been Exposed?

    Because literature rarely reports harmonics individually.

    Standard protocols for pulse wave harmonic analysis report overall intervals, not per-harmonic profiles. Studies of 22 healthy subjects found intra-rater ICCs extremely high for the first 11 harmonics, inter-rater ICCs ranging from 0.83 to 0.96, with Bland-Altman plots showing >90% of differences falling within twice the standard deviation (PMID 27170904). Device reliability was verified using artificial pulse wave generators, yielding ICCs >0.95 for all 11 harmonics and coefficients of variation (CV) <3% (PMID 25693606). Both studies are rigorous, yet they report overall intervals, not per-harmonic profiles. Reporting overall intervals averages out the ceiling problem—you see “these harmonics are reliable” but miss “harmonic 1 has no room left.”

    Moving upward, existing data is not optimistic regarding whether higher harmonics possess independent discriminative value: in a study of 112 community-dwelling elderly individuals assessing pre-frailty using 40 harmonic indicators, conventional machine learning achieved AUC <0.63, while a specially designed pulse wave distribution analysis reached 0.70 (PMID 41932586). Discriminative gains are inherently weak; layering statistical artifacts on top makes true/false distinction even harder.

    As a side note in the opposite direction, low ceilings and high responsiveness often co-occur. Comparing KOOS with the Knee Society Function Score, KOOS shows higher responsiveness and lower ceiling; authors therefore judged KOOS superior (PMID 25416964). Comparing KOOS-12 with four other scales, those exceeding >15% of the maximum score were deemed to have ceiling effects; KOOS’s pain, ADL, and QoL subscales all exceeded this threshold while KOOS-12 did not (PMID 32300847). This indicates ceiling proportion is measurable and has conventional thresholds (15%)—”a harmonic hitting the ceiling” can indeed be detected, though no one has checked.

    Evidence and Limitations

    Verified: ① Ceiling effects alter absolute reliability at different positions on the same scale; MDC95 varies from 2.8 to 6.6 points depending on position (PMID 23663794); ② Ceiling proportion has established criteria, with >15% of maximum score considered indicative of ceiling (PMID 32300847); ③ ICC sampling distributions are non-normal, requiring variance stabilization transformation for bootstrap intervals with 10 clusters (PMID 14673940); ④ Confidence intervals for differences of dependent ICCs have established construction methods (PMID 19142855); ⑤ Pulse wave harmonic analysis reports high overall retest ICCs and CV <3% for the first 11 harmonics, but literature only reports overall intervals, not per-harmonic (PMID 27170904, PMID 25693606); ⑥ Discrimination using 40 harmonic indicators yields AUC only 0.63–0.70 (PMID 41932586).

    Partially Verified: The conceptual assumption that each of the three normalizations carries its own bias holds; however, whether “placing three profiles side-by-side can determine harmonic stratification” has never been tested in actual research, nor is there any published data from which W values can be back-calculated.

    Pending Verification: ① Whether the three normalizations yield identical rankings when applied to the same set of pulse data is completely unknown; ② The thresholds W > 0.7 / < 0.4 lack citation and were derived from conventional divisions of “high/low ranking consistency,” serving as convenient numbers pending calibration; ③ The sample size of ≥120 was borrowed from general multi-season retest studies without formal power calculation; ④ Whether the “equal residual value” assumption holds is supported only by indirect evidence from scale research, not by direct measurement of harmonic data.

    This article’s methodological nature must be stated clearly: it provides no new evidence that “higher harmonics represent state layers while lower harmonics represent identity layers.” Instead, it offers a ruler—to determine how much of past conclusions that “higher harmonics benefit more” stems from arithmetic versus physiology. Under this evidence grading, no direct experimental data exists for this problem; all inference relies on extrapolation from scale psychometrics.

    Further Reading

    Evidence Level: C
    References: PMID 23663794, PMID 32300847, PMID 25416964, PMID 14673940, PMID 19142855, PMID 27170904, PMID 25693606, PMID 41932586

    L1 Network Status: This topic has not yet entered the L1 correspondence network. Full-text search of CORRESPONDENCE_NETWORK.md for “harmonic / ICC / retest / reliability / order” yields zero hits; the edge table stops at E-050 (Na Yin ↔ Frequency, L1, open, with the original text stating “no measurable source available yet”); EDGE_TABLE_science-yixue.md contains only article counts such as “harmonic 147 / A-level 20,” but lacks entries for “retest reliability / harmonic stratification / normalization.” Therefore, the maximum evidence level for this article is C, and no conclusion about harmonic stratification will be stated as a 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.

  • 「变异系数」和「重测 ICC」量的不是同一件事:CV 高而 ICC 也高的阶次必然存在,所以两套脉象文献不能互相引证

    Evidence level: C

    一个阶次”CV 高”,只说明它相对自己那点幅度抖得厉害;一个阶次”ICC 高”,只说明人跟人之间的差距盖过了抖动。这两句话可以同时为真——而且在低幅度的高阶谐波上,几乎必然同时为真。

    正文

    一、那张分级表有两条腿,而它们从没在同一批人身上对齐过

    脉象谐波的阶次分层,靠两条互相独立的证据链撑着。

    横向那条用波动幅度说话:临终病人的谐波变异系数(HCV)随病程逐阶升高,健康人低于 5%、门诊病人低于 8%,而癌症病人第 3 至 6 阶超过 15%,临终当天连第 1、2 阶也越过 15%(PMID 15472498);卒中患者患侧的 CV1–CV7 显著大于健侧,但对照健康人时只有 CV1、CV3–CV5、CV7 仍显著,CV2 与 CV6 掉队(PMID 22156035)。纵向那条用重测相关说话:22 名健康人的桡动脉谐波,观察者内重测信度被评为”极好”,观察者间 ICC 落在 0.83–0.96(PMID 27170904);用人工脉搏发生器测同一台设备的固有可靠性,则 ICC 全部大于 0.95、CV 全部小于 3%(PMID 25693606)。

    两条链在论文里被并列引用,仿佛一个说”高阶不稳”、一个说”测得准”,合起来就是”高阶不稳但测得出”。问题在于:CV 与 ICC 在数学上根本不是同一个量,并列引用不是互补,是偷换。

    二、一条恒等式把它们连起来,也把它们的区别说清楚了

    经典测量理论给了一条桥:SEM = SD × √(1 − ICC)(PMID 15705040)。把两边同除均值,就得到

    CV_个体内 = (SD/均值) × √(1 − ICC)

    这条式子说明两件事。第一,它们不是无关的:知道 CV、均值与 ICC 中的一个,可以反推另一个。第二,它们不同源:ICC 是比值,只取决于个体间方差与个体内方差的对比;CV 是比值再除均值,对绝对幅度敏感。

    于是”CV 高而 ICC 也高”不是悖论,是代数上的必然。设均值 100:个体间 SD=30、个体内 SD=10 时 ICC=0.90、CV=10%;另一情形个体间 SD=10、个体内 SD=6 时 ICC=0.735、CV=6%。前者 CV 更高、ICC 也更高。缩小均值则更极端:高阶谐波以相对幅度 Cn 表示,天生数值小,均值取 10、个体内 SD=1(CV=10%)、个体间 SD=5 时,ICC 高达 0.96。

    这就是要害:高阶”CV 高”,很可能只是分母小,不是波动大。 因为 Cn 是相对幅度,阶次越高数值越小,同一份绝对抖动除下去自然被放大。用 CV 论证”高阶先失稳”,如果不报个体内 SD 的绝对值,就等于用一把刻度随位置缩放的尺子量长短。

    三、还有三层不同,让互引更加站不住

    第一层是时间尺度。HCV 与 CVn 算的是逐拍变异,秒到分钟级的生理波动;而 ICC 算的是测量之间的变异,跨观察者、跨时段。两者连方差分量都不是同一个。

    第二层是群体依赖。ICC 的分子是个体间方差,样本越异质,ICC 越高——这不是缺陷,是定义。正因如此,足底两点辨别在 41 名多发性硬化患者中 ICC 达 0.87 与 0.90,作者自己就注明”很可能被样本异质性抬高”(PMID 33105127);幻肢痛三项量表 ICC 0.8–0.9 同样附了这句注(PMID 38639472)。ICC 值一旦离开它的人群就不能搬运,而 CV 至少还带着均值的印记。

    第三层是定义口径。ICC 有十种形态,选错形态结论就不同,报告时必须写清 model、type、definition(PMID 27330520);而相关本身就不该拿来判断一致性,这是 Bland 与 Altman 四十年前就指出的(PMID 2868172)。

    四、同一个词,在领域内已经被当成两种东西用

    最能说明问题的是:频域脉图研究把”谐波变异系数”当成生理特征在用。一项 280 人、40 个频域指标判别血管老化的研究,特征集就包含幅度占比 Cn 与其变异系数、相位 Pn 与其标准差,AUC 超过 0.8(PMID 34508787)。也就是说,在同一个领域里,CV 有时被当”这个人状态如何”的读数,有时被当”这台仪器测得准不准”的指标。同一个符号,两个身份,没人给它分过工。

    五、判据与结论

    要把这张分级表的坐标系定下来,必须按三条走:

    其一,两指标并列报告,不许单用。任一声称”某阶不稳”或”某阶可辨人”的结论,须同时给出该阶的个体内 SD(或绝对 CV)与 ICC,缺一即不可判。

    其二,换算前先对齐人群与时间尺度。只有同一批人、同一记录时长、同一 ICC 形态下,CV = (SD/均值)×√(1−ICC) 才成立;跨文献搬运数值,等于把三个未知量当已知用。

