Author: joe

  • 那个比值的误差根本没被放大:误差翻倍的是你以为自己在比较的东西,不是比值本身

    Evidence level: C

    我们把怕的东西搞错了整整一轮。比值 Δ/D_min 的抽样误差不但没被放大,反而被系统性压小;真正会翻转判决的,是分母里那个你从来没报过的固定项 B。

    正文

    一、先说结论:怕反了

    上一轮我给出的操作规则是「先算比值 R = Δ 除以 D_min,落在 0.7 到 2.5 之间才谈打折」。我还留了一条待验证:D_min 里的 δ 和 ρ 都是估的,两个估计量一除,误差会怎么走。

    现在算完了,结论跟预想的相反。比值不但没有把误差翻倍,它的相对误差永远比分子自己更小,而且有一个干净的上界。

    设 R = Δ /(B + 1.96·u·SE),其中 u 是根号下 2 减 2ρ,B 是边际带宽这个固定项。做一次 delta 方法展开,R 的相对标准误恰好是

    相对 SE = 1 / (1.96 + B/(u·SE))
    

    这是本文最要紧的一条。分母里那个 1.96 是拍下来的 95% 系数,B/(u·SE) 恒为正。所以相对 SE 永远小于 1/1.96 = 0.510。比值做不出一条 50% 以上的相对误差带。所谓「一除就把噪声翻倍」在代数上不成立。

    n 等于 60、m 等于 0.85 时实测 0.482,n 等于 600 时降到 0.431,n 等于 2600 时 0.369。从 60 例加到 2600 例,四十三倍的样本,相对精度只从 48% 改善到 37%。

    二、为什么压不下去:结构项在当家

    看那个式子的分母。B/(u·SE) 里,SE 随根号 n 缩小,B 是常数。样本量一大,B 项就接管了分母。

    m 等于 0.85 时 B 是 0.0166,n 等于 60 时 B/(u·SE) 只有 0.114,B 在 D_min 里占 5.5%;n 等于 2600 时占 27.7%;n 等于 50000 时占 62.7%。B 从配角变成主角。

    这解释了一件反直觉的事:比值的精度不是「样本够就准」,而是被一个跟样本量无关的结构常数锁住。要让半宽降到 0.45,ρ 等于 0.5 时需要约两万六千例;降到 0.30 要九万六千例。而整个判读文献的样本量中位数只有 71 —— 那是从 2022 年 82 项一致性研究的综述里读出来的数。

    三、判决其实只有一个旋钮,它叫 λ

    把 R 的式子重排一下,会浮出一个恒等式:

    Δ / D_min = z / λ,其中 z = Δ/(u·SE),λ = 1.96 + B/(u·SE)

    我用四组 n 乘三组 ρ 共十二个点验算过,R 与 z 除以 λ 的差最大 8.9×10⁻¹⁶,纯浮点。

    这个恒等式的后果比它的形式重要得多。它说明「R 在 0.7 到 2.5 之间」这件事完全等价于「z 在 0.7λ 到 2.5λ 之间」。 而 λ 只由 (B, SE, ρ) 决定,与你要判的那个差 Δ 一点关系都没有。

    于是那条判决带,在显著性尺度上是一条会滑动的窗口。n 等于 60 时窗口是 z 落在 1.45 到 5.19;n 等于 600 时 1.62 到 5.80;n 等于 10000 时 2.40 到 8.58。同一个真差,样本量一大,就不再落进「可打折」这一档。

    具体点:m 等于 0.85、ρ 等于 0.5、真差 0.15。n 等于 60 判「太细、别打」,n 等于 125 翻成「可打折」,一路到 n 等于 2589 还在这一档,n 等于 5000 再翻成「不必打折」。三段判决、同一个真差、只换样本量。

    两个研究各做各的、真差一模一样,判决不一致的概率:n 取 60 对 600 是 64.9%,60 对 2000 是 75.7%,200 对 10000 是 100%。

    四、还有一条更硬的天花板

    R 的上界是 Δ 除以 B。因为 1.96·u·SE 恒为正,R 永远小于 Δ/B。

    m 等于 0.85 时 B 是 0.0166,所以真差 0.02 的 R 上限只有 1.205、真差 0.03 上限 1.807、真差 0.04 上限 2.410 —— 这三种情况无论做多少人、ρ 多高,永远进不了「不必打折」那一档。规则给了一档,数据里根本没有那一档。

    反过来说,真差 0.05 就能到 2.5,0.15 能到 9.0。2.5 这条上界对小于 0.0415 的真差是一句空话。

    这里要补一句方法学上的旧账:比值型统计量的区间从来是个坑,不是本篇才发现的。Fieller 定理的名字就来自这件事——它的区间有三种形态,闭区间、闭区间的补集、以及整条实轴,具体是哪一种取决于分母能不能显著拒绝零。1999 年一项一万次模拟的对照研究直接说,当效果差接近零时,Fieller 与三种 bootstrap 全都给不出可用结果。工具变量文献里的说法更直白:比值估计量的均值不存在有限值,因为分母有非零概率贴着零,样本一多均值就发散。

    本篇的分母好在不接近零——B 与 1.96·u·SE 都是正数之和,D_min 恒有下界 B。这正是它比典型比值估计安全的地方。 反过来说,一旦有人把分母换成不带宽带底的项目(比如只用 SE),那套无界区间的麻烦立刻回来。

    五、ρ 的误差:比想象的小得多

    原假设担心 δ 与 ρ 正相关会「系统性互相放大」。实测:n 等于 60 时相关系数从 负 0.5 走到 正 0.5,半宽从 0.930 变成 1.029,只差 10.6%;n 等于 600 时差 8.9%。

    原因在方差分解。R 等于 1 处,ρ 项对总方差的贡献 n 等于 60 时是 3.6%、n 等于 600 时 0.36%、n 等于 2600 时 0.08%。ρ 估不准这件事,在比值这个用途上几乎可以忽略。 跟上一轮 #1410 的结论对上了:ρ 的抽样误差在那条判据里也从来不是瓶颈。

    六、真正该改的三处

    第一,判决必须带 λ,不能只报 R。λ 一个数就把 (B, SE, ρ) 三样东西压进去了,而它是唯一决定判决带位置的东西。报告里不写 λ,等于不写这条规则的适用条件。

    第二,0.7 与 2.5 这两条线是按「R 的尺度」画的,而 R 的尺度本身随 n 滑动。正确写法是把它翻译成 z 尺度:z 要落在 0.7λ 到 2.5λ 之间。原规则在 n 从 60 变到 10000 的过程里,实际要求的 z 下界从 1.45 涨到 2.40 —— 严了六成半,而没人知道这件事。

    第三,n 小的时候不是「区间宽一点」的问题,是「判决本身不存在」的问题。n 等于 60 时半宽 0.945、区间总宽 1.89,比整条判决带(宽 1.8)还宽 —— 区间把三档全盖住了。这种情况下给一个「可打折」或「太细」的结论,是在报噪声。

    顺带说一句 n 该怎么规划。既然相对精度等于 1.96 除以 z,那「区间总宽不超过判决带宽」等价于半宽不超过 0.9,m 等于 0.85 时需约 219 例。但那只保证区间窄于判决带,不保证能分清三档。 要半宽压到 0.45,ρ 等于 0.5 时需两万六千例。这两个数差两个量级,文献里的实际做法是把它当第一个数在用。

    证据与局限

    已验证:① 相对 SE 等于 1 除以(1.96 加 B/(u·SE))这条式子在四组 n 乘三组 ρ 上逐点吻合,上界 0.510 是代数结论;② R 恒等于 z 除以 λ 十二个格点最大偏差 8.9×10⁻¹⁶;③ R 的 95% 区间上限是 Δ 除以 B,与 n、ρ 无关;④ ρ 项对总方差的贡献在各档均低于 6%;⑤ Δ 与 ρ 的相关性在正负 0.5 范围内对半宽的影响不超过 11%(二十万次蒙特卡洛)。

    部分验证:① SE 的闭式沿用前作 Delta 方法口径,边际相等假设未在本篇重检;② B 取自 m 等于 0.60、0.85、0.90、0.95 四点的插值,中间值有插值误差;③ 判决带 0.7 与 2.5 本身是上一轮拍的,本篇只算它的误差、不辩护它的取值。

    待验证:① ρ̂ 的抽样标准差用了 (1 减 ρ 平方) 除以根号 n 的近似,小样本下未必准;② 本篇假定 Δ 与 ρ 独立可分解,若两者真同源(即 #1436 那个问题),方差结构要重算;③ 比值形式在分母可能接近零时 Fieller 型区间会变成无界集,本文的线性展开在 B/(u·SE) 很小时可能低估尾部;④ 三档之外的有序数据未涉及。

    L1 网络状态:L1 对应网络中未找到与本主题(比值估计的误差传播与判决带稳定性)对应的边。两个源文件内检索「比值 / 置信区间 / 误差 / 阈值概率 / 决策曲线 / 净收益 / 判读 / 一致性」均无命中。唯一结构相关的是 §3.8 至 §3.11 反复出现的「E-060 标签前提」,状态为 conflict(未裁决):「中医医师间诊断一致性在开放人群中低(low),作者称为令人担忧(PMID 30383398);同系列第二篇指出偶然校正一致性统计常被误用、建议 Gwet AC2(PMID 31730403、31634000、31730404);但训练可改善(PMID 19538100)、量表化可行」。冲突双方在本篇的意义是:这条测量链的输入端本身未被裁决,本篇讨论的比值再准,也只是在一个尚未定论的量表上做精度分析。该冲突不构成对本篇数学结论的否证,但限定了它的适用范围。

    相关阅读

    证据等级: C
    参考文献: PMID 32865227, PMID 10874387, PMID 26282889, PMID 11512137, PMID 22560852, PMID 11890316, PMID 36123642


    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 error of the ratio itself was never amplified: what doubled in error was what you thought you were comparing, not the ratio itself.

    Evidence level: C

    We got the thing we feared wrong by exactly one full cycle. The sampling error of the ratio Δ/D_min was not only not amplified; it was systematically suppressed. What actually flips the verdict is the fixed term B in the denominator, which you never reported.

    Body

    I. Conclusion First: The Fear Was Reversed

    In the previous round, my operational rule was “first calculate the ratio R = Δ divided by D_min, and only discuss discounting if it falls between 0.7 and 2.5.” I also left one item to verify: since δ and ρ inside D_min are both estimates, dividing two estimators will affect the error in what way.

    Now that the calculation is done, the conclusion is opposite to expectation. The ratio did not double the error; its relative error is always smaller than that of the numerator itself, and it has a clean upper bound.

    Let R = Δ /(B + 1.96·u·SE), where u is the square root of 2 minus 2ρ, B is the fixed term for marginal bandwidth. Expanding via the delta method, R’s relative standard error is exactly

    Relative SE = 1 / (1.96 + B/(u·SE))
    

    This is the most critical equation in this paper. The 1.96 in the denominator is the captured 95% coefficient; B/(u·SE) is always positive. Thus, the relative SE is always less than 1/1.96 = 0.510. The ratio cannot produce a relative error band exceeding 50%. The claim that “dividing doubles the noise” fails algebraically.

    When n equals 60 and m equals 0.85, the measured value is 0.482; when n equals 600 it drops to 0.431; when n equals 2600 it is 0.369. Increasing the sample from 60 cases to 2,600 cases—a forty-three-fold increase—improves relative precision only from 48% to 37%.

    II. Why It Cannot Be Suppressed: Structural Terms Dominate

    Look at the denominator of that equation. In B/(u·SE), SE shrinks with the square root of n, while B is constant. As sample size grows, the B term takes over the denominator.

    When m equals 0.85, B is 0.0166; when n equals 60, B/(u·SE) is only 0.114, meaning B accounts for 5.5% of D_min; when n equals 2600 it accounts for 27.7%; when n equals 50,000 it accounts for 62.7%. B shifts from supporting role to lead.

    This explains a counterintuitive fact: the precision of the ratio is not “accurate when sample size is sufficient,” but locked by a structural constant independent of sample size. To bring the half-width down to 0.45, ρ must equal 0.5, requiring approximately 26,000 cases; to reach 0.30 it requires 96,000 cases. Yet the median sample size across all judgment literature is only 71 — a number read from a 2022 review of 82 consistent studies.

    III. The Verdict Actually Has Only One Knob, Called λ

    Rearranging the equation for R reveals an identity:

    Δ / D_min = z / λ, where z = Δ/(u·SE), λ = 1.96 + B/(u·SE)

    I verified this across four groups of n times three groups of ρ, totaling twelve points; the maximum difference between R and z divided by λ is 8.9×10⁻¹⁶, pure floating-point noise.

    The consequence of this identity is far more important than its form. It shows that “R falls between 0.7 and 2.5” is completely equivalent to “z falls between 0.7λ and 2.5λ.” Since λ depends only on (B, SE, ρ), it has nothing to do with the difference Δ you are judging.

    Thus, that judgment band is a sliding window on the significance scale. When n equals 60, the window is z falling between 1.45 and 5.19; when n equals 600, between 1.62 and 5.80; when n equals 10,000, between 2.40 and 8.58. The same true difference, with larger sample size, no longer falls into the “discountable” category.

    Specifically: m equals 0.85, ρ equals 0.5, true difference 0.15. When n equals 60 it is judged “too fine, do not discount”; when n equals 125 it flips to “discountable”; continuing up to n equals 2,589 it remains in this category; when n equals 5,000 it flips again to “no need to discount.” Three verdicts, the same true difference, changing only sample size.

    If two studies proceed independently with identical true differences, the probability of inconsistent verdicts: n taking 60 versus 600 is 64.9%, 60 versus 2,000 is 75.7%, 200 versus 10,000 is 100%.

    IV. Another Harder Ceiling

    R’s upper bound is Δ divided by B. Since 1.96·u·SE is always positive, R is always less than Δ/B.

    When m equals 0.85, B is 0.0166, so for true differences of 0.02, 0.03, and 0.04, the R upper limits are only 1.205, 1.807, and 2.410 respectively — in no case do these enter the “no need to discount” category, regardless of how many people are tested or how high ρ is. The rule provides one tier, but the data lacks that tier entirely.

    Conversely, a true difference of 0.05 can reach 2.5, and 0.15 can reach 9.0. The upper bound of 2.5 is an empty statement for true differences less than 0.0415.

    Here, let me add a methodological debt: ratio-type statistics intervals have always been a pitfall, not something discovered in this paper. The name Fieller’s theorem comes precisely from this issue — its interval has three forms: closed interval, complement of a closed interval, and the entire real axis, depending on whether the denominator can significantly reject zero. A 1999 study with 10,000 simulations directly stated that when the effect difference approaches zero, neither Fieller nor three bootstrap methods yield usable results. The statement in instrument variable literature is even clearer: the mean of the ratio estimator does not have a finite value, because the denominator has non-zero probability of touching zero, causing the mean to diverge as sample size increases.

    The denominator in this paper is fortunate in not approaching zero — B and 1.96·u·SE are sums of positive numbers, so D_min always has a lower bound B. This is exactly where it is safer than typical ratio estimators. Conversely, once someone replaces the denominator with items without bandwidth bottoms (e.g., using only SE), the trouble of unbounded intervals immediately returns.

    V. ρ’s Error: Much Smaller Than Imagined

    The original hypothesis worried that positive correlation between δ and ρ would “systematically amplify each other.” Measured results: when n equals 60, the correlation coefficient moves from negative 0.5 to positive 0.5, half-width shifts from 0.930 to 1.029, a difference of only 10.6%; when n equals 600 the difference is 8.9%.

    The reason lies in variance decomposition. At R equals 1, the ρ term’s contribution to total variance is 3.6% when n equals 60, 0.36% when n equals 600, and 0.08% when n equals 2,600. The uncertainty in estimating ρ is almost negligible for this ratio application. This matches the conclusion from previous #1410: ρ’s sampling error was never the bottleneck in that judgment criterion.

    VI. Three Places That Truly Need Revision

    First, the verdict must include λ, not just report R. A single number λ compresses (B, SE, ρ), and it is the only factor determining the position of the judgment band. Omitting λ from the report equals omitting the applicability conditions of this rule.

    Second, the lines at 0.7 and 2.5 are drawn according to “R’s scale,” but R’s scale itself slides with n. The correct way is to translate it into z scale: z must fall between 0.7λ and 2.5λ. The original rule, as n increases from 60 to 10,000, actually requires the z lower bound to rise from 1.45 to 2.40 — a tightening of sixty-five percent, yet no one knew this.

    Third, when n is small, the problem is not “the interval is wider,” but “the verdict itself does not exist.” When n equals 60, the half-width is 0.945 and the total interval width is 1.89, which is wider than the entire judgment band (width 1.8) — the interval covers all three tiers. Under such circumstances, giving a verdict of “discountable” or “too fine” is merely reporting noise.

    Incidentally, let me say how n should be planned. Since relative precision equals 1.96 divided by z, “total interval width not exceeding judgment band width” is equivalent to “half-width not exceeding 0.9,” requiring approximately 219 cases when m equals 0.85. But this only may support the interval is narrower than the judgment band, not that it can distinguish the three tiers. To press the half-width down to 0.45, ρ must equal 0.5, requiring 26,000 cases. These two numbers differ by two orders of magnitude; actual practice in the literature treats it as the first number.

    Evidence and Limitations

    Verified: ① The equation “Relative SE equals 1 divided by (1.96 plus B/(u·SE))” matches point-by-point across four groups of n times three groups of ρ; the upper bound 0.510 is an algebraic conclusion; ② R always equals z divided by λ, with maximum deviation of 8.9×10⁻¹⁶ across twelve grid points; ③ R’s 95% interval upper bound is Δ divided by B, independent of n and ρ; ④ the ρ term’s contribution to total variance is below 6% in all tiers; ⑤ the influence of correlation between Δ and ρ on half-width, within ±0.5 range, does not exceed 11% (200,000 Monte Carlo simulations).

    Partially Verified: ① The closed-form for SE follows the Delta method convention of prior work; the assumption of equal margins was not rechecked in this paper; ② B is interpolated from four points with m equal to 0.60, 0.85, 0.90, and 0.95, introducing interpolation error in intermediate values; ③ the judgment bands of 0.7 and 2.5 themselves were set in the previous round; this paper only calculates their error, not defends their values.

    To Be Verified: ① the sampling standard deviation of ρ̂ uses the approximation (1 minus ρ squared) divided by square root of n, which may not be accurate for small samples; ② this paper assumes Δ and ρ are independent and decomposable; if they truly share the same source (i.e., the issue in #1436), the variance structure must be recalculated; ③ when the denominator in ratio form approaches zero, Fieller-type intervals become unbounded sets; this paper’s linear expansion may underestimate tails when B/(u·SE) is very small; ④ ordered data outside the three tiers were not involved.

    L1 Network Status: L1 corresponds to edges not found in the network matching this topic (error propagation of ratio estimation and stability of judgment bands). Searching “ratio / confidence interval / error / threshold probability / decision curve / net gain / judgment / consistency” in both source files yielded no hits. The only structurally related element is the recurring “E-060 label premise” in §3.8 to §3.11, status conflict (unresolved): “Diagnostic consistency among TCM physicians is low in open populations (low), authors call this worrying (PMID 30383398); the second paper in the same series notes that chance-corrected agreement statistics are often misused, recommending Gwet AC2 (PMID 31730403, 31634000, 31730404); but training can improve (PMID 19538100), and scaling is feasible.” The significance of this conflict for this paper is: the input end of this measurement chain itself remains undecided; the ratio precision discussed here is merely an accuracy analysis on a scale not yet finalized. This conflict does not refute the mathematical conclusions of this paper, but limits its scope of application.

    Related Reading

    Evidence Level: C
    References: PMID 32865227, PMID 10874387, PMID 26282889, PMID 11512137, PMID 22560852, PMID 11890316, PMID 36123642


    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.