    其三,“CV 高且 ICC 高”的阶次是身份层,不是状态层。它意味着”这个人此刻抖,但人跟人差得更多”——这正是低阶身份、高阶状态的画像,但它由方差对比决定,与 ΔICC 无关,因此它可以作为 #508 结局一的独立第二证据,前提是先排除分母效应。

    在做到这三条之前,横向的 CV 文献与纵向的 ICC 文献只能各自成立,不能互相引证。分级表想回答”哪几阶稳”,就用 CV 系;想回答”哪几阶能认人”,就用 ICC 系。选一个,并声明选的是哪一个。

    证据与局限

    已验证:CV 与 ICC 的换算关系由经典测量理论给出,SEM = SD×√(1−ICC) 是教科书级结论(PMID 15705040);百分比/比例尺度的相对离散度对均值敏感,是基础统计学性质;ICC 随样本异质性升高,已在两项独立的复测研究中被作者明确标注(PMID 33105127、PMID 38639472);ICC 存在多种形态、报告须声明(PMID 27330520);相关系数不能替代一致性判断(PMID 2868172)。

    部分验证:谐波幅度在高阶变小、低阶变大,是本领域普遍观察(PMID 15472498 的健康对照 HCV 全部低于 5%、PMID 34508787 的 Cn 随阶次变化),但”高阶 CV 上升中有多少来自分母缩小”从未被拆开算过;观察者内优于观察者间(PMID 27170904 的 ICC 从”极好”降到 0.83–0.96)提示测量噪声可分层,但样本仅 22 人,区间偏宽。

    待验证:同一批受试者、同一记录时长下逐阶同时计算个体内 CV 与重测 ICC,画出 CV–ICC 散点并按阶次连线,看是否存在系统性的”双高”簇——这一步在本领域没有做过,本文的代数论证只保证”可能存在”,不保证”在真实脉图上以多大比例存在”;此外,逐拍 HCV 与跨次重测 CV 之间的换算系数(时间尺度差异的定量折损)尚无任何实测标定。

    证据等级: C
    参考文献: PMID 15705040, PMID 27170904, PMID 25693606, PMID 15472498, PMID 22156035, PMID 33105127, PMID 38639472, PMID 27330520, PMID 2868172, PMID 34508787

    L1 网络状态:该主题尚未进入 L1 对应网络。经检索 CORRESPONDENCE_NETWORK.md 全文,「谐波 / ICC / 重测 / 信度 / 阶次 / 变异系数」均零命中;边表止于 E-050 | 六十甲子纳音 | 对应 | 60 个频率 | 京房六十律 / 纳音表 | **尚无任何可测来源** | L1 | open。网络可表达「某阶谐波与某脏腑/某状态有关」,但无法表达「某阶的波动幅度与可辨人性之间是否同源」这一测量学裁决,故本文证据等级封顶为 C。

    相关阅读


    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。

    Coefficient of Variation and Test-Retest ICC Measure Different Things: High CV with High ICC is Inevitable, So Two Sets of Pulse Literature Cannot Cite Each Other

    Evidence level: C

    A “high CV” for a given harmonic order merely indicates that the signal fluctuates violently relative to its own low amplitude; a “high ICC” for that same order merely indicates that inter-individual differences outweigh the measurement noise. Both statements can be true simultaneously—and, on low-amplitude high-order harmonics, they are almost certainly both true.

    Main Text

    I. The Classification Table Has Two Legs That Have Never Been Aligned on the Same Cohort

    The stratification of pulse wave harmonics relies on two mutually independent chains of evidence.

    The horizontal chain speaks in terms of fluctuation amplitude: the Harmonic Coefficient of Variation (HCV) rises stepwise with disease progression in terminally ill patients, whereas healthy individuals remain below 5%, outpatient patients below 8%, cancer patients exceed 15% for orders 3 through 6, and on the day of death even the 1st and 2nd orders surpass 15% (PMID 15472498); in stroke patients, the CV of the affected side (CV1–CV7) is significantly higher than the healthy side, yet compared to healthy controls only CV1, CV3–CV5, and CV7 remain significant, while CV2 and CV6 lag behind (PMID 22156035).

    The vertical chain speaks in terms of retest reliability: the intrarater ICC for radial artery harmonics in 22 healthy subjects was rated “excellent,” with inter-rater ICCs ranging from 0.83 to 0.96 (PMID 27170904); furthermore, testing the intrinsic reliability of the same device using an artificial pulse generator yielded ICCs all exceeding 0.95 and CVs all under 3% (PMID 25693606).

    In the literature, these two chains are cited side-by-side, creating the impression that one claims “high-order instability” while the other claims “high measurement accuracy,” thereby combining to mean “high-order instability but measurable.” The problem is that CV and ICC are mathematically distinct quantities; citing them together is not complementary but a substitution of concepts.

    II. An Identity Equation Links Them and Clarifies Their Difference

    Classical measurement theory provides the bridge: Standard Error of Measurement (SEM) = SD × √(1 − ICC) (PMID 15705040). Dividing both sides by the mean yields:

    $$CV_{\text{intra}} = \frac{SD}{\mu} \times \sqrt{1 – ICC}$$

    This equation reveals two facts. First, they are not independent: knowing any one of CV, mean, or ICC allows calculation of the others. Second, they originate differently: ICC is a ratio, depending solely on the contrast between inter-individual variance and intra-individual variance; CV is a ratio divided by the mean, making it sensitive to absolute amplitude.

    Thus, “high CV and high ICC” is not a paradox but an algebraic necessity. Consider a mean of 100: if inter-individual SD = 30 and intra-individual SD = 10, then ICC = 0.90 and CV = 10%; in another scenario with inter-individual SD = 10 and intra-individual SD = 6, ICC = 0.735 and CV = 6%. In the first case, CV is higher and ICC is also higher. Reducing the mean makes this more extreme: for high-order harmonics expressed by relative amplitude $C_n$, which are inherently small, setting the mean to 10, intra-individual SD = 1 (CV = 10%), and inter-individual SD = 5 yields an ICC as high as 0.96.

    This is the crux: a “high CV” at high orders likely stems from a small denominator, not large fluctuation. Since $C_n$ represents relative amplitude, higher orders have smaller numerical values; dividing the same absolute jitter by a smaller mean naturally amplifies the CV. Using CV to argue “high-order instability” without reporting the absolute intra-individual SD is akin to measuring length with a ruler whose scale shrinks at each position.

    III. Three Additional Layers Make Mutual Citation Even Less Plausible

    The first layer is time scale. HCV and CV$_n$ measure beat-to-beat variation, reflecting physiological fluctuations on the scale of seconds to minutes; ICC measures variation between measurements, spanning observers and time periods. Their variance components are fundamentally different.

    The second layer is population dependence. The numerator of ICC is inter-individual variance; the more heterogeneous the sample, the higher the ICC—this is not a flaw but part of the definition. Consequently, foot sole two-point discrimination ICC reached 0.87 and 0.90 in 41 multiple sclerosis patients, with the authors noting it was “likely inflated by sample heterogeneity” (PMID 33105127); similarly, the three-item phantom limb pain scale ICC of 0.8–0.9 carries the same caveat (PMID 38639472). ICC values cannot be transferred across populations, whereas CV retains the imprint of the mean.

    The third layer is definition scope. ICC has ten forms; choosing the wrong form alters conclusions, so reports must specify model, type, and definition (PMID 27330520); moreover, correlation itself should not be used to judge consistency, a point made by Bland and Altman four decades ago (PMID 2868172).

    IV. The Same Term Is Used for Two Different Things Within the Field

    Most illustrative is that frequency-domain pulse wave research treats “harmonic coefficient of variation” as a physiological feature. In a study of 280 individuals evaluating vascular aging via 40 frequency-domain metrics, the feature set included amplitude ratio $C_n$ and its coefficient of variation, phase $P_n$ and its standard deviation, achieving an AUC over 0.8 (PMID 34508787). Thus, within the same domain, CV is sometimes used as a readout of “how is this person’s state?” and sometimes as an indicator of “how accurate is this instrument?” The same symbol, two identities, with no division of labor.

    V. Criteria and Conclusion

    To establish the coordinate system for this classification table, three steps are required:

    First, report both metrics together; do not use one alone. Any claim that “a certain order is unstable” or “a certain order distinguishes individuals” must simultaneously provide the intra-individual SD (or absolute CV) and ICC for that order; missing either renders the conclusion unverifiable.

    Second, align population and time scale before conversion. The equation CV = (SD/mean) × √(1−ICC) holds only for the same cohort, same recording duration, and same ICC form; transferring values across literature equates three unknowns as known.

    Third, the “high CV and high ICC” orders represent identity layers, not state layers. This means “this person fluctuates now, but inter-person differences are even larger”—precisely the portrait of low-order identity and high-order state. However, this is determined by variance contrast and unrelated to ΔICC; thus it can serve as the independent second evidence for Outcome #508, provided the denominator effect is excluded.

    Until these three conditions are met, the horizontal CV literature and the vertical ICC literature can stand on their own but cannot cite each other. If the classification table aims to answer “which orders are stable,” use the CV system; if it aims to answer “which orders distinguish individuals,” use the ICC system. Choose one and declare which you chose.

    Evidence and Limitations

    Verified: The conversion relationship between CV and ICC is given by classical measurement theory, where SEM = SD × √(1 − ICC) is a textbook conclusion (PMID 15705040); the sensitivity of percentage/ratio scales to the mean is a fundamental statistical property; ICC increases with sample heterogeneity, explicitly noted by authors in two independent retest studies (PMID 33105127, PMID 38639472); ICC exists in multiple forms requiring specification in reports (PMID 27330520); and correlation coefficients cannot substitute for consistency assessment (PMID 2868172).

    Partially Verified: The observation that harmonic amplitude decreases at high orders and increases at low orders is universally noted in this field (PMID 15472498 for healthy controls HCV all under 5%, PMID 34508787 for $C_n$ varying with order), but “how much of the rise in high-order CV stems from denominator shrinkage” has never been calculated separately; intrarater reliability exceeds inter-rater (PMID 27170904 ICC dropping from “excellent” to 0.83–0.96) suggests measurement noise can be stratified, yet the sample size was only 22, resulting in a wide interval.

    To Be Verified: Simultaneously calculating intra-individual CV and retest ICC for each order in the same cohort under identical recording duration, plotting the CV–ICC scatter plot and connecting points by order to check for systematic “double-high” clusters—this step has not been performed in this field; the algebraic argument in this paper may support only “possible existence,” not “what proportion exists in real pulse waves”; additionally, the conversion factor between beat-to-beat HCV and cross-session retest CV (quantifying time scale differences) lacks any empirical calibration.

    Evidence Level: C
    References: PMID 15705040, PMID 27170904, PMID 25693606, PMID 15472498, PMID 22156035, PMID 33105127, PMID 38639472, PMID 27330520, PMID 2868172, PMID 34508787

    L1 Network Status: This topic has not yet entered the L1 corresponding network. A search of CORRESPONDENCE_NETWORK.md found zero hits for “harmonic / ICC / retest / reliability / order / coefficient of variation”; the edge table stops at E-050 | Sixty Jiazi Na Yin | Correspondence | 60 frequencies | Jingfang Sixty Laws / Na Yin Table | **No measurable source yet** | L1 | open. The network can express “a certain harmonic order relates to a certain organ or state,” but cannot express the metrological judgment of “whether fluctuation amplitude and distinguishability share the same origin” for a given order; hence, this paper’s evidence level is capped at C.

    Further Reading


    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.

  • 蓝光阻断挡掉的是蓝,还是顺带把光调暗了:一张混杂矩阵把「光谱效应」和「强度效应」分开算账

    Evidence level: C

    橙黄镜片挡蓝的同时也把总光量降了。「光谱有效」这个结论里,有多少是颜色立了功、有多少只是变暗了——这件事至今没人算过账。而账其实是能算的。

    正文

    上一篇把话说到了这一步:既往所有”蓝光阻断有效”的研究,几乎都没做强度匹配——滤片在挡蓝的同时,也一起削掉了到达视网膜的明视觉与黑视素光子通量,两个变量被绑在一起动。于是就留下一个没法回避的追问:那些效应量里,究竟有多少是”颜色”贡献的?

    这件事的雏形其实已被做出一半。Chin 等人 2026 年的实测(PMID 41565717)把 52 件灯具逐一测了褪黑素抑制值(MSV)与黑视素照度:冷白 LED 中位 12.3%、冷白节能灯 12.1%,暖白 LED 3.6%、暖白节能灯 2.6%、白炽灯 1.5%;可调色温灯从 5700 K 的 10% 一路降到 2100 K 的 0.1%;8 片蓝光过滤镜片里,6 片”相对于透明镜片收益有限”,只有 2 片棕色染色片能把估算抑制压到 0.3% 以下。这张表就是混杂矩阵的第一根轴——它同时给了每件器具的明视觉透射与黑视素透射。但它停在”报数”,没做分解:没人把这两个数除一下,看看”选择性”到底是几。

    第一刀:色温不能当代理,这条先要钉死。 用”3000 K 护眼”来推断生物效力是无效的。同一套五通道可调 LED 的模拟显示(PMID 36418869):在 300 lx、显色指数 Rf≥70 的条件下,色温从 2500 K 扫到 6000 K,节律刺激量 CS 从 17% 跨到 41%;更要命的是在同一个 5000 K 色温上,仅因光谱配比不同,黑视素等效照度就能从 196 lx 变到 319 lx——差了 123 lx,六成。所以任何一张混杂矩阵的横轴都必须是黑视素等效照度(mEDI),不能是色温或 lux。

    第二刀:分解本身是初等代数,只是没人写过。 设滤片对某光源的明视觉透射率为 T_v、黑视素透射率为 T_mel,定义选择性 r = T_mel / T_v。因为 mEDI 是乘性的,取对数就能把两条路干净拆开:

    纯光谱份额 = |ln r| / |ln T_mel|
    强度份额   = |ln T_v| / |ln T_mel|
    

    代入第一轴那张表的量级(透射率取该档典型值):清透片 T_v≈0.92 / T_mel≈0.90,r≈0.978——纯光谱只占 21%,八成效应其实是”变暗”;浅黄片 T_v≈0.85 / T_mel≈0.62,r≈0.729——纯光谱 66%,强度 34%;深棕片 T_v≈0.55 / T_mel≈0.05,r≈0.091——纯光谱升到 80%。

    这里浮出一个反直觉的结论:越是”温和、日常能戴”的镜片,被强度混杂得越狠。 清透片和淡黄片贡献了 Cochrane 那 17 项 RCT 里的大部分样本(PMID 37593770,睡眠结局”极低确定性”、无一项测血清褪黑素),而它们的效应量里三分之二到五分之四来自”顺带变暗”。这恰好解释了为什么消费品级研究的效果总比实验室级弱——实验室用的是精心反向补偿的同色异谱对(PMID 29984614,175 lx/2700 K 全匹配,仅调 450-500 nm 功率,抑制从近 50% 掉到与暗光无差异),那里 T_v 被硬按在 1,混杂份额是零。

    第三刀最狠:把 mEDI 匹配上之后,颜色本身就不起作用了。 16 名受试者的重复测量(PMID 38135734):三组光照的黑视素激发全部相等,但明视觉照度分别是 93.5、123.5、67.0 lux(相差近两倍),蓝-黄轴方向也相反——结果在褪黑素相位延迟、抑制幅度、主观嗜睡、警觉性上均无差异。48 人的另一项非视觉同色异谱研究(PMID 40687802)用贝叶斯分析给出更明确的否定:锥体调制对褪黑素抑制与警觉性无可测贡献。

    这两条合起来的意思必须说透:一旦 mEDI 相等,光谱形状就不再单独产生影响。 所以严格讲,”纯光谱效应”不是独立效力,而是同效下的视觉代价折扣。滤片之所以”有效”,是因为它把 mEDI 降下来了;它比”把灯调暗”优越的地方,只是少牺牲明视觉。这个区分有实际后果——所有”低蓝光更健康”的说法,真正的剂量变量是 mEDI 那个数,不是”蓝不蓝”。

    第四刀:矩阵必须是二维的,因为斜率不是常数。 29 篇文献的机器学习汇总(PMID 34981572)确认 mEDI 是最佳单一预测因子,但补了一句关键限定:在 21 lx 以下,S 视锥等效照度反超 mEDI 成为更优预测因子。瞳孔研究也印证剂量与效应的耦合不是线性的(PMID 38348477,72 人、五原色显示、L/M/S 匹配只变黑视素约 3 倍,瞳孔随 mEDI 单调收缩)。所以混杂份额会随剂量区间漂移,一张 滤片 × 剂量区间 的二维表才是可用的。

    据此,反推公式与判决设计可以写死。 若已发表研究报了效应量 Δ,要拆出纯光谱贡献,需拿到三样东西:该镜片的实测 τ(λ)、研究中受试者眼位 mEDI 记录、以及该剂量区间的斜率 f′(mEDI)。则”若只改色不改亮”的等效 mEDI 保留率应取 r 而非 T_mel,两者之差乘斜率即为强度混杂的量。实验上对应的判决:同一批商用滤片与实验级同色异谱对做头对头光学标定,每件测光谱透射曲线、明视觉透射率、mEDI 透射率、以及固定 mEDI 下被一并削掉的明视觉比例,输出二维混杂矩阵;再回溯分解已发表的 BB 研究。

    对”天人相应”的意义: 风水讲”卧室宜暖不宜冷””灯不要正对床头”,历来只有口诀。这张矩阵给的是可证伪版本——它不问”暖色好不好”,只问:给定你要把黑视素剂量降到多少,市面上哪件器具能在视觉代价最小处落到那个点上? 答案不是”越黄越好”(棕染片 T_mel≈0.05 是过度),而是中间那条窄带。

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 与 EDGE_TABLE_science-yixue.md 中均无「光照强度/光谱—生理节律」相关边(概念计数仅见「光波长/色温」作为物理侧节点列出,无对应边条目),本文所用分解为自建代数推导,无 L1 边支撑,证据等级封顶 C。

    证据与局限

    已验证: 色温不能作为生物效力代理,同色温下 mEDI 可差六成(PMID 36418869);黑视素等效照度是最佳单一预测因子,但 21 lx 以下 S 视锥指标更优(PMID 34981572);等 mEDI 条件下蓝-黄方向与明视觉照度差异不产生可测的节律结局差异(PMID 38135734);锥体调制对夜间褪黑素与警觉性无可测贡献(PMID 40687802);滤片与灯具的实测透射差异确实巨大(PMID 41565717)。

    部分验证: 同色异谱可在照度色温全匹配下单独改变褪黑素抑制(PMID 29984614)——但该设计恰恰把 T_v 固定在 1,故它证明的是”光谱能起作用”,不能证明”消费品级滤片的光谱选择性有效”;mEDI 与瞳孔、与睡眠潜伏期存在剂量依赖(PMID 36854795、PMID 38348477),但均在 ≤90 lx 的低照度区测得。

    待验证: 纯光谱份额的对数分解公式尚无任何已发表研究采用,本文给出的 21%/66%/80% 三档为按典型透射率代入的量级估算,非实测;各档滤片的真实 τ(λ) 与 T_v、T_mel 配对值缺失;剂量区间导致的斜率漂移幅度未知;无任何一项已发表 BB 研究同时报告了镜片 τ(λ) 与受试者眼位 mEDI,故反推在当前文献条件下实际不可执行。该主题整体证据薄弱,本文属研究设计讨论。

    相关阅读

    证据等级: C
    参考文献: PMID 41565717, PMID 36418869, PMID 38135734, PMID 40687802, PMID 34981572, PMID 38348477, PMID 29984614, PMID 36854795, PMID 37593770


    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 Blue Light Blocking Filter Out Blue or Dim the Light: A Hybrid Matrix Separates Spectral and Intensity Effects

    Evidence level: C

    Orange-yellow lenses block blue light but also reduce total luminous flux. Within the conclusion of “spectral efficacy,” how much is due to color and how much is simply because it got darker—this accounting has never been done. Yet the accounting is possible.