  • 给尺子加刻度是对的,但校准的人只抽了十分之一:锚本身的误差只占一成,真正吃掉精度的是「拿少数人代表全部人」

    Evidence level: C

    上一篇说「加一个自带刻度的外部锚,秩就补齐」。这在代数上无懈可击,在统计上只答了一半:锚自己带误差。但把账算清之后,最要紧的一条并不是「锚太差会拖垮校正」——锚误差在总代价里只占一到两成,真正吃掉精度的是另一件事:校准只抽了一部分人,那部分人对全体的代表性才是主项。

    正文

    上一篇(#1403)的结论是:R 位读者只给出 R 个观测,未知量却是 R+1 个,亏的秩恒为 1,加读者补不上;只有引入一个自带单位的物理锚(皮温的摄氏度数、握力的牛顿值),把均值钉住,秩才补齐。

    这是线性代数的直接推论。但它只保证「有唯一解」,不保证「解误差小」。把锚真的接进模型之后,有两件事必须分开算,而它们相差一个数量级。

    一、好消息:两人之差里,锚的误差精确抵消

    设读者读数 z = x + b + ε,锚 Y = k·x + η。有锚之后,每位读者的位置估计是 b̂ = z̄ − Ȳ/k。两人之差是两个 b̂ 相减,Ȳ/k 这一项在减法里被完整抵掉。

    实测(n=240、4 位读者、每人重复 2 次、读者噪声 0.5):锚误差从 0 加到 0.3、1.0、3.0,两人之差的抽样标准差始终停在 0.031 至 0.034 之间,逐样本差的绝对值上界是 5.55e-17——不是「小到可以忽略」,是浮点运算的舍入底噪。

    加锚对「差」这件事既没有精度代价、也没有精度收益。 那个 0.032 完全由读者噪声和病例数决定,锚误差对此一无所知。

    二、绝对值那一侧:锚误差确实是全额落账的,但它只是第二项

    同一次模拟里,单个读者位置估计的标准差从无锚时的 0.0223,随锚误差 0.354 涨到 0.0325(+46%),0.6 时 0.0445(翻倍),1.0 时 0.0678(三倍)。

    方差可以拆成两项相加:读者项 se²/(mn) = 0.000521,锚项 s²/n。两项相等处是 s* = k·se/√m = 0.354。低于这个值锚误差基本白给,高于这个值就变成「用一把更差的尺子标定」。

    但请注意这条曲线的形状:s 从 0 涨到 1.0,锚项从 0 变成 0.00417,是读者项的 8 倍——SD 从 0.0223 到 0.0678,只涨 3 倍。误差是平方相加的,最差的那一项只能占主导,不能叠加成灾。

    三、真正的主项:校准只抽了一部分人

    真实做法很少给全样本都上锚,通常抽一部分人做校准。这时最要紧的不是锚有多准,而是「这 n_cal 个人能不能代表全体 n 个人」。

    把两件事拆开验证(n=240、se=0.5、m=2、锚误差 s=0.5):

    • 所有人共用同一批校准对象(所有读者的 b̂ 都减掉同一个 Ȳ_cal):校准 10 人时差的标准差 0.99 倍,30 人 1.00 倍,60 人 1.01 倍。完全不动。 因为差值把共同项又减掉了一次。
    • 每位读者各抽各的校准对象(跨中心、跨设备时这是常态):校准 10 人时 15.2 倍,30 人 8.7 倍,60 人 5.7 倍,120 人 3.6 倍,180 人 2.5 倍。

    同一个锚、同一份锚误差,「共用」与「各抽」在 10 人校准点上差 15 倍。 这不是锚准不准的问题,是校准对象的身份问题。

    四、而且这个代价几乎全是抽样误差,不是锚误差

    把锚换成完美(s=0)再算一遍,膨胀倍数:

    • 校准 10 人:13.5 倍(s=0.5 时 15.2 倍)
    • 校准 30 人:7.6 倍(8.7 倍)
    • 校准 60 人:5.0 倍(5.7 倍)
    • 校准 120 人:3.0 倍(3.6 倍)

    即便用的是零误差的完美锚,校准 10 人也要付出 13.5 倍的精度代价。 锚误差从 0 加到 0.5,在 10 人校准点上只把代价从 13.5 推到 15.2,净增 12%;在 60 人点上是 5.00 到 5.74,净增 15%;只有在接近全样本(240 人)时,锚误差才成为主项(1.00 到 1.73,净增 73%)——而那时抽样误差已经归零,剩下的自然全是锚的账。

    这是本轮最实用的一条判据:校准人数不到样本的 60%,讨论「锚要多准」是抓错了主要矛盾;该讨论的是「校准要抽多少人」。

    五、解析式:膨胀的平方等于 1 加上两个代价之比

    上面所有数字都能用一个式子复现:

    膨胀² = 1 + [ (1/n_cal − 1/n) + s²/n_cal ] ÷ (se²/(mn))

    方括号里第一项是「用 n_cal 人估计全体均值」的抽样误差,第二项是锚自身的误差,分母是基线(不标定时差值的方差)。六组参数核到小数点后两位,最大偏差 0.07。

    反解:在 n=240、se=0.5、m=2 的条件下,想把膨胀控制在 2 倍以内,锚完美时需要校准 175 人(占样本 73%),锚误差 0.5 时需要 218 人(91%)。想让膨胀不超过 1.5 倍,锚误差 0.5 时无解——哪怕全校准也救不回来,因为那时 s > s*,锚本身就成了最差那一项。

    换句话说:所谓「抽 10% 做校准」在文献里是常规做法,在这套设计下等于把精度砍掉 15 倍。 而系统综述早就写过这条线:内部验证子样本的抽样策略研究表明,即便抽到主样本的 40%,分层随机与极值抽样也只比纯随机好 10% 到 12%,而且这点效率增益要以更高的偏倚百分比和更低的覆盖率为代价——效率不是免费的,而「抽多少」这件事比「怎么抽」重要得多。 另有方法学研究明确指出,参考测量未必是无偏金标准,其误差与被测仪器的误差可能相关,这种相关会以不可预测的方式高估或低估校正系数。

    六、共模误差:这部分谁救不了

    若锚误差里有一部分是全批共享的(同一台仪器、同一批试剂、同一位资深者按曲线判断),这部分不随 n 缩小,会留一个地板。

    实测地板 0.13 时,样本量 60、240、960 三档的标准差是 0.1379、0.1292、0.1297——样本量涨 16 倍,只降了 5%。

    加读者同样救不了。地板 0.13 下,读者从 2 位加到 16 位,绝对位置的标准差从 0.1303 只到 0.1442(反而略升,因为每人病例数被摊薄)。而两人之差那一侧纹丝不动:0.0325、0.0462、0.0643、0.0930——它只受病例数影响,完全不受共模误差影响。

    这就是加锚这件事的完整账:对「差」无影响(既不加分也不减分),对「绝对值」有影响但只占一到两成,而真正的主项是校准子样本的代表性。

    证据与局限

    已验证: 共锚时两人之差的抽样标准差与锚误差完全无关(锚误差 0 至 3.0,标准差恒在 0.031 至 0.034,逐样本差的绝对值上界 5.55e-17);单个读者绝对位置的方差可分解为读者项 se²/(mn) 与锚项 s²/n,两项相等处 s*=k·se/√m=0.354,七组参数实测与解析最大偏差 0.0007;独锚且各读者各抽校准子样本时,差值的膨胀平方等于 1 加「抽样项 (1/n_cal−1/n) 与锚项 s²/n_cal 之和」除以基线方差,六组参数实测与解析最大偏差 0.07;共模锚误差的地板不随样本量下降(地板 0.13 时三档样本量标准差 0.1379/0.1292/0.1297);共模误差下加读者无效(2 位到 16 位绝对位置 SD 0.1303 至 0.1442,而两人差 SD 只随病例数变化)。

    部分验证: 「锚误差只占总代价一到两成」这个结论是在 n=240、se=0.5、m=2、s=0.5 这一组参数下算的,参数换成小样本或高噪声时占比会上升(解析式给出占比等于 (s²/n_cal) ÷ [(1/n_cal−1/n)+s²/n_cal],n_cal 越接近 n 占比越高);「共模误差不随样本量缩小」是按「批内共同偏移 + 个体随机项」的简化两层结构模拟的,真实共享误差可能有中心—设备—批次三层,三层各自的方差占比未测;「共用同一批校准对象」和「各抽各的」是两种极端,中间态(部分重叠)的膨胀曲线未算。

    待验证: ① 三位以上读者加外部锚的完整设计中,锚误差与读者噪声若正相关(锚也由人判读),有效独立精度会按 s/√(1−ρ²) 放大,而 ρ 本身从未在判读数据里测过;② 「以资深者按曲线判读为锚」这类退化情形下,锚与读者的零点是否共享同一个任意起点,未检验;③ 5pp 当量阈值在共模锚误差存在时是否仍可执行,本篇只给了地板值,没给判据失效的临界地板。

    相关阅读

    证据等级: C
    参考文献: PMID 25627982, PMID 33878166, PMID 8793367, PMID 22468848, PMID 30402814, PMID 37704705, PMID 30383398, PMID 19538100
    L1 网络状态:CORRESPONDENCE_NETWORK.md 第 202 行「E-060 标签前提:五脏辨证(作为分组标签)→ 标签信度」原记为 conflict(未裁决)——「中医医师间诊断一致性在开放人群中『低』(PMID 30383398)」与「训练可改善(PMID 19538100)」两方并列;第 223 行 KB-317 追加两处错配(样本量单位错配、精度区间错配),仍标 conflict。该边只把一致性当整体数值看,未登记「锚误差与校准抽样误差如何分别传播进该数值」这一层;全表 grep「锚 / 标定 / 零点 / 刻度 / 可识别 / 精度 / 方差 / 校准」在 591 行中仅 2 处命中且均无关(第 37 行「五脏是中医侧的锚点」、第 353 行「强度锚点=可见肌肉抽动」)。该主题尚未进入 L1 对应网络,已登记至 science-yixue-l1-gaps.md(「测量一致性」簇第 35 条)。


    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。

    Adding scale marks to a ruler is right, but adding them incorrectly makes it worse: the anchor’s error doesn’t advance into \”difference,\” but instead presses entirely onto the u…

    Evidence level: C

    The previous post stated that “adding an external anchor with its own scale completes the rank.” This is algebraically flawless but statistically answers only half the question: the anchor itself carries error, and this error precisely cancels out on the “how much do the two differ” side (becoming zero), but is fully accounted for on the “how far off is the zero point” side. The trouble is that what research truly needs is precisely the side that cancels out.

    Main Text

    The conclusion of the previous post (#1403) is: R readers provide only R observations, while the unknown quantity is R+1; the rank deficit is constantly 1, which readers cannot make up; only by introducing a physical anchor with its own unit (e.g., Celsius degrees for skin temperature, Newtons for grip strength) to fix the mean can the rank be completed.

    This is a direct consequence of linear algebra, which cannot be denied. But it only may support “a unique solution exists,” not “the solution error is small.” Once the anchor is truly integrated into the model, three things happen simultaneously, and in different directions.

    1. Good news: the anchor’s error precisely cancels out in the difference between two readers

    Let the reader reading be $z = x + b + \varepsilon$ and the anchor be $Y = k \cdot x + \eta$. With the anchor, each reader’s position estimate is $\hat{b} = \bar{z} – \bar{Y}/k$. The difference between two readers is obtained by subtracting their $\hat{b}$ values; the term $\bar{Y}/k$ is completely cancelled out in this subtraction.

    In a real test ($n=240$, 4 readers, each repeated 2 times, reader noise 0.5): as the anchor error increases from 0 to 0.3, 1.0, and 3.0, the sampling standard deviation of the difference between two readers remains constantly at 0.032; the first decimal place stays at 0.0 unchanged, and the upper bound of the absolute difference per sample is $5.55 \times 10^{-17}$—not “so small it can be ignored,” but the rounding floor noise of floating-point computation.