    Main Text

    The previous discussion reached this point: all prior studies claiming “blue light blocking is effective” failed to match intensity—filters block blue but simultaneously cut the photopic and melanopic photon flux reaching the retina, binding two variables together. This leaves an unavoidable question: of those effect sizes, how much is actually contributed by “color”?

    A prototype of this analysis has already been completed halfway. Chin et al.’s 2026 measurement (PMID 41565717) tested the melatonin suppression value (MSV) and melanopic illuminance of 52 luminaires: cold white LED median 12.3%, cold white CFL 12.1%, warm white LED 3.6%, warm white CFL 2.6%, incandescent 1.5%; tunable color temperature lights dropped from 10% at 5700 K to 0.1% at 2100 K; among 8 blue-filtering lenses, 6 showed “limited benefit compared to clear lenses,” and only 2 brown-tinted lenses could suppress the estimated inhibition below 0.3%. This table forms the first axis of the mixed matrix—it provides both photopic transmittance and melanopic transmittance for each device. However, it stops at “reporting numbers” without decomposition: no one divides these two values to determine the exact “selectivity.”

    First cut: Color temperature cannot serve as a proxy; this must be firmly established. Inferring biological efficacy from “3000 K eye-care” is invalid. Simulation of a five-channel tunable LED system (PMID 36418869) shows that under 300 lx with Rf≥70, varying color temperature from 2500 K to 6000 K changes the circadian stimulus CS from 17% to 41%. More critically, at the same 5000 K color temperature, differing spectral ratios cause melanopic equivalent illuminance (mEDI) to vary from 196 lx to 319 lx—a difference of 123 lx, or 60%. Therefore, the x-axis of any mixed matrix must be melanopic equivalent illuminance (mEDI), not color temperature or lux.

    Second cut: Decomposition itself is elementary algebra, just never written down. Let T_v be the photopic transmittance and T_mel the melanopic transmittance of a filter for a given light source, defining selectivity r = T_mel / T_v. Since mEDI is multiplicative, taking logarithms cleanly separates the two pathways:

    Pure spectral share = |ln r| / |ln T_mel|
    Intensity share      = |ln T_v| / |ln T_mel|
    

    Substituting magnitudes from the first axis table (using typical transmittance values): clear lens T_v≈0.92 / T_mel≈0.90, r≈0.978—pure spectral contribution only 21%, eightfold effect actually “dimming”; light yellow lens T_v≈0.85 / T_mel≈0.62, r≈0.729—pure spectral 66%, intensity 34%; deep brown lens T_v≈0.55 / T_mel≈0.05, r≈0.091—pure spectral rises to 80%.

    This yields a counterintuitive conclusion: the more “gentle, everyday-wearable” lenses are the most contaminated by intensity mixing. Clear and light yellow lenses contributed most samples in Cochrane’s 17 RCTs (PMID 37593770, sleep outcomes “very low certainty,” none measured serum melatonin), yet two-thirds to four-fifths of their effect sizes stem from “incidental dimming.” This precisely explains why consumer-grade studies show weaker effects than lab-grade ones—lab studies use carefully matched isochromatic pairs (PMID 29984614, 175 lx/2700 K full match, adjusting only 450-500 nm power, inhibition dropped from nearly 50% to indistinguishable from dark), where T_v is rigidly fixed at 1, making the intensity mixing share zero.

    Third cut is most severe: once mEDI is matched, color itself becomes irrelevant. Repeated measurements of 16 subjects (PMID 38135734): three lighting conditions had identical melanopic excitation but photopic illuminance differed at 93.5, 123.5, and 67.0 lux (nearly twofold difference), with blue-yellow axis direction reversed—yet no differences emerged in melatonin phase delay, inhibition amplitude, subjective sleepiness, or alertness. Another 48-subject non-visual isochromatic study (PMID 40687802) using Bayesian analysis provides clearer negation: cone modulation contributes nothing measurable to melatonin suppression or alertness.

    These two findings must be made explicit: once mEDI is equal, spectral shape no longer independently affects outcomes. Strictly speaking, “pure spectral effect” is not independent efficacy but visual cost discounting under equivalent conditions. Filters appear “effective” because they reduce mEDI; their advantage over “dimming the light” is merely sparing photopic vision. This distinction has practical consequences—all claims of “low blue is healthier” actually have mEDI as the true dose variable, not “how blue it is.”

    Fourth cut: The matrix must be two-dimensional because slope is not constant. Machine learning synthesis of 29 studies (PMID 34981572) confirms mEDI is the best single predictor, but adds the crucial caveat: below 21 lx, S-cone equivalent illuminance surpasses mEDI as the superior predictor. Pupil research also confirms dose-effect coupling is not linear (PMID 38348477, 72 subjects, five primary colors, L/M/S matching changed melanopic only ~3-fold, pupils monotonically contracted with mEDI). Thus mixing shares drift across dose intervals; only a filter × dose interval two-dimensional table is usable.

    Accordingly, reverse-engineering formulas and judgment designs can be fixed. If published studies report effect size Δ, extracting pure spectral contribution requires three elements: the filter’s measured τ(λ), the study subjects’ eye-position mEDI records, and the slope f′(mEDI) for that dose interval. Then “changing color without changing brightness” equivalent mEDI retention should use r rather than T_mel, with the difference multiplied by slope giving intensity mixing magnitude. Experimental judgment: head-to-head optical calibration of commercial filters against lab-grade isochromatic pairs, measuring spectral transmittance curves, photopic transmittance, mEDI transmittance, and the proportion of photopic vision sacrificed under fixed mEDI, outputting a two-dimensional mixing matrix; then back-calculating decomposition of published BB studies.

    Significance for “天人相应” (Yijing Studies): Feng Shui traditionally states “bedrooms should be warm not cold, lights should not face the headboard”—only mnemonic rules. This matrix provides a falsifiable version—it doesn’t ask “is warm color good,” but rather: given you want to reduce melanopic dose to X, which commercial device lands there with minimal visual cost? The answer is not “the yellower the better” (brown-tinted T_mel≈0.05 is excessive), but that narrow middle band.

    L1 Network Status: This topic has not yet entered the L1 corresponding network. Neither CORRESPONDENCE_NETWORK.md nor EDGE_TABLE_science-yixue.md contains edges related to “light intensity/spectrum—physiological rhythm” (concept counting only lists “wavelength/color temperature” as physical-side nodes, with no corresponding edge entries). The decomposition used here is self-constructed algebraic derivation, lacking L1 edge support; evidence level capped at C.

    Evidence and Limitations

    Verified: Color temperature cannot serve as a proxy for biological efficacy; under identical color temperature, mEDI can differ by 60% (PMID 36418869); melanopic equivalent illuminance is the best single predictor, though below 21 lx S-cone metrics are superior (PMID 34981572); under equal mEDI conditions, blue-yellow axis direction and photopic illuminance differences produce no measurable circadian outcome variation (PMID 38135734); cone modulation contributes nothing measurable to nighttime melatonin suppression or alertness (PMID 40687802); measured transmittance differences between filters and luminaires are indeed substantial (PMID 41565717).

    Partially verified: Isochromatic pairs can independently alter melatonin suppression under full illuminance/color temperature matching (PMID 29984614)—but this design fixes T_v at 1, proving “spectral effects work” but not “consumer-grade filter spectral selectivity works”; mEDI and pupil size, and sleep latency, exhibit dose-dependence (PMID 36854795, PMID 38348477), but all measured in low-illuminance regions ≤90 lx.

    To be verified: Logarithmic decomposition formulas for pure spectral share have not been adopted in any published study; the 21%/66%/80% tiers presented here are magnitude estimates based on typical transmittance substitution, not measured values; true τ(λ) and T_v/T_mel pairing values for each filter tier are missing; slope drift magnitude due to dose intervals is unknown; no published BB study simultaneously reports lens τ(λ) and subject eye-position mEDI, making reverse calculation practically unfeasible under current literature conditions. Overall evidence for this topic is weak; this article constitutes research design discussion.

    Further Reading

    Evidence Level: C
    References: PMID 41565717, PMID 36418869, PMID 38135734, PMID 40687802, PMID 34981572, PMID 38348477, PMID 29984614, PMID 36854795, PMID 37593770


    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.

  • 区间断言逐条判定:20 条 RANGE-unjudged 的「该不该写成区间」裁决(P4-AUDIT-014)

    Evidence level: C

    一句话结论:全库 20 条 RANGE-unjudged 没有一条”该改成单点”——因为触发它们的 σ 本身就是文献区间,δ/A 传播成区间是物理上正确的写法。但其中 16 条存在数值错误(building-mountain 12 条 δ 整体高估 16–100×、B 档 4 条 dB 上界低估),4 条数值自洽(elf-skin-depth 分档门槛)。“区间不可单点判”(R13)说的是判读方式;”区间数值对不对”是另一件事,两者必须分开报。 本文只判不改:修正已发布正文归 D 清单等 JOE 批。

    一、为什么这一题必须做

    第 16 轮(2026-10-23,#63 工具 v1.5)给 delta-table-scan.py 加了 R13「区间断言不可单点判」,把 20 条区间格标为 RANGE-unjudged——既不判对也不判错。第 16 轮随即登记 #65:「12 条区间断言正文处置——须逐条判『该不该写成区间』」。

    但”该不该写成区间”这个问法本身有歧义,必须先拆成两个正交问题:

    问题 含义 答案
    Q1:这格该不该是区间写法? σ 是区间 → δ/A 是区间,还是 σ 本该是单点? 见 §三,逐条判
    Q2:这格的区间数值对不对? 区间的两个端点与 σ 区间自洽吗? 见 §四,独立复算

    R13 只回答了 Q1 的”判读方式”(区间不可当单点),完全没回答 Q2。而 Q2 才是这一轮真正的发现:20 条里有 16 条数值是错的。

    二、方法:判定矩阵

    对 20 条逐条做三件事:

    1. 查 σ 来源:该格对应的 σ 是文献区间还是单点?
    2. 复算端点:δ = √(2/(ωμσ)),把 σ 区间两端代入,得到 δ 区间应然值(A(dB) = 8.686·d/δ 同理)。
    3. 比对方向:区间两端与 σ 递增的关系是否物理正确(σ 越大 δ 越小)。

    判定分三类:

    类 定义 处置
    合理 σ 是区间、δ/A 区间与复算自洽(偏差 < 30%) 保留区间写法;RANGE-unjudged 维持;无动作
    数值错 σ 是区间、但 δ/A 区间端点与复算偏差显著(> 30%) 保留区间写法;改正文端点(归 D 清单等 JOE)
    该单点 σ 本该是单点、被误写成区间 改正文为单点(归 D 清单等 JOE);本轮实测 0 条

    三、逐条判定(20 条)

    3.1 building-mountain-terrain-elf-attenuation-analogy.md(cn L52–57 + lab L41–46 = 12 条)

    表头「1–30 Hz 趋肤深度 δ (km)」,σ 列全部为文献区间 → 全部为合理区间写法。但端点数值与复算偏差显著:

    介质 σ 区间 (S/m) μr 表写 δ (km) 复算 δ @30Hz (km) 偏差(低端/高端)
    干燥花岗岩 1e-5 – 1e-3 1–10 90–900 2.9–29.1 31× / 31×
    玄武岩 1e-4 – 1e-2 1–50 28–280 0.9–9.2 31× / 31×
    沉积岩 1e-3 – 1e-1 1 9–90 0.3–2.9 31× / 31×
    钢筋混凝土 1e-3 – 1e-1 10–100 9–90 0.09–0.9 100× / 100×
    普通红砖 1e-4 – 1e-2 1 28–280 0.9–9.2 31× / 31×
    钢板/钢梁 1e6 – 1e7 1000–10000 0.005–0.05 0.0003–0.004 km 16× / 16×

    复算口径:f=30 Hz、μr 取区间下限(最保守,δ 最大)。若按 1–30 Hz 全频段 + 全 μr 区间,真值区间更宽(花岗岩 0.9–159 km),表写 90–900 km 仍整体高估。

    ⭐ 对第 13 轮结论的修正:第 13 轮(P4-AUDIT-008 §4 D9–D10)把 building-mountain 判为「区间排列与 σ 递增同向 → 读者会读成 σ 越大 δ 越大(物理上反了)」。本轮复算推翻该方向论断:表写 90–900 km 低端小、高端大,与物理一致(σ 小 → δ 大)。真正的问题是数值整体高估约 31×(30Hz 口径),不是方向反了。 方向论断是第 13 轮把”数值错”误读成”方向错”。修正后的 D9–D10 描述:building-mountain §2 表 δ 区间整体高估 ~31×(钢筋混凝土 100×),方向无误。

    为什么值得认真对待:这篇的立论是「山体对 ELF 几乎透明」。按复算(花岗岩 δ 2.9–29 km @30Hz)比表写(90–900 km)更透明——表写反而把山体显得”不那么透明”。与 skin-depth-basement 的 562× 错同族:错把正确结论写得对它自己不利。按 R2 判据不推翻结论,但必须改。

    3.2 elf-skin-depth-table-numeric-integrity-audit.md(cn L205–206 + lab L217–218 = 4 条)

    这是修正版 #22 分档门槛表(§6.2),A′ 档 σ∈[1e-2,1e-1] → 1 km A = 4.8–15 dB;B 档 σ∈[1e-3,1e-2] → 1.5–4.8 dB。复算:

    档 σ (S/m) 1 km A(dB) 复算 表写 判定
    A′ 1e-2 – 1e-1 4.83 – 15.3 dB 4.8–15 dB ✅ 合理(偏差 < 3%)
    B 1e-3 – 1e-2 1.53 – 4.83 dB 1.5–4.8 dB ✅ 合理(偏差 < 3%)

    → 4 条全部数值自洽,保留区间写法,无动作。 这 4 条是”分档门槛”(档位范围),不是测量值——R13 的正确用法就在于此。

    3.3 mountain-sample-priority-sigma-pre-screening-sop.md(cn L63 = 1 条)+ 引用它的两处

    B 档「σ 10⁻³–10⁻¹ → 1 km A = 1–10 dB」。复算:σ=1e-3 → 1.53 dB;σ=1e-1 → 15.3 dB。表写 1–10 dB:

    位置 表写 应然 偏差
    mountain-sop cn L63 1–10 dB 1.5–15 dB 上界低估 5 dB(15→10)
    sigma-endpoint cn L41(引用) 1–10 dB 1.5–15 dB 同上
    sigma-endpoint lab L48(引用) 1–10 dB 1.5–15 dB 同上
    schumann-station lab L37(引用) 1–10 dB 1.5–15 dB 同上

    → 4 条全部”合理区间写法、数值上界低估”。其中 mountain-sop 是源,其余 3 条是引用(R14:引用的不是主张,但引用的错值仍要随源修正)。

    这个错在 elf-skin-depth 审计 §6.1 已独立登记过:「湿花岗岩 10⁻³–10⁻¹ → 1–10 dB,⚠️ 上界低估」——本轮把它的影响面追全了:1 源 + 3 引用 = 4 处。

    为什么值得认真对待:B 档上界 10 dB 写成 10,导致”σ=1e-1 的含盐灰岩 1 km 能挡 15 dB”被低估成 10 dB——低估了高 σ 岩性的屏蔽能力,会影响 #42 高 σ 岩性路径的 K1 门槛判断。

    四、汇总判定表

    文件 条数 类 处置
    building-mountain cn + lab 12 数值错(δ 高估 16–100×) 改正文端点(D9/D10 修正后)
    elf-skin-depth cn + lab 4 合理(自洽 <3%) 无动作
    mountain-sop cn 1 数值错(上界低估) 改正文端点(并入 D1/D2)
    sigma-endpoint cn + lab 2 数值错(引用 mountain-sop) 随源修正
    schumann-station lab 1 数值错(引用 mountain-sop) 随源修正
    合计 20 合理 4 / 数值错 16 / 该单点 0 —

    ⭐ 本轮核心结论:
    1. 20 条全部”该保留区间写法”——触发它们的 σ 全是文献区间,写成区间是物理正确。R13 的”不可单点判”是判读纪律,不是”区间写法有问题”。
    2. 16 条数值错,其中 12 条(building-mountain)是δ 整体高估 16–100×,4 条(B 档)是 dB 上界低估。
    3. 对第 13 轮”方向反了”的修正:building-mountain 方向无误,错在数值。这修正了 D9–D10 的描述。
    4. R13 的盲区被补上:工具把区间标”不可判”后,区间本身的数值对不对就没人看了。本文证明:不可判 ≠ 没问题。这与 R3「未解析 ≠ 无问题」是同一族教训。

    五、证据边界

    • 全部为公式复算 + 库内文本核算,无任何实测。
    • δ = √(2/(ωμσ)) 与 A = 8.686·d/δ 假设均匀半空间、无位移电流、μr/σ 为常数;真实山体分层各向异性且 σ 随频率弥散(Kariya & Shankland 1983 10.1190/1.1441407 / Laštovičková & Bochníček 1992 10.1007/BF01613900 / Revil 2013 10.1029/2012WR012700 / Mehran & Arulanandan 1977 10.1346/CCMN.1977.0250107,4 DOI Crossref 回查通过,均非 PubMed → 无 PMID → E3)。
    • 判定”数值错”仅指”与表格自述 σ/μr 区间不自洽”,不指真实地下 δ 为另一值(若本意是某未写出的等效 σ,问题从”算错”变”没写清 σ”)。
    • 复算口径 f=30 Hz(表头 1–30 Hz 的上端);若作者本意是 7.83 Hz,真值区间为表中值的 1/1.96(σ 同)——无论哪档口径,表写值都显著高估。
    • 本文未改任何已发布正文;修正动作全部登记 D 清单等 JOE 批。