    This is the most important sentence in this post: adding an anchor imposes no precision cost and yields no precision gain for the “difference” itself. That 0.032 is entirely determined by reader noise and case count; the anchor error knows nothing about it.

    2. Bad news: on the absolute value side, the error is fully loaded

    In the same simulation, the standard deviation of a single reader’s position estimate rises from 0.0224 (when the anchor error is 0) to 0.0328 (when the anchor error is 0.354, a 46% increase), and further to 0.0680 (a doubling, when the anchor error is 1.0).

    Breaking this down reveals two additive terms: the reader term (0.000521) and the anchor term ($s^2/n$). These two terms are equal at $s^* = k \cdot se / \sqrt{m}$; in this example, $s^* = 0.354$. When the anchor error is below this value, reader noise dominates, and adding an anchor is essentially wasted effort; when it exceeds this value, the anchor term dominates, meaning adding an anchor is equivalent to handing over precision to a worse ruler.

    And “how far off is the zero point”—this absolute value—is originally something nobody can use: it shifts arbitrarily with population composition (the true mean $\mu$). If $\mu$ shifts by 5 units, $\hat{b}$ shifts in the opposite direction by 5 units. The significance of adding an anchor lies entirely in fixing this $\mu$; the product after fixing it is precisely the half that is least needed.

    3. What really goes wrong is “each reader gets its own anchor”

    The cancellation in the common-anchor case occurs because everyone shares the same $\bar{Y}$, so subtraction eats up the common term. If we change to each reader calibrating against its own anchor (which is normal when crossing centers or devices), cancellation no longer occurs, and error becomes the sum of two independent terms:

    When anchor error is 0.175, the standard deviation of the difference rises from 0.0323 to 0.0362 (1.12 times), requiring 269 cases to regain the original precision; at 0.354 it rises to 0.0457 (1.41 times), requiring 479 cases; at 0.600 it becomes 0.0559 (2.00 times), requiring 960 cases; at 1.000 it becomes 0.0968 (2.99 times), requiring 2148 cases.

    The inflation factor is $1 + (s/s^*)^2$, which is independent of $n$, so one can only compensate by adding samples. When anchor error equals reader noise ($s = 0.354$), the sample size must double to recover; when it is three times, nine times.

    This is the most practical criterion in this round: if anchor error does not fall below reader noise, adding an anchor is a trade-off of doubling the sample size to gain an unnecessary absolute value.

    4. Calibration subsamples: the amplification factor is $n/n_{cal}$

    In practice, it is rare to apply an anchor to the full sample; usually a subset is drawn for calibration. In this case, the anchor term becomes $s^2/n_{cal}$, while the reader term is $se^2/(mn)$; their ratio is the amplification factor. With $n=240$, $se=0.5$, $m=2$, and anchor error 0.5:

    • Calibrating 10 people: SD 0.1598, which is 5 times the uncalibrated value (0.0323); converted to 5pp equivalents, it is 1.3 times worse than not calibrating
    • Calibrating 30 people: SD 0.0941, 0.7 equivalents
    • Calibrating 60 people: SD 0.0685, 0.54 equivalents

    Systematic reviews have long noted this line: if the calibration subsample is fewer than 30 people, the confidence interval after correction widens significantly due to uncertainty in the correction coefficient itself. Other simulation studies comparing sampling strategies for internal validation subsamples found that even drawing 40% of the main sample, stratified random sampling and extreme-value sampling improve over pure random sampling by only 10% to 12%, and this efficiency gain comes at the cost of higher bias percentages and lower coverage—efficiency is not free.

    5. Adding readers cannot save common-mode anchor error

    If part of the anchor error is shared across the batch (e.g., same instrument, same batch of reagents, or the same senior person judging by curve), this part does not shrink with $n$ and leaves a floor.

    In the real test with a floor of 0.13, the standard deviations for sample sizes of 60, 240, and 960 are 0.1371, 0.1316, and 0.1307 respectively; even though the sample size increases 16-fold, the standard deviation drops only by 5%.

    Adding readers also cannot save this. When anchor error is 1.0, increasing the number of readers from 2 to 16 reduces the standard deviation only from 0.0666 to 0.0645, exactly hitting the floor of 0.0645: the $1/\sqrt{n}$ part is averaged out by the number of readers, while the other half remains unchanged.

    This also explains why regression calibration literature repeatedly emphasizes that “the reference measurement is not necessarily an unbiased gold standard”—the error of the reference measurement and that of the tested instrument may be correlated, which can unpredictably overestimate or underestimate the correction coefficient; using a flawed reference measurement can lead to the true effect being underestimated by about half and the required sample size doubling.

    6. When $k$ is unknown, do not perform division; directly take the intercept

    All the algorithms above assume the anchor conversion coefficient $k$ is known. If $k$ is unknown, the intuitive approach is to first regress to estimate the slope and then divide; the problem is that division also divides the anchor error by $k$—when $k=0.5$, the anchor error is doubled.

    The correct approach is to take the regression intercept: regress the reading against the anchor, and the intercept itself is $b$. In the real test with $k=0.5$ and anchor error 0.6, the division route yields a standard deviation of 0.0803, while the intercept route yields 0.0539, which is 1.49 times more accurate; when $k=2.0$, the two are almost identical (0.0291 vs. 0.0287).

    The worry that “if $k$ is unknown, calibration cannot be done” is unnecessary, provided one uses the intercept instead of division. This corresponds to the second step of the standard three-step regression calibration in literature: regress the gold standard against observed values, and use the regression coefficients (including the intercept) themselves as the source of the correction amount.

    Evidence and Limitations

    Verified: When using a common anchor, the sampling standard deviation of the difference between two readers is completely independent of the anchor error (anchor error ranging from 0 to 3.0, standard deviation constantly at 0.032; the upper bound of absolute difference per sample is $5.55 \times 10^{-17}$); the variance of a single reader’s absolute position decomposes into the sum of the reader term and the anchor term, with equality at $s^ = k \cdot se / \sqrt{m} = 0.354$, and the real data matches the analytical prediction at six points ($s=0$ to $2.0$) with a maximum deviation of 0.0011; when using individual anchors, the variance inflation factor equals $1 + (s/s^)^2$, with $s=0.175/0.354/0.600/1.000$ yielding real values of 1.12/1.41/2.00/2.99 versus predicted 1.12/1.42/1.97/3.00; the amplification factor for calibration subsamples equals $n/n_{cal}$; the floor of common-mode anchor error does not decrease with sample size (with $s=0.13$, the standard deviations for three sample sizes are 0.1371/0.1316/0.1307); adding readers is ineffective under common-mode error (increasing from 2 to 16 reduces SD only from 0.0666 to 0.0645, hitting the floor); when $k$ is unknown, the intercept route outperforms the division route by 1.49 times when $k=0.5$.

    Partially verified: the 5pp equivalent (0.126 standard deviations) comes from numerical solutions in the upstream of this series; this post only uses it as a precision conversion scale and did not re-derive it; the statement “common-mode error does not shrink with sample size” is based on a simplified structure of “within-batch common shift + individual random term”; real shared error may have hierarchical structure (center–device–batch three levels), and the variance proportions of each level were not measured; the intercept route outperforms the division route only under two-parameter linear relationships; if the relationship between anchor and reading is nonlinear (saturation, upper-limit truncation), the meaning of the intercept changes.

    To be verified: ① In the complete design with more than three readers and an external anchor, if anchor error and reader noise are positively correlated (since the anchor is also judged by humans), effective independent precision will be amplified by $s/\sqrt{1-\rho^2}$, but $\rho$ itself has never been measured in the judgment data; ② In degenerate cases such as “using a senior person judging by curve as the anchor,” whether the zero points of the anchor and readers share the same arbitrary origin has not been tested; ③ Whether the 5pp equivalent threshold remains executable when common-mode anchor error exists; this post only provides the floor value, not the critical floor where the criterion fails.

    Related Reading

    Evidence level: C
    References: PMID 25627982, PMID 33878166, PMID 23553407, PMID 37704705, PMID 8082971, PMID 22468848, PMID 30402814
    L1 network status: In CORRESPONDENCE_NETWORK.md, line 202, “E-060 label premise: five-zang differentiation (as grouping label) → label credibility” was originally recorded as conflict (unresolved)—”consistency among TCM physicians in an open population is ‘low’ (PMID 30383398)” and “training can improve it (PMID 19538100)” stand side by side; line 223, KB-317 adds two mismatches (sample size unit mismatch, precision interval mismatch), yet still marked conflict. This edge treats consistency as an overall numerical value without registering “how anchor error propagates into the components of this number”; grepping “anchor / calibration / zero point / scale / identifiable / precision / variance” in 591 lines hits only 2 places and both are irrelevant (line 37 “five-zang is the anchor point on the TCM side”, line 353 “intensity anchor point = visible muscle twitch”). This topic has not yet entered the corresponding L1 network, and has been registered to science-yixue-l1-gaps.md (cluster “measurement consistency”, item 35).


    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.

  • 止损闸门要留多少余量:n=30 的 ICC 闸门连自己的门槛都判不准

    Evidence level: B

    医疗免责声明:本文为研究设计与统计方法学讨论,不构成诊断标准或临床建议。任何心功能评估须由执业医师依据规范完成。

    池③ 有一张卡叫 P3-EXP-057(#187):投入 N=480 的大队列之前,先用 n=30 做一次复测信度预实验,把 C2/C3 比值这个”尺子”先量一遍。卡里写得很干净——H1 要求同日 ICC(2,1) ≥ 0.75,F1 规定”同日 ICC < 0.75 → 终止,课题以负结果收尾”。

    这条规则读起来毫无问题。问题是它判不准。而且判不准的方向不是”判松了”,是”判死了”。

    一、先把判据的分辨率算出来

    n=30、同日 3 次重复(k=3),ICC(2,1) 的 95% CI 用 McGraw-Wong 那套 F 分布法可以解析算。当点估恰好在 0.75 时:

    指标 数值
    点估 0.750
    95% CI [0.598, 0.861]
    CI 全宽 0.264
    CI 半宽 0.132

    也就是说,你测出一个 0.75,真值落在 0.60 到 0.86 之间都不奇怪。而这张卡用来分档的”带宽”是 0.75 与 0.60(”中等”下界)之间的 0.15。

    按池④ 第 23 轮立的 R26(判据分辨率必须匹配输入分辨率)算一遍比值:判据半带宽 0.075 ÷ 输入半宽 0.132 = 0.57。小于 1,就是伪精度。这条闸门的刻度比它的噪声还细。