    六、相关阅读

    • 同池 #65 pool4-range-unjudged-20-assertions-adjudication-governance.md(P4-AUDIT-014)
    • building-mountain-terrain-elf-attenuation-analogy.md(被判定 12 条)
    • elf-skin-depth-table-numeric-integrity-audit.md(被判定 4 条,自洽)
    • mountain-sample-priority-sigma-pre-screening-sop.md(B 档源,上界低估)
    • pool4-sigma-endpoint-direction-gate.md(R7/R8 端点判读规则来源)

    七、参考文献

    1. PMID 36676419 — Cheng Y, Materials (Basel) 2023, 低剩磁屏蔽材料测试(工程级 ELF 屏蔽基线)
    2. PMID 28654641 — Binhi VN, Prato FS, PLoS One 2017, 低磁环境生物效应综述(”局部地磁微扰 ≠ 生物体整体低磁”边界)
    3. PMID 38132339 — Sarimov RM, Biology (Basel) 2023, 低磁条件生物效应综述(补充)
    4. PMID 24131394 — Sirav B, Electromagn Biol Med 2014, ELF 变压器磁场生物效应(建筑 ELF 暴露基线)
    5. Kariya & Shankland 1983, 10.1190/1.1441407(岩石电导率频散)
    6. Laštovičková & Bochníček 1992, 10.1007/BF01613900(岩石电导率频散)
    7. Revil 2013, 10.1029/2012WR012700(孔隙介质电导率)
    8. Mehran & Arulanandan 1977, 10.1346/CCMN.1977.0250107(黏土频散)

    稿签:
    – evidence_level: C(公式复算,无实测)
    – review_status: draft
    – 不得直接标 accepted
    – 进入 779 intake 需先做地形 MT 至少 3 处剖面


    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。

    证据与局限

    本文证据等级为 C,属 pure,结论限于研究层面的推断,不构成对个体情形的判定。不可单点判”(R13)说的是判读方式;”区间数值对不对”是另一件事,两者必须分开报。

    局限:本文未覆盖全部相关文献,效应量与结论随样本与方法变化;传统体系与现代研究的对照属结构类比而非等价证明,请结合专业判断参考。

    Segmented Assertions Evaluated Individually: 20 Decisions on Whether to Write as Ranges for RANGE-unjudged Assertions (P4-AUDIT-014)

    Evidence level: C

    Range-Unjudged Item-by-Item Assessment: Deciding Whether to Express as a Range

    One-sentence conclusion: None of the 20 RANGE-unjudged items in the full corpus should be converted to single-point values. This is because the σ (sigma) triggering them originates from documented intervals, and propagating δ/A into an interval is physically correct notation. However, 16 items contain numerical errors: the building-mountain category has 12 δ values overestimated by factors of 16–100×; the B-grade category has 4 dB upper bounds underestimated. Conversely, 4 items are numerically self-consistent (elf-skin-depth categorization thresholds). The statement “Ranges cannot be judged as single points” (R13) refers to the method of interpretation, whereas “Are the numerical values within the range correct?” is a separate question. These two aspects must be reported separately. This document assesses but does not alter: corrections to published text have been assigned to the D-list for JOE approval.

    Item ID Category / Type Issue Description Error Magnitude / Nature Evidence Level
    1–12 building-mountain δ values overestimated 16–100× higher than reference D (numerical error)
    13–16 B-grade dB upper bounds underestimated Lower than expected limits D (numerical error)
    17–20 elf-skin-depth Categorization thresholds Self-consistent with model C (validated)

    Key Distinctions

    • Interpretation vs. Accuracy: The decision to use a range (R13) is about how we read the result, not whether the numbers inside it are right.
    • Scope of Correction: Only 16 items need numerical revision; 4 remain as published.
    • Next Steps: All identified errors are logged for formal review and correction in the official record.

    I. Why This Question Must Be Addressed

    In Round 16 (2026-10-23, #63 tool v1.5), R13 “Interval Assertion Cannot Be Judged by Single Points” was added to delta-table-scan.py, marking 20 interval cells as RANGE-unjudged—neither judged correct nor incorrect. Round 16 immediately logged #65: “Handling of the main text for 12 interval assertions—each must be judged individually on ‘whether it should be written as an interval’.”

    However, the phrasing “whether it should be written as an interval” itself is ambiguous, requiring decomposition into two orthogonal questions:

    Question Meaning Answer
    Q1: Should this cell be written in interval form? Is σ an interval → δ/A is an interval, or should σ itself be a single point? See §III, judged individually
    Q2: Are the interval values correct for this cell? Do the two endpoints of the interval cohere with σ’s interval? See §IV, independently recalculated

    R13 only addressed the “judgment method” of Q1 (intervals cannot be treated as single points), completely failing to address Q2. Yet Q2 represents the true discovery of this round: 16 out of 20 values are numerically incorrect.

    II. Method: Judgment Matrix

    For each of the 20 items, three actions are performed:

    1. Trace σ source: Is the corresponding σ for this cell derived from a literature range or a single point?
    2. Recalculate endpoints: δ = √(2/(ωμσ)); substitute both ends of the σ range to obtain the expected δ interval (A(dB) = 8.686·d/δ follows similarly).
    3. Compare direction: Does the relationship between the interval endpoints and the σ increase align with physical expectations (i.e., larger σ yields smaller δ)?

    Judgments fall into three categories:

    Category Definition Action
    Reasonable σ is a range, and the δ/A interval is internally consistent with the recalculated values (deviation < 30%) Retain range notation; maintain RANGE-unjudged status; no action required
    Numerical error σ is a range, but the δ/A interval endpoints show significant deviation from the recalculated values (> 30%) Retain range notation; correct the interval endpoints (add to D list, etc. JOE)
    Single point σ should have been a single point but was incorrectly written as a range Convert the term to a single point (add to D list, etc. JOE); zero items verified in this round

    III. Item-by-Item Verification (20 items)

    3.1 building-mountain-terrain-elf-attenuation-analogy.md (cn L52–57 + lab L41–46 = 12 items)

    The table header “1–30 Hz skin depth δ (km)” lists σ columns as literature intervals → all written as reasonable intervals. However, endpoint values and recalculated deviations are significant:

    Medium σ interval (S/m) μr Written δ (km) Recalculated δ @30Hz (km) Deviation (low/high end)
    Dry granite 1e-5 – 1e-3 1–10 90–900 2.9–29.1 31× / 31×
    Basalt 1e-4 – 1e-2 1–50 28–280 0.9–9.2 31× / 31×
    Sedimentary rock 1e-3 – 1e-1 1 9–90 0.3–2.9 31× / 31×
    Reinforced concrete 1e-3 – 1e-1 10–100 9–90 0.09–0.9 100× / 100×
    Common red brick 1e-4 – 1e-2 1 28–280 0.9–9.2 31× / 31×
    Steel plate/beam 1e6 – 1e7 1000–10000 0.005–0.05 0.0003–0.004 km 16× / 16×

    Recalculation scope: f=30 Hz, μr takes interval lower bound (most conservative, δ largest). If using 1–30 Hz full band + full μr interval, true range is wider (granite 0.9–159 km); written 90–900 km remains overall overestimated.

    ⭐ Correction to Round 13 conclusion: Round 13 (P4-AUDIT-008 §4 D9–D10) judged building-mountain as “interval arrangement and σ increase same direction → readers will read σ larger δ larger (physically reversed)”. This round’s recalculation overturns that directional claim: written 90–900 km low end small, high end large, consistent with physics (σ small → δ large). The real issue is numerical overall overestimation by ~31× (30Hz scope), not direction reversed. Directional claim is Round 13 misreading “numerical error” as “directional error”. Corrected D9–D10 description: building-mountain §2 table δ interval overall overestimated ~31× (reinforced concrete 100×), direction correct.

    Why worth serious attention: This paper’s thesis is “mountain almost transparent to ELF”. By recalculation (granite δ 2.9–29 km @30Hz) compared to written (90–900 km), it is more transparent—written instead makes mountain appear “less transparent”. Same family as skin-depth-basement 562× error: writing correct conclusion against its own interest. By R2 criterion do not overturn conclusion, but must revise.

    3.2 elf-skin-depth-table-numeric-integrity-audit.md (cn L205–206 + lab L217–218 = 4 items)

    This is revised #22 tier threshold table (§6.2), A′ tier σ∈[1e-2,1e-1] → 1 km A = 4.8–15 dB; B tier σ∈[1e-3,1e-2] → 1.5–4.8 dB. Recalculation:

    Tier σ (S/m) 1 km A(dB) recalc Written Verdict
    A′ 1e-2 – 1e-1 4.83 – 15.3 dB 4.8–15 dB ✅ reasonable (deviation < 3%)
    B 1e-3 – 1e-2 1.53 – 4.83 dB 1.5–4.8 dB ✅ reasonable (deviation < 3%)

    → All 4 items numerically self-consistent, retain interval notation, no action. These 4 items are “tier thresholds” (tier ranges), not measured values—R13’s correct usage lies here.

    3.3 mountain-sample-priority-sigma-pre-screening-sop.md (cn L63 = 1 item) + two places citing it

    B tier “σ 10⁻³–10⁻¹ → 1 km A = 1–10 dB”. Recalculation: σ=1e-3 → 1.53 dB; σ=1e-1 → 15.3 dB. Written 1–10 dB:

    Location Written Should be Deviation
    mountain-sop cn L63 1–10 dB 1.5–15 dB upper bound underestimated 5 dB (15→10)
    sigma-endpoint cn L41 (cited) 1–10 dB 1.5–15 dB same as above
    sigma-endpoint lab L48 (cited) 1–10 dB 1.5–15 dB same as above
    schumann-station lab L37 (cited) 1–10 dB 1.5–15 dB same as above

    → All 4 items “reasonable interval notation, numerical upper bound underestimated”. Among them mountain-sop is source, remaining 3 are citations (R14: cited not claims, but cited wrong values still must follow source correction).

    This error independently logged in elf-skin-depth audit §6.1: “wet granite 10⁻³–10⁻¹ → 1–10 dB, ⚠️ upper bound underestimated”—this round traces its impact scope fully: 1 source + 3 citations = 4 locations.

    Why worth serious attention: B tier upper bound 10 dB written as 10, leads to “σ=1e-1 salt-bearing limestone 1 km can block 15 dB” underestimated to 10 dB—underestimates high σ rock shielding capability, affecting #42 high σ rock path K1 threshold judgment.

    IV. Summary Judgment Table

    File Count Category Action
    building-mountain cn + lab 12 Numerical error (δ overestimated 16–100×) Correct endpoint (post-D9/D10 revision)
    elf-skin-depth cn + lab 4 Reasonable (self-consistent <3%) No action
    mountain-sop cn 1 Numerical error (upper bound underestimated) Correct endpoint (merge into D1/D2)
    sigma-endpoint cn + lab 2 Numerical error (cites mountain-sop) Correct per source
    schumann-station lab 1 Numerical error (cites mountain-sop) Correct per source
    Total 20 Reasonable 4 / Numerical error 16 / Single-point 0 —

    ⭐ Core conclusions for this round:
    1. All 20 entries use “retainable interval notation”—their σ values correspond to literature intervals; writing them as intervals is physically correct. R13’s “cannot judge single-point” is a reading discipline, not an indication that “interval notation is problematic.”
    2. 16 entries show numerical errors, of which 12 (building-mountain) involve δ overestimated by 16–100×, and 4 (Category B) involve dB upper bound underestimated.
    3. Correction for Round 13 “direction reversed”: building-mountain direction is correct; the error lies in the numerical value. This revises the description of D9–D10.
    4. R13 blind spot is filled: after the tool marks intervals as “unjudgeable,” no one checks whether the numerical values of those intervals are correct. This paper demonstrates that “unjudgeable ≠ no problem.” This aligns with the lesson from R3 that “unresolved ≠ no problem.”

    V. Evidence Boundaries

    • All calculations are formula re-computations + in-library text verification; no field measurements were conducted.
    • δ = √(2/(ωμσ)) and A = 8.686·d/δ assume a uniform half-space, no displacement current, and constant μr/σ; real mountain strata are anisotropic with σ dispersing across frequency (Kariya & Shankland 1983 10.1190/1.1441407 / Laštovičková & Bochníček 1992 10.1007/BF01613900 / Revil 2013 10.1029/2012WR012700 / Mehran & Arulanandan 1977 10.1346/CCMN.1977.0250107, 4 DOI Crossref checks passed, none PubMed → no PMID → E3).
    • “Numerical error” refers only to inconsistency with the table’s own stated σ/μr range, not to a different true subsurface δ (if the intent was an unstated equivalent σ, the issue shifts from “miscalculated” to “σ not clearly defined”).
    • Re-computation uses f = 30 Hz (upper end of the table header 1–30 Hz); if the author intended 7.83 Hz, the true range is 1/1.96 of the table values (σ unchanged) — under either convention, the tabulated values significantly overestimate.
    • This paper does not alter any published text; all corrections are logged in the D-list and JOE batch.

    VI. Further Reading

    • Pool4 #65 pool4-range-unjudged-20-assertions-adjudication-governance.md (P4-AUDIT-014)
    • building-mountain-terrain-elf-attenuation-analogy.md (judged 12 assertions)
    • elf-skin-depth-table-numeric-integrity-audit.md (judged 4 assertions, internally consistent)
    • mountain-sample-priority-sigma-pre-screening-sop.md (Tier B source, upper-bound underestimation)
    • pool4-sigma-endpoint-direction-gate.md (source for R7/R8 endpoint adjudication rules)

    VII. References

    1. PMID 36676419 — Cheng Y, Materials (Basel) 2023, testing of low remanence shielding materials (engineering-grade ELF shielding baseline)
    2. PMID 28654641 — Binhi VN, Prato FS, PLoS One 2017, review of low-magnetic environment biological effects (boundary: “local geomagnetic micro-perturbations ≠ whole-body low magnetic state”)
    3. PMID 38132339 — Sarimov RM, Biology (Basel) 2023, review of low-magnetic condition biological effects (supplemental)
    4. PMID 24131394 — Sirav B, Electromagn Biol Med 2014, ELF transformer magnetic field biological effects (building ELF exposure baseline)
    5. Kariya & Shankland 1983, 10.1190/1.1441407 (rock conductivity dispersion)
    6. Laštovičková & Bochníček 1992, 10.1007/BF01613900 (rock conductivity dispersion)
    7. Revil 2013, 10.1029/2012WR012700 (porous medium conductivity)
    8. Mehran & Arulanandan 1977, 10.1346/CCMN.1977.0250107 (clay dispersion)

    Draft Notes:
    – evidence_level: C (formula recalculation, no experimental measurement)
    – review_status: draft
    – do not mark as accepted directly
    – entry into 779 intake requires at least three MT terrain profiles


    Disclaimer: This article constitutes an organization, cross-reference, and research design discussion of traditional knowledge and scientific literature, intended solely for understanding and reference. It does not constitute medical advice and cannot replace professional diagnosis or treatment. The effect sizes, frequency ranges, and exposure conclusions mentioned herein are largely inferences at the research report level, with significant individual variation; if experiencing physical discomfort or health concerns, please seek medical attention and have them evaluated by a qualified physician.

  • 复核不是重扫一遍:审计工具升级后该查的四种东西

    Evidence level: C

    方法学记录 · 池④ #67 · 2026-10-24

    上一轮(第 17 轮)我们修好了审计工具的两个缺陷——R16(判定”谁在主张”从文件名改为文字形态)和 R17(排除规则必须对称施加)。修完工具只是第一步。真正要回答的问题是:修好了,然后呢? 这一轮做的就是这件事——不是再扫一遍就收工,而是按四个角度验证”修复是否真的有效、有没有留下新的坑”。

    一、同一把尺子量两次:修复的稳定性

    判断一个审计工具修复有没有效,最便宜的方法是隔一轮什么都不改,用同一把尺子再量一次。

    第 17 轮修复后,R5X(跨文件自相矛盾检查)报出 6 条违反。本轮在不改任何正文的情况下重跑,数字是:

    • R5X 违反:6 条(一条不多一条不少)
    • 跨文件配对:14 组(不变)
    • 因转述/审计文排除的点:7 个(不变)
    • R5(同文件检查):1 条(不变)
    • selftest:92/92(不变)

    数字完全一样,这本身就是信息。上一轮第 17 轮我们刚抓到过”什么都没改、隔一轮重跑,违反数从 6 涨到 16″——那是工具在指控自己上一轮写的文章。本轮数字纹丝不动,说明第 10 类失败模式(自指污染)没有复发,R16 的修复是真修好了,不是碰巧。

    二、换一个方向再判:6 条违反是不是同一个根

    数字稳定还不够。还得问:这 6 条违反,到底是一个问题还是六个问题?

    本轮逐条核对 R5X 报出的 6 条:4 条的一侧是同一文件同一张表(skin-depth-basement 的 §3 衰减表 L81/L82/L85),第 6 条的另一端也是它。这和第 17 轮的判断完全一致——6 条不是 6 个笔误,是 1 处根错派生的 6 个症状。

    独立手算复现(f=7.83 Hz,μ0=4π×10⁻⁷):

    材料(按文件自述 σ) δ @7.83 Hz 100 m 的 A 1 km 的 A
    花岗岩 σ=1e-4 17.986 km 0.048 dB 0.48 dB
    花岗岩 σ=1e-5 56.88 km 0.015 dB 0.15 dB
    玄武岩(湿)σ=1e-2 1.799 km 0.48 dB 4.84 dB

    对照文件 §2 表声明:花岗岩 δ @7.83 Hz 写的是「32 m」。真值 17.986 km → 差 562×。同一份文件 §3 表把 100 m 花岗岩写成「27 dB」,真值 0.048 dB,也差 562×。两个数字差同一个倍数,不是巧合——它们共用同一个错的 δ。

    为什么这个错值得认真对待:这份文件的立论是「山体对 ELF 几乎透明、靠山屏蔽不成立」。按错值(100 m 27 dB),山体像强屏蔽体;按真值(0.048 dB),几乎不挡。这个错把正确的结论写得在它自己的文件里显得不成立。 更实际的是:δ 低估 562× ⇒ A 高估 562× ⇒ 撑起”普通花岗岩山值得测”,而按 #22 的分档,普通花岗岩是 C 档(最不该优先测)——会把测量预算投到错误样本上。

    三、查工具自己:版本自述与实况一致吗

    第三件事是查工具自己有没有说谎。本轮发现一个很隐蔽的小 bug:

    扫描报告头部自述 delta-table-scan.py v1.5,但文件的字节数和 selftest 数量都指向 v1.6。查源码定位:报告标题行的版本号是硬编码的,v1.6 升级时漏改了(delta-table-scan.py L1243)。

    这看起来只是个展示问题,但它的性质值得警惕:工具自述失真。如果下游有人把”报告标题 v1.5″当成版本依据——尤其当磁盘上同时躺着 v1.5 的备份(bak-20261023-v15)时——就会误判你实际用的是哪一版工具。本轮已把标题改为从 docstring 动态提取最大版本号,改后 selftest 仍 92/92、全库扫描数字与改前完全一致(判定逻辑零改动)。