    H2 更薄。跨 7 天只有 2 个时点(k=2),n=30 时 CI 全宽 0.386,半宽 0.193,比值掉到 0.26。

    二、真值正好卡在门槛上时,会发生什么

    CI 宽不是”误差大一点”,它会直接改写决策。对 H1 做蒙特卡洛(每档 30 万次模拟):

    真值 ICC 观测中位 判”通过” 判”F1 终止”
    0.60 0.597 2.7% 97.3%
    0.70 0.697 23.1% 76.9%
    0.75 0.748 48.6% 51.4%
    0.80 0.798 77.8% 22.2%
    0.85 0.849 95.7% 4.3%

    看 0.75 那一行。真值恰好等于门槛,有一半的几率被判”终止”。 不是判”不确定”,是判死——卡里 F1 的处置是”课题以负结果收尾”。

    再看 0.85 那行:真值比门槛高 0.10,仍有 4.3% 被枪毙。这意味着照这张卡执行,一个完全合格的比值方案,每二十次里会有一次被自己的预实验否掉,而且理由写得很正当——”数据显示 ICC 未达 0.75″。

    三、错在哪:把”没有结论”硬塞进”终止”

    根本问题不是样本量小(n=30 是刻意选的最小代价),而是二值决策规则少了一个出口。

    ICC 点估落在 0.60–0.85 这个区间时,数据其实什么都没说清。但规则只有两条路:≥0.75 继续,<0.75 终止。于是所有”说不清”都被归进”终止”。

    而这两条路的代价是不对称的:

    • 终止:课题死掉,写进档案。若判错,是整个研究方向被误杀,而且事后没人会回头复查一个已成定论的负结果。
    • 继续:多花一点钱。若判错,后面 N=480 的数据会自己暴露问题——那张大卡里本来就有阶段 2 的检验。

    不可逆的一侧,门槛必须偏保守;可逆的一侧,可以放宽。 这是标准的非对称代价原则,而现卡把门槛恰恰设在了目标值上——等于默认两侧代价相同。

    四、怎么改:把阈值挪开,再加一道筛

    把”终止线”从 0.75 挪到 0.60,规则改成:点估 < 0.60 才终止,其余一律继续(含加 n 重测)。同样模拟:

    终止阈值 真值 0.50 检出 真值 0.60 检出 真值 0.75 假终止 真值 0.85 假终止
    0.75(现卡) 99.9% 97.3% 51.3% 4.3%
    0.70 98.7% 88.8% 25.9% 1.0%
    0.60 84.9% 51.3% 4.1% 0.0%

    阈值放在 0.60,假终止率从 51% 掉到 4%,代价是”真值 0.50 的差尺子”只能抓出 85%——但那 15% 漏网的会在阶段 2 被 N=480 抓出来,损失是可回收的。

    配套的一条是顺序设计:n=30 只当第一道粗筛,落在中间区就加人再判。

    累计 n CI 半宽(k=3) 能分辨的最小差
    30 0.132 0.264
    60 0.092 0.185
    100 0.071 0.142
    200 0.050 0.100

    n=100 时 CI 半宽降到 0.071,”0.75 与 0.60 之间”这个判断才第一次变得可做。样本量需求本身也不是拍脑袋——Bonett 2002 给过 ICC 精度与样本量的解析关系(PMID 12111881),可以直接反解。

    五、同一个毛病还有两处

    H3 的 ΔICC ≥ −0.05。 单个 ICC(n=30, k=3)的抽样 SD 约 0.071;两个指标相减,差值 SD 最差到 0.100(若两者完全相关则可降到接近 0,方向未知)。判据带宽 0.05 只有最差 SD 的一半。用一条容差比噪声还窄的规则去判”比值是否比单值更差”,结论必然由噪声决定。

    F5 的”两层 ICC 差异 > 0.20 判依赖人群”。 每层只有 n=15,CI 半宽 0.191,判据半带宽 0.100,比值 0.52。模拟显示:真实层间差为 0 时,仍有 17.3% 的概率被判”依赖人群”;真实差 0.10 时检出率只有 26–36%。这条规则既会误报也会漏报,而且误报的方向是”要求分亚组”,会让本来能合并的分析被不必要地拆散。

    六、这条能不能推广

    池④ 上一轮在 K1–K4 判档体系里立 R26,池③ 这一轮从 ICC 闸门独立撞到同一件事——两个池、两套完全不同的判据,比值都掉到 0.3–0.6。这不是巧合,是同一个失败形状:先定门槛,再看数据够不够支撑这个门槛。

    所以除了 R26,本轮再补一条 R27「闸门的方向不对称」:止损闸门的”终止”与”通过”代价不对称时,阈值不得设在目标值上。设在目标值上,等于默认把所有”说不清”都归入”终止”。

    R26 与 R27 分工明确:R26 管带宽够不够(分辨率问题),R27 管阈值该放在哪(代价问题)。带宽不够时,靠挪阈值只能止损、不能救回精度——两件事不可互替。

    证据与局限

    证据等级 B(不是 A,也不是 E1)。

    • 五条引文全部经 NCBI E-utilities 逐条回查:Koo & Li 2016 的 ICC 选择与报告指南(PMID 27330520)是 H1/H2 阈值 0.75/0.70 的出处;Bland & Altman 1986(PMID 2868172)是 H4 一致性分析方法;Shrout & Fleiss 1979(PMID 18839484)是 ICC 分类的原始文献;Bonett 2002(PMID 12111881)给样本量-精度关系;Stoner 2017(PMID 27861248)是脉搏波形参数信度的间接支撑。
    • 但没有任何一条文献直接研究”HRI-C2/C3 复测信度”,也没有一条给出”止损闸门阈值应如何随 n 调整”的现成公式。本文的门槛数值全部是在”同一真值、独立重复”假设下的解析推导与模拟,不构成对 C2/C3 实际 ICC 的任何估计。
    • 模拟假设 ICC(2,1) 的正态、无时间趋势、无操作者变异(卡里已明确排除操作者变异)。若同日 3 次采集间隔 ≥1 h 导致受试者状态自相关,真 ICC 会随 k 变化,本文的不确定性会被低估。
    • risk_level: medium + disclaimer_required: true——涉及医疗测量,不得进入面向患者的表达。
    • 本文没有回答”C2/C3 比值是否可用”,只回答”这张卡的闸门能不能判出这个答案”。R26/R27 只经本库两处实例检验,按 R11 只承担”发现伪精度”的证伪职能,不得反过来当”判据已合格”的证明。

    本轮建议(待 JOE 决策,不停在建议层不动手)

    编号 建议 优先级
    D27 P3-EXP-057 F1 的终止阈值 0.75 → 0.60,并显式写出”中间区不终止,加 n 重测” P1
    D28 P3-EXP-057 H3 的 ΔICC ≥ −0.05 改为”报差值 95% CI”,不设点容差 P1
    D29 P3-EXP-057 F5 的”层间差 > 0.20″改为”两层 CI 不重叠”或补足每层 n ≥ 60 P2
    — H1/H2 门槛数值本身(0.75/0.70)属领域惯例,本轮不建议改动 —

    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 much margin to leave for the stop-loss gate: ICC gates with n=30 cannot even determine their own thresholds

    Evidence level: B

    Medical Disclaimer: This article discusses research design and statistical methodology. It does not constitute diagnostic criteria or clinical advice. Any assessment of cardiac function must be performed by a licensed physician according to established standards.

    Card Pool ③ contains a card named P3-EXP-057 (#187): Before investing in the large cohort of N=480, it first conducts a pilot reliability pre-experiment with n=30 to measure the C2/C3 ratio “ruler.” The card states clearly—H1 requires ICC(2,1) ≥ 0.75 on the same day; F1 stipulates “ICC < 0.75 on the same day → terminate, and the project ends with a negative result.”

    This rule reads perfectly fine. The problem is that it cannot judge accurately. Moreover, the direction of inaccuracy is not “too lenient,” but rather “too strict.”

    I. First, calculate the resolution of the criterion

    With n=30 and 3 repeated measures on the same day (k=3), the 95% CI for ICC(2,1) can be analytically calculated using McGraw and Wong’s F-distribution method. When the point estimate is exactly 0.75:

    Metric Value
    Point estimate 0.750
    95% CI [0.598, 0.861]
    Full CI width 0.264
    Half CI width 0.132

    In other words, if you measure 0.75, it is not surprising that the true value lies between 0.60 and 0.86. Yet, the “bandwidth” used for classification in this card is 0.15 (the gap between the threshold 0.75 and the lower bound of “medium” at 0.60).

    According to R26 established in Round 23 of Card Pool ④ (the criterion’s resolution must match the input resolution), let us calculate the ratio: criterion half-width 0.075 ÷ input half-width 0.132 = 0.57. Less than 1, this is pseudo-precision. The scale of this gate is finer than its own noise.

    H2 is even thinner. With only 2 time points across 7 days (k=2) and n=30, the full CI width is 0.386, half-width 0.193, dropping the ratio to 0.26.

    II. What happens when the true value lands exactly on the threshold?

    The wide CI does not merely mean “large error”; it directly alters decision-making. Monte Carlo simulations for H1 (300,000 iterations per tier):

    True ICC Observed median Pass Fail (F1 terminate)
    0.60 0.597 2.7% 97.3%
    0.70 0.697 23.1% 76.9%
    0.75 0.748 48.6% 51.4%
    0.80 0.798 77.8% 22.2%
    0.85 0.849 95.7% 4.3%

    Look at the row for 0.75. When the true value equals the threshold, there is a 50% chance of being judged “terminated.” It is not judged “uncertain”; it is judged dead—the card’s F1 disposition states “the project ends with a negative result.”

    Now look at the row for 0.85: even though the true value is 0.10 higher than the threshold, 4.3% are still shot down. This means that under this card’s rules, a perfectly qualified ratio scheme has a 1-in-20 chance of being rejected by its own pre-experiment, and the reason stated is entirely valid—”the data shows ICC did not reach 0.75.”

    III. Where the error lies: forcing “no conclusion” into “terminate”

    The fundamental issue is not small sample size (n=30 was deliberately chosen as the minimum cost). Rather, the binary decision rule lacks a third exit.

    When ICC point estimates fall between 0.60 and 0.85, the data actually tells us nothing clearly. Yet the rules offer only two paths: ≥0.75 continue, <0.75 terminate. Thus, all “unclear” cases are lumped into “terminate.”

    The cost of these two paths is asymmetric:

    • Terminate: The project dies, recorded in the archive. If judged wrongly, an entire research direction is mistakenly killed, and no one will revisit a settled negative result afterward.
    • Continue: Spend a bit more money. If judged wrongly, the N=480 data will eventually expose the problem—the large card already includes Stage 2 validation.