    方法学增量:审计工具的升级清单里,应该有一条”版本自述同步”检查——升级判定逻辑是功能,升级自述是卫生。两者都要做。

    四、查修复的代价:形态屏蔽有没有误伤

    第四件事最容易被跳过:新规则有没有误伤不该伤的东西。

    R16 把”引号/代码块内的数字”一律不算主张。它会不会把表格里确实在主张的数字也屏蔽掉?本轮核查:

    • 全库表格内引号数字 118 处,绝大多数是 DOI 反引号引用(10.xxxx)与审计文里的对照说明,不构成裸主张,屏蔽正确。
    • 未发现”表格里写 「27 dB」 且确实在主张”的案例。
    • 上轮自指边界示例(pool4-selftest-contamination-single-root-error.md L103)已规避——只描述形状、不给具体值,实测无误报。

    已知盲区(非本轮新增):把数字写进引号但确实在主张的文件,R16 会漏(实测 0 例,理论盲区);R14 排除名单仍按文件名正则(#66 正向化待做)。两者都不阻塞本轮结论。

    证据与局限

    • 证据等级 C(E3 结构类比):全部为解析复算 + 库内文本核算,无任何实测。A ≈ 8.7d/δ 与 δ=√(2/(ωμσ)) 假设均匀半空间、无位移电流、μr/σ 为常数;真实山体分层各向异性且 σ 随频率弥散(Kariya & Shankland 1983 10.1190/1.1441407 / Laštovičková & Bochníček 1992 10.1007/BF01613900 / Revil 2013 10.1029/2012WR012700 / Mehran & Arulanandan 1977 10.1346/CCMN.1977.0250107,4 DOI Crossref 回查通过,均非 PubMed → 无 PMID)。
    • 局限:R5X 判”违反”仅指两文件数字不满足 A ∝ d,不指真实地下 A 为另一值;§1 自述 σ 区间本身未审计;分母极小(35 点 / 14 组配对),不得外推”全库矛盾率”;只覆盖等式断言(R13);屏蔽规则是形态启发式,把数字写进引号但确实在主张的文件会被漏(已知盲区,实测 0 例)。

    下一步(留给 JOE 决策)

    1. #68(P1,须 JOE 批):改 skin-depth-basement §2 表花岗岩 δ 一格(32 m → 17.99 km,5 频点连带)——唯一一项”改一格消 6 条违反”的动作,且直接决定 #22/#23/#25 的采样点名。建议与 D1′ 合并提请(同源:都是”σ/δ 低估导致误以为普通花岗岩值得测”)。
    2. #65(P2,可自主):12 条区间断言逐条判”该不该写成区间”。
    3. #59(P2,可自主):F2 余 4 + F3 余 6 逐格复算。

    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。

    pool4-r16-r17-full-rescan-review

    Evidence level: C

    Methodological record · Pool④ #67 · 2026-10-24

    In the previous round (Round 17), we fixed two defects in the audit tool—R16 (changing “who is making a claim” from filename to textual form) and R17 (ensuring exclusion rules apply symmetrically). Fixing the tool is only the first step. The real question to answer is: Now that it’s fixed, what next? This round addresses exactly that—not simply re-running the scan and calling it done, but verifying from four angles whether “the fix is truly effective and no new pitfalls have been left behind.”

    I. Measuring twice with the same ruler: stability of the fix

    The cheapest way to judge whether an audit tool fix is effective is to run it again without any changes in the next round, using the same ruler.

    After Round 17’s fix, R5X (cross-file self-contradiction check) reported 6 violations. Running the same scan again without modifying any text, the numbers are:

    • R5X violations: 6 (one more, one less—exactly the same)
    • Cross-file pairs: 14 (unchanged)
    • Points excluded due to paraphrase/audit text: 7 (unchanged)
    • R5 (same-file check): 1 (unchanged)
    • selftest: 92/92 (unchanged)

    The numbers are exactly the same, which itself is information. In Round 17, we just caught “nothing changed, re-run next round, violation count rose from 6 to 16″—that was the tool accusing itself of its own previous round’s articles. This round’s numbers remain unchanged, indicating that failure mode #10 (self-reference contamination) has not recurred, and R16’s fix is truly fixed, not coincidental.

    II. Checking from another direction: are the 6 violations all rooted in one problem

    Stable numbers aren’t enough. We must also ask: are these 6 violations all one problem or six separate ones?

    This round, we checked each of the 6 violations reported by R5X individually: 4 violations have one side in the same file and same table (skin-depth-basement §3 attenuation table L81/L82/L85), and the other end of violation #6 is also in it. This matches exactly the judgment from Round 17—these 6 violations are not 6 typos, but 6 symptoms stemming from 1 root error.

    Independent manual recalculation (f=7.83 Hz, μ0=4π×10⁻⁷):

    Material (as stated in file σ) δ @7.83 Hz A at 100 m A at 1 km
    Granite σ=1e-4 17.986 km 0.048 dB 0.48 dB
    Granite σ=1e-5 56.88 km 0.015 dB 0.15 dB
    Basalt (wet) σ=1e-2 1.799 km 0.48 dB 4.84 dB

    Comparing with the file’s §2 table declaration: granite δ @7.83 Hz is stated as “32 m“. True value 17.986 km → off by 562×. In the same file §3 table, 100 m granite is stated as “27 dB“, true value 0.048 dB, also off by 562×. Both numbers differ by the same factor, not coincidence—they share the same wrong δ.

    Why this error matters: This file’s thesis is “mountains are nearly transparent to ELF, shielding by mountains doesn’t hold.” Based on the wrong value (100 m, 27 dB), mountains act like strong shields; based on the true value (0.048 dB), they barely block anything. This error makes the correct conclusion appear invalid within its own file. More practically: δ underestimated by 562× ⇒ A overestimated by 562× ⇒ supporting “ordinary granite mountains worth measuring,” while according to #22’s classification, ordinary granite is Category C (least priority)—this would misallocate measurement budget to wrong samples.

    III. Checking the tool itself: does the version statement match reality?

    The third thing to check is whether the tool itself is lying. This round found a subtle bug:

    The scan report header states delta-table-scan.py v1.5, but the file’s byte count and selftest numbers both point to v1.6. Tracing the source code: the version number in the report title line is hard-coded, and v1.6 upgrade missed updating it (delta-table-scan.py L1243).

    This looks like just a display issue, but its nature is worth caution: tool self-statement is distorted. If downstream users take “report title v1.5” as the version basis—especially when both v1.5 backup (bak-20261023-v15) and v1.6 files coexist on disk—they may misjudge which version of the tool is actually in use. This round has changed the title to dynamically extract the maximum version from docstring; after change, selftest remains 92/92, full-library scan numbers match before modification exactly (judgment logic unchanged).

    Methodological increment: In the audit tool upgrade checklist, there should be an item “version statement synchronization”—upgrade judgment logic is functionality, upgrade statement is hygiene. Both must be done.

    IV. Checking the cost of the fix: did new rules wrongly shield things they shouldn’t?

    The fourth thing is most easily skipped: did the new rules wrongly shield things they shouldn’t have.

    R16 excludes “numbers inside quotes/code blocks” from being claims. Could it also shield numbers in tables that are indeed making claims? This round checked:

    • 118 instances of quoted numbers in tables across the whole library, most are DOI backtick references (10.xxxx) and audit text cross-references, not bare claims, so shielding is correct.
    • No cases found where “table says 「27 dB」 and it is indeed making a claim.”
    • Previous self-reference boundary example (pool4-selftest-contamination-single-root-error.md L103) has been avoided—only describing shape, not giving specific values, tested with no false alarms.

    Known blind spots (not added this round): files that write numbers into quotes but are indeed making claims, R16 will miss them (tested 0 cases, theoretical blind spot); R14 exclusion list still uses filename regex (#66 forwardization pending). Neither blocks this round’s conclusion.

    Evidence and limitations

    • Evidence level C (E3 structural analogy): All are parsed recalculation + library text calculation, no actual measurement. A ≈ 8.7d/δ and δ=√(2/(ωμσ)) assume uniform half-space, no displacement current, μr/σ constant; real mountain layers are anisotropic and σ disperses with frequency (Kariya & Shankland 1983 10.1190/1.1441407 / Laštovičková & Bochníček 1992 10.1007/BF01613900 / Revil 2013 10.1029/2012WR012700 / Mehran & Arulanandan 1977 10.1346/CCMN.1977.0250107, 4 DOI Crossref lookup passed, none are PubMed → no PMID).
    • Limitations: R5X judging “violation” only means two files’ numbers don’t satisfy A ∝ d, not that true underground A is another value; §1 stated σ interval itself was not audited; denominator extremely small (35 points / 14 pairs), cannot extrapolate “whole-library contradiction rate“; only covers equation assertions (R13); shielding rule is morphological heuristic, files that write numbers into quotes but are indeed making claims will be missed (known blind spot, tested 0 cases).

    Next steps (for JOE decision)

    1. #68 (P1, requires JOE approval): Fix skin-depth-basement §2 table granite δ by one cell (32 m → 17.99 km, 5 frequency points linked)—the only action of “fixing one cell to eliminate 6 violations”, and directly determines #22/#23/#25 sampling names. Suggested to merge with D1′ submission (same origin: both “σ/δ underestimated leads to thinking ordinary granite worth measuring”).
    2. #65 (P2, can be done independently): Judge each of 12 interval assertions “should it be written as an interval.”
    3. #59 (P2, can be done independently): Recalculate F2 remainder 4 + F3 remainder 6 cell by cell.

    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.