    The irreversible side must have a conservative threshold; the reversible side can be relaxed. This is the standard principle of asymmetric cost, yet the current card sets the threshold exactly at the target value—effectively assuming equal costs on both sides.

    IV. How to fix it: shift the threshold and add a sieve

    Move the “termination line” from 0.75 to 0.60, changing the rule to: terminate only if point estimate < 0.60; otherwise continue (including adding n for retesting). Simulations:

    Termination threshold Detect true 0.50 Detect true 0.60 False terminate at true 0.75 False terminate at true 0.85
    0.75 (current card) 99.9% 97.3% 51.3% 4.3%
    0.70 98.7% 88.8% 25.9% 1.0%
    0.60 84.9% 51.3% 4.1% 0.0%

    Setting the threshold at 0.60 reduces false termination rate from 51% to 4%, at the cost that “poor rulers with true value 0.50” can only be caught 85% of the time—but those 15% missed will be caught in Stage 2 by N=480, and the loss is recoverable.

    A complementary element is sequential design: n=30 serves only as the first coarse screen; those falling in the middle zone add more subjects for re-evaluation.

    Cumulative n Half CI width (k=3) Minimum detectable difference
    30 0.132 0.264
    60 0.092 0.185
    100 0.071 0.142
    200 0.050 0.100

    At n=100, the half CI width drops to 0.071, making the judgment between “0.75 and 0.60” finally feasible for the first time. Sample size requirements are not arbitrary—Bonett (2002) provided an analytical relationship between ICC precision and sample size (PMID 12111881), which can be directly inverted.

    V. Two other flaws in the same vein

    H3’s ΔICC ≥ −0.05. The sampling SD of a single ICC (n=30, k=3) is about 0.071; the difference between two indicators has an SD as bad as 0.100 (if both are perfectly correlated it could drop near zero, direction unknown). The criterion bandwidth of 0.05 is only half the worst-case SD. Using a tolerance narrower than the noise itself to judge “whether the ratio is worse than the single value” inevitably leads to conclusions driven by noise.

    F5’s “two-layer ICC difference > 0.20 indicates population dependence.” Each layer has only n=15, half CI width 0.191, criterion half-width 0.100, ratio 0.52. Simulations show: even when the true inter-layer difference is zero, there is still a 17.3% chance of being judged “population-dependent”; when the true difference is 0.10, detection rate is only 26–36%. This rule both falsely alarms and misses detections, and the false alarm direction is “require subgroup splitting,” which unnecessarily breaks apart analyses that could be combined.

    VI. Can this be generalized?

    In the previous round, Card Pool ④ established R26 within the K1–K4 tiering system. Now, Card Pool ③ independently hits the same issue via ICC gates—two pools, two completely different criteria, yet ratios both drop to 0.3–0.6. This is not coincidence; it is the same failure shape: first set a threshold, then check whether the data supports that threshold.

    Therefore, besides R26, this round adds R27 “asymmetric direction of the gate”: when the costs of “terminate” and “pass” for a stop-loss gate are asymmetric, the threshold must not be set at the target value. Setting it at the target value equals assuming all “unclear” cases go into “terminate.”

    R26 and R27 have clear divisions: R26 manages whether the bandwidth is sufficient (resolution issue), while R27 manages where to place the threshold (cost issue). When bandwidth is insufficient, shifting the threshold can only mitigate loss—it cannot recover precision. These two matters cannot substitute for each other.

    Evidence and Limitations

    Evidence level B (not A, nor E1).

    • All five citations were individually verified via NCBI E-utilities: Koo & Li (2016) on ICC selection and reporting guidelines (PMID 27330520) is the source for H1/H2 thresholds 0.75/0.70; Bland & Altman (1986) (PMID 2868172) provides the consistency analysis method for H4; Shrout & Fleiss (1979) (PMID 18839484) is the original paper on ICC classification; Bonett (2002) (PMID 12111881) gives the sample size-precision relationship; Stoner (2017) (PMID 27861248) offers indirect support for pulse waveform parameter reliability.
    • However, no single literature directly studies “HRI-C2/C3 retest reliability,” nor does any provide a ready-made formula for “how the stop-loss gate threshold should adjust with n.” All threshold values in this article are analytically derived and simulated under the assumption of “same true value, independent repetition,” and do not constitute any estimate of actual ICC for C2/C3.
    • Simulations assume ICC(2,1) is normally distributed, without time trend, and without operator variation (the card explicitly excludes operator variation). If same-day three collections have intervals ≥1 h leading to subject autocorrelation, the true ICC will vary with k, and this article’s uncertainty will be underestimated.
    • risk_level: medium + disclaimer_required: true—involving medical measurement, must not enter patient-facing expression.
    • This article does not answer “whether the C2/C3 ratio is usable,” only answers “can this card’s gate judge this answer.” R26/R27 have been tested only in two instances of this repository; according to R11, they bear only the falsification function of “discovering pseudo-precision” and must not be used as proof that “the criterion is qualified.”

    This round’s recommendations (pending JOE decision, do not remain at the recommendation level without action)

    Number Recommendation Priority
    D27 Change P3-EXP-057 F1 termination threshold from 0.75 → 0.60, and explicitly state “middle zone does not terminate; add n for retest” P1
    D28 Change P3-EXP-057 H3 ΔICC ≥ −0.05 to “report difference 95% CI,” without point tolerance P1
    D29 Change P3-EXP-057 F5 “inter-layer difference > 0.20” to “two-layer CIs do not overlap” or supplement each layer with n ≥ 60 P2
    — H1/H2 threshold values themselves (0.75/0.70) belong to domain convention; not recommended to change in this round —
  • 五道证伪闸门串成一张门票:F1–F5 的刻度比它要量的间隙还粗

    Evidence level: B

    先给结论

    池⑤ 的《三质↔气虚 RSA 反证边界》设计里,F1–F5 被写成五条并列的”证伪条件”,而 §11 又写”进入 779 intake 的条件:H1 + H2 同时通过且 F1–F5 全部未触发”。这一句把五道止损阀直接当成了准入证。两件事的阈值本该不一样,这里用了同一套。

    本轮只读复核(不改正文)算下来,最硬的一句是:

    F2 的间隙是 Δr 0.10 → 0.15,宽 0.05;
    F2 自己的 95% 噪声半宽 0.136(ρ=0.5);
    F3 每条件 n=12,噪声半宽 0.49。
    间隙只有噪声的 1/2.7 和 1/9.8 —— "中等衰减"这个区间,物理上判不出来。
    

    一、五把尺子,五个刻度,被串成一个 AND

    F1–F5 分别长在五个不同的量纲上:|r|(相关值)、Δr(差值)、票数(k of n)、ΔCFI(拟合指标差)、比例(受试者误解率)。五者没有共同刻度,却被一个 AND 串起来当门票用。

    这带来的第一个后果是可算的:若四道独立闸门各自的假杀率分别是 0.50、0.50、0.131、0.370(取”真值恰好压在门槛上”这个决策最相关的点),全部通过的概率是 13.7%。也就是说,一个”刚好合格”的项目有 86.3% 概率被至少一道闸门判死。而”刚好合格”恰恰是准入判断唯一真正要处理的区间——明显好和明显差的项目,用一道闸门就够。

    多道止损阀串联不等于”更严格”,只等于”更容易否决”。 严格与否要看单闸的判据带宽,不看串联的道数。

    二、F1:门槛压在目标值上,n=120 只分得开 0.50 与 0.35

    F1 是”H1 不满足即降级 E3″,而 H1 的门槛是 |r| ≥ 0.50。真值恰好 0.50 时,被判失败的概率是 50.0%;真值 0.55 时仍有 22.7%。这不是精度差,是阈值设在了目标值上,等于把所有”说不清”都归进不可逆的那一侧(R27)。

    n=120/组时,r=0.50 处的 95% CI 半宽是 0.136。按 Cohen 的锚点(0.30 与 0.50 差 0.20),判据半带宽是 0.10,R26 比值 0.74 < 1;按更严的锚点 0.40,半带宽 0.05,比值降到 0.37。要把”0.50 与 0.40″分开,需要 SE_z ≤ 0.0255,即 n ≥ 1541/组——现设计 120,差 12.8 倍。

    两个锚点都给出来,是因为乘性/非线性尺度上的”±X”只在一个锚点成立,不写锚点的百分比会被读成全局。

    三、F3:多数表决的票数门槛没随票数重算

    F3 是”5 个扰动探针中 ≥ 3 个使 |r| 衰减 ≥ 0.15 即判失败”。这是 (k,n) 多数表决。

    每条件 n=12(60 人 / 5 条件,每条件 ≥12),SE_dz = 1/√(12−3) = 0.333,Δr 的 95% 半宽 0.49。真 Δr = 0 时,单个探针”假报 ≥0.15″的概率是 0.274(ρ=0.5)到 0.336(ρ=0)。三个探针凑够票,概率 13.1%–21.3%。

    反直觉的一条:探针从 5 个加到 8 个,若票数门槛仍是 3 票(37.5%),假报率从 0.130 升到 0.382;加到 10 个仍是 3 票(30%),升到 0.545。加探针而不改票数门槛,会让闸门更容易假报。 要保住 60% 的票数比,10 个探针需 ≥6 票,假报率才回落到 0.031。

    这是因为多数表决的假报率不是 n 的单调函数——它由”票数比 k/n”和 n 共同决定,加票不改比例就是在放松。

    四、F3 的阶梯是倒挂的:结论越重,证据越薄

    F3 一旦触发,结论从 E3 降到 E2 并”必须说明扰动敏感性”——这是全卡里后果最重的一档。但它的样本最小:每条件 12 人。

    Δr 的标准误是 F1(n=120)的 3.61 倍,而门槛只放宽 3.0 倍(0.15 vs 0.05)。要让 F3 达到与 F1 同等的分辨率,每条件需 3460 人(ρ=0.5),5 个条件共 17,302 人——现设计是 60 人。

    越往”最弱假设”走,需要的证据越硬,而设计给的样本越少。 这个倒挂不是笔误,是”证伪条件按严重程度排序、样本量按方便程度分配”两套逻辑没有对齐的结果。

    五、F5:”>20%” 在 n=20 时实际只能落到 25%

    认知访谈 n=20,规则写”出现 >20% 受试者语义误解”。20% × 20 = 4,而 > 不含等号,所以只有 X ≥ 5 才算触发,实现阈值是 5/20 = 25%,比声明值高 1.25 倍。

    真误解率恰好 20% 时,触发 F5 的概率是 0.370。这个数在 n=20 时最小,随 n 增大反而上升到 0.466(n=300)——因为大 n 下实现阈值逼近 20%,判据变”灵敏”,而 20% 本身就是门槛。离散计票的阈值实化误差与统计功效的方向是相反的,加样本并不能同时改善两者。

    六、顺带把 #22 的三档也扫了一遍:R29 不适用,但另有形态

    池④ 上一轮在 K1–K4 上立了 R29「判据单侧自动收紧」,并把 #22 的 A/B/C 列为”疑同一缺陷,待查”。本轮补算:#22 不触发 R29。

    按 1 km / 7.83 Hz 复算,A 档真实边界是 σ ≥ 4.283×10⁻²(A = 10 dB,D1′ 记 4.29×10⁻²,吻合),C 档边界 σ ≤ 10⁻³(A = 1 dB)。σ 的不确定度是 ×q:

    A 档门槛(σ 下界)随 q 上移:σ_A·q   —— 更难进 A
    C 档门槛(σ 上界)随 q 下移:σ_C/q   —— 更难进 C
    B 档 = (σ_C/q, σ_A·q),宽度比从 42.8× 涨到 q=10 时的 4282.9×
    

    两侧同时外扩,方向对称,因此没有单侧偏心。但出现了另一种形状:

    R29-b「中间档吸收」——判定力不偏向某一侧,而是全部塌进中间那一档。精度变差时 A 档与 C 档同时消失,所有站都落进”条件性测”。后果与 R29 一样(永远不说”不用测”,永远说”以后再说”),机制不同:R29 是门槛函数形式不同导致单侧收紧,R29-b 是宽档两侧外扩导致全域吸收。

    两者不得互相替代:R29 的解法是统一门槛的函数形式;R29-b 的解法是取消”宽档”这个吸收态,或给它设上限。

    七、新增两条硬规矩

    R30「合取门槛必须报联合功效」:多道止损闸门串联作为准入条件时,必须同时报”单闸假杀率 → 全过概率”。只报”我们有五道检查”会被读成严格,而实际是 13.7% 的通过率。

    R31「多数表决的票数门槛必须随票数重算」:(k,n) 结构的假报率不是 n 的单调函数。加探针不改票数比会让假报率上升;任何”多设几个探针更稳”的写法都要重算 k。

    八、引用核查中的两个硬伤(如实报)

    复核过程中顺手扫了本设计 §12 的引文,发现两处:

    原写 实况 处置
    PMID 24233341(Lehrer PM, HRV biofeedback: effects across domains, Appl Psychophysiol Biofeedback 2013) PubMed 无此记录(efetch 返回空集) 建议换 25101026(Lehrer PM, Heart rate variability biofeedback: how and why does it work?, Front Psychol 2014;5:756,已核)
    PMID 39563397(Li Z, acupuncture… functional dyspepsia, Chin Med 2024) 该号真实为 Loix M, Stearoyl-CoA desaturase-1, Mol Neurodegener 2024;19 —— 主题完全无关 建议换 30246392(Liu Y, 电针 ST36 大鼠消化不良模型, Neurogastroenterol Motil 2019)或直接删除该条
    PMID 34479992(Chung JW, HRV and TCM patterns) 假号,已登记 P-1 层 I-2/I-3 仍在等 JOE 批,本轮未动

    本轮不改这个文件:它属 779 intake 三件套范围,且引文治理全在 JOE 待批。只登记,不执行。

    证据与局限

    本文全部为解析推导与二项/正态模拟,不含任何实测数据。三条必须写明的边界:

    1. 「判不出」不等于「假说不成立」。 本文只说”这套闸门判不出它想判的答案”,不说”Vata-low ≈ 气虚-low”为真或为假。按 R11,零参数/纯解析判据只承担证伪职能,不得反过来当”判据合格”的证明。
    2. ρ(两次测量间的相关)未实测。 F2/F3 的噪声全部依赖 ρ,本文给了 ρ = 0 到 0.9 的区间。真实 ρ 取决于同一批人重复测量的间隔与体位控制,未做过预实验。
    3. 有效推进的边界:本文提出的是判据层的结构问题,不是研究结论。F1–F5 的具体修订(阈值、票数、样本量)全部停在建议层,等 JOE 批。

    相关阅读

    • 池④ #77 K1–K4 判档体系结构审计(R26 起源)
    • 池③ #187 ICC 闸门 CI 宽度前置补全(R27 起源)
    • 池④ #26 六站逐站过闸(R29 起源)
    • 池⑤ #19 #20 证伪优先方法学 / 预注册规范(F1–F5 出处)

    引用锚点

    • Task Force of ESC/NASPE. Heart rate variability: standards of measurement, physiological interpretation and clinical use. Circulation. 1996. PMID: 8598068.
    • Cohen J. A power primer. Psychol Bull. 1992. PMID: 19565683.
    • Lehrer PM, Gevirtz R. Heart rate variability biofeedback: how and why does it work? Front Psychol. 2014;5:756. PMID: 25101026.
    • Lakens D. Equivalence Tests: A Practical Primer for t Tests, Correlations, and Meta-Analyses. Soc Psychol Personal Sci. 2017. PMID: 28736600.
    • Nosek BA, Ebersole CR, DeHaven AC, Mellor DT. The preregistration revolution. Proc Natl Acad Sci USA. 2018. PMID: 29531091.
    • Koo TK, Li MY. A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research. J Chiropr Med. 2016. PMID: 27330520.
    • Chen FF. Sensitivity of Goodness of Fit Indexes to Lack of Measurement Invariance. Struct Equ Modeling. 2007. doi:10.1080/10705510701301834.

    免责声明:本文为方法学审计,涉及中医体质、阿育吠陀三质与自主神经指标,属健康相关主题。文中所有结论均为统计判据层面的推导,不构成任何诊断、治疗或体质判定建议,不得用于临床决策或产品化。

    review_status: draft —— 未经 779 intake 审核,不得标 accepted。


    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。

    Five falsification gates strung into one ticket: the scale of F1–F5 is coarser than the gap it measures

    Evidence level: B

    Conclusion First

    In the design of Pool ⑤’s “Three Constitutions ↔ Qi Deficiency RSA Falsification Boundary”, F1–F5 are written as five parallel “falsification conditions,” while §11 states: “Condition to enter the 779 intake: H1 + H2 both pass AND F1–F5 all do NOT trigger.” This sentence treats the five stop-loss valves directly as an admission ticket. The thresholds for these two functions should differ, yet the same set is used.

    This round involves only read-and-review (no changes to the main text). The most critical calculation is:

    The gap for F2 is Δr 0.10 → 0.15, width 0.05;
    F2's own 95% noise half-width is 0.136 (ρ=0.5);
    F3 per condition n=12, noise half-width 0.49.
    The gap is only 1/2.7 and 1/9.8 of the noise — "moderate attenuation" cannot be physically distinguished.
    

    I. Five Rulers, Five Scales, Strung Together into One AND

    F1–F5 grow on five different dimensions: |r| (correlation), Δr (difference), votes (k of n), ΔCFI (fit index difference), and proportion (misunderstanding rate). These five lack a common scale yet are strung together by an AND gate to serve as the ticket.

    The first consequence is calculable: if the false-kill rates of the four independent gates are 0.50, 0.50, 0.131, and 0.370 (taking the decision-relevant point where “truth lies exactly on the threshold”), the probability of passing all is 13.7%. Thus, a “just-qualified” project has an 86.3% chance of being killed by at least one gate. Note that “just-qualified” is precisely the interval that admission judgment must truly handle; obviously good and obviously bad projects can be judged by a single gate.

    Serialing multiple stop-loss valves does not mean “stricter”; it means “easier to reject.” Whether it is strict depends on the bandwidth of the single-gate criterion, not the number of gates in series.

    II. F1: Threshold Set on Target Value, n=120 Distinguishes 0.50 from 0.35

    F1 is “if H1 fails, downgrade to E3,” while H1’s threshold is |r| ≥ 0.50. When truth is exactly 0.50, the probability of failure is 50.0%; at truth 0.55, it remains 22.7%. This is not poor precision; it is because the threshold is set on the target value, effectively casting all “unclear” cases into the irreversible side (R27).

    With n=120/group, the 95% CI half-width at r=0.50 is 0.136. Using Cohen’s anchor (difference of 0.20 between 0.30 and 0.50), the criterion half-bandwidth is 0.10, yielding R26 ratio 0.74 < 1; using the stricter anchor 0.40, half-bandwidth is 0.05, ratio drops to 0.37. To separate “0.50 from 0.40,” we need SE_z ≤ 0.0255, i.e., n ≥ 1541/group — the current design of 120 is off by 12.8 times.

    Both anchors are provided because “±X” on multiplicative/nonlinear scales holds only at one anchor; percentages without the anchor will be read as global.

    III. F3: Majority Vote Threshold Not Recalculated with Votes

    F3 is “among 5 disturbance probes, ≥3 cause |r| attenuation ≥0.15 → failure.” This is (k,n) majority voting.

    Per condition n=12 (60 people / 5 conditions, each ≥12), SE_dz = 1/√(12−3) = 0.333, Δr’s 95% half-width 0.49. When true Δr = 0, the probability of a single probe “falsely reporting ≥0.15” ranges from 0.274 (ρ=0.5) to 0.336 (ρ=0). Three probes reaching the threshold yields probability 13.1%–21.3%.

    Counterintuitively: Adding probes from 5 to 8, if the vote threshold remains 3 (37.5%), the false-report rate rises from 0.130 to 0.382; adding to 10 with threshold 3 (30%) rises to 0.545. Adding probes without changing the vote threshold makes the gate easier to falsely report. To maintain 60% of votes, 10 probes need ≥6 votes, bringing false-report rate back to 0.031.

    This occurs because the false-report rate of majority voting is not a monotonic function of n — it is determined by the ratio k/n and n. Adding votes without adjusting the proportion relaxes the criterion.

    IV. F3’s Steps Are Inverted: Heavier Conclusions, Thinner Evidence

    Once F3 triggers, the conclusion drops from E3 to E2 and “must explain disturbance sensitivity” — this is the most severe consequence in the whole card. Yet its sample size is smallest: 12 per condition.

    Δr’s standard error is 3.61 times that of F1 (n=120), while the threshold relaxes only 3.0 times (0.15 vs 0.05). To match F1’s resolution, each condition would need 3460 people (ρ=0.5), totaling 17,302 for five conditions — the current design is 60.

    Moving toward “weakest hypotheses” requires harder evidence, yet the design provides less sample. This inversion is not a typo; it results from misalignment between “falsification conditions ranked by severity” and “sample allocation by convenience.”

    V. F5: “>20%” at n=20 Actually Means 25%

    The cognitive interview uses n=20, with the rule “trigger if >20% of subjects show semantic misunderstanding.” Since 20% × 20 = 4, and “>” excludes equality, only X ≥ 5 triggers, making the actual threshold 5/20 = 25%, 1.25 times higher than stated.

    When true misunderstanding rate is exactly 20%, the probability of triggering F5 is 0.370. This is minimal at n=20; as n increases, it rises to 0.466 (n=300) — because under large n the realized threshold approaches 20%, making the criterion “more sensitive,” while 20% itself is the threshold. The discretization error of the voting threshold and statistical power move in opposite directions; adding sample size does not improve both simultaneously.

    VI. Incidentally Reviewing #22’s Three Tiers: R29 Inapplicable, Yet Another Form Emerges

    In Pool ④’s previous round, R29 “criterion one-sided automatic tightening” was established on K1–K4, and #22’s A/B/C were listed as “suspected same defect, pending check.” This round completes the calculation: #22 does not trigger R29.

    Recalculating at 1 km / 7.83 Hz:
    – Tier A’s true boundary is σ ≥ 4.283×10⁻² (A = 10 dB, D1′ recorded as 4.29×10⁻², matching);
    – Tier C’s boundary is σ ≤ 10⁻³ (A = 1 dB).
    – σ’s uncertainty scales by q:

    Tier A threshold (σ lower bound) shifts up with q: σ_A·q   — harder to enter A
    Tier C threshold (σ upper bound) shifts down with q: σ_C/q   — harder to enter C
    Tier B = (σ_C/q, σ_A·q), width ratio grows from 42.8× to 4282.9× when q=10
    

    Both sides expand outward symmetrically, so no single-sided bias occurs. However, another shape emerges:

    R29-b “Middle Tier Absorption” — judgment power does not lean to one side but collapses entirely into the middle tier. When precision worsens, Tier A and Tier C both vanish, and all stations fall into “conditional measurement.” The consequence matches R29 (always saying “not yet measured,” always saying “later”), but the mechanism differs: R29 arises from different threshold function forms causing single-sided tightening, whereas R29-b stems from both sides of the wide tier expanding outward causing global absorption.

    These two cannot substitute each other: R29’s solution is unifying the threshold function form; R29-b’s solution is eliminating the “wide tier” absorbing state or setting its upper limit.

    VII. Two New Hard Rules Added

    R30 “Conjunction Threshold Must Report Joint Power”: When multiple stop-loss gates are serialized as admission conditions, one must simultaneously report “single-gate false-kill rate → all-pass probability.” Reporting only “we have five checks” will be read as strict, while the actual pass rate is 13.7%.

    R31 “Majority Vote Threshold Must Be Recalculated with Votes”: The false-report rate of (k,n) structure is not a monotonic function of n. Adding probes without adjusting the vote ratio increases false reports; any claim “adding more probes makes it more stable” must recalculate k.

    VIII. Two Hard Flaws Found in Citation Review (Reported Truthfully)

    During review, we scanned the citations in §12 of this design and found two issues:

    Original Reality Action
    PMID 24233341 (Lehrer PM, HRV biofeedback: effects across domains, Appl Psychophysiol Biofeedback 2013) PubMed has no record (efetch returns empty set) Recommend replacing with 25101026 (Lehrer PM, Heart rate variability biofeedback: how and why does it work?, Front Psychol 2014;5:756, verified)
    PMID 39563397 (Li Z, acupuncture… functional dyspepsia, Chin Med 2024) This number actually refers to Loix M, Stearoyl-CoA desaturase-1, Mol Neurodegener 2024;19 — topic completely unrelated Recommend replacing with 30246392 (Liu Y, electroacupuncture ST36 rat dyspepsia model, Neurogastroenterol Motil 2019) or deleting this entry entirely
    PMID 34479992 (Chung JW, HRV and TCM patterns) Fake number, already registered in P-1 layer I-2/I-3 Still awaiting JOE approval; not touched this round

    This round does not change this document: it falls within the 779 intake triad scope, and citation governance is pending JOE approval. We only register it, do not execute.

    Evidence and Limitations

    All content in this paper consists of analytical derivation and binomial/normal simulations, containing no empirical data. Three boundaries must be explicitly stated:

    1. “Cannot judge” does not mean “hypothesis false.” This paper states only that “this set of gates cannot judge the answer it wants to judge,” not whether “Vata-low ≈ Qi Deficiency-low” is true or false. Per R11, zero-parameter/pure-analytic criteria bear only falsification function and cannot serve as proof of “criterion qualified.”
    2. ρ (correlation between two measurements) was not empirically measured. F2/F3’s noise entirely depends on ρ; this paper provides the range ρ = 0 to 0.9. True ρ depends on repeat measurement intervals and posture control of the same batch, which were not pre-tested.
    3. Boundary of effective advancement: This paper addresses structural issues at the criterion level, not research conclusions. Specific revisions to F1–F5 (thresholds, vote counts, sample sizes) remain at the suggestion level, pending JOE approval.

    Related Reading

    • Pool ④ #77 K1–K4 tiering system structure audit (origin of R26)
    • Pool ③ #187 ICC gate CI width pre-completion (origin of R27)
    • Pool ④ #26 six-station sequential gate passage (origin of R29)
    • Pool ⑤ #19 #20 falsification-priority methodology / preregistration norms (source of F1–F5)

    Reference Anchors

    • Task Force of ESC/NASPE. Heart rate variability: standards of measurement, physiological interpretation and clinical use. Circulation. 1996. PMID: 8598068.
    • Cohen J. A power primer. Psychol Bull. 1992. PMID: 19565683.
    • Lehrer PM, Gevirtz R. Heart rate variability biofeedback: how and why does it work? Front Psychol. 2014;5:756. PMID: 25101026.
    • Lakens D. Equivalence Tests: A Practical Primer for t Tests, Correlations, and Meta-Analyses. Soc Psychol Personal Sci. 2017. PMID: 28736600.
    • Nosek BA, Ebersole CR, DeHaven AC, Mellor DT. The preregistration revolution. Proc Natl Acad Sci USA. 2018. PMID: 29531091.
    • Koo TK, Li MY. A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research. J Chiropr Med. 2016. PMID: 27330520.
    • Chen FF. Sensitivity of Goodness of Fit Indexes to Lack of Measurement Invariance. Struct Equ Modeling. 2007. doi:10.1080/10705510701301834.

    Disclaimer: This paper is a methodological audit involving TCM constitution, Ayurvedic three constitutions, and autonomic nervous system indicators — all health-related topics. All conclusions herein are statistical criterion-level derivations and do not constitute any diagnostic, therapeutic, or constitution judgment advice; they must not be used for clinical decision-making or productization.

    review_status: draft — Not yet reviewed by the 779 intake; cannot be marked accepted.

  • 「三条判据都同意」到底是不是三个独立证据:把输入重叠度算出来,一致率就只剩一个自由度

    Evidence level: C

    三条判据读的是同一批变量、同一批人、同一台仪器的数据——它们”都同意”很可能只是同一份输入被读了三次,而不是三份独立证据指向同一个结论。

    正文

    在方法学写作里有一句话几乎是条件反射:「合成指数、逐变量诊断、比例口径三条判据都指向同一结论,所以结论稳健。」这句话读起来像证据叠加,但在大多数实际场合它不是。三条判据用的是同一批候选变量、同一批受试者、同一套设备采集的数据;如果它们的输入集合重叠八成,那”三口径互证”就不是三个证人各自作证,而是同一个人换了三套说辞。

    先看重叠度怎么算。 最直接的是 Jaccard 系数:两条判据实际用到的输入变量集合,交集除以并集。这是集合相似度的标准量,在特征选择稳定性评价里被当作可复现性的代理——文献里常见用法是拿它比较不同重抽样下选出的特征集重合程度,分布集中在 1 附近才算稳定(PMID 33375940 的多视图特征选择即以此类集合相似度评估所选标记的非冗余性)。但 Jaccard 有个已知的坑:它对集合大小敏感——这一点在特征选择文献里被明确点出为”list-size bias”,即两条清单越接近特征全集,越容易凑出高重叠(PMID 16077740 同族的有效检验数思路正是为了处理这种”看似很多、实则高度相关”的冗余)。所以 Jaccard 只能当重叠的下界参照,不能单独当独立性证据。

    再看重叠度能不能直接换算成一致率。 可以做一步自算,不需要任何额外假设之外的东西。设两条判据的输入共享份额为 f,把它看成裁决分数的方差里共享的比例(即相关系数 ρ=f);裁决被压成二值(保留/降级)时,两条独立正态、相关 ρ 的分数符号相同的概率是:

    一致率下界 = 1 − arccos(ρ)/π
    

    把 Jaccard 换成 f 需要一步:等大小两集合,交集份额 f = 2J/(1+J)。代进去得到一张表:

    输入 Jaccard 共享方差 f = 2J/(1+J) 二值裁决一致率下界
    0.20 0.333 60.8%
    0.40 0.571 69.4%
    0.60 0.750 77.0%
    0.70 0.824 80.8%
    0.80 0.889 84.9%
    0.90 0.947 89.6%

    反过来解更刺眼:要让三条判据的一致率达到 90%,输入 Jaccard 得高到 0.907;要到 95%,J 得 0.976。也就是说,只要两条判据的输入集合有一成不同,它们的一致率在结构上就已经上不了 90%。反过来说,若观察到一致率 95% 以上,先不要庆祝——几乎可以肯定它的输入集合是同一个。

    第二个量是有效独立信息条数。 这才是”是不是三个证据”的直接答案。三种算法都给同一个方向的答案:

    • Cheverud 型(用相关矩阵特征值方差):特征值越分散,独立检验数越少(PMID 11678987 在区间定位基因组扫描里提出这条思路);
    • Li–Ji 型(特征值超 1 的部分抵扣掉):M_eff = m − Σ(λᵢ−1),只要 λᵢ>1(PMID 16077740 指出 Cheverud 的估计偏大、偏保守,给出了更准的版本);
    • 特征值熵(Hill 数):exp(−Σpᵢln pᵢ),pᵢ 为归一化特征值。

    三条判据等相相关 ρ 时,特征值是 1+2ρ、1−ρ、1−ρ。代进去:

    判据间相关 ρ Li–Ji 有效条数 特征值熵有效条数
    0.00 3.000 3.000
    0.30 2.400 2.758
    0.50 2.000 2.381
    0.70 1.600 1.895
    0.90 1.200 1.338
    0.99 1.020 1.046

    ρ = 0.75 时 Li–Ji 给 1.50——三条判据合起来不到两条独立信息;ρ=0.8 时是 1.4。而”输入重叠八成”在 Jaccard 上是 0.7、换成 ρ 是 0.82,已经落在这条线上。换句话说:只要输入 Jaccard 过 0.7,「三条判据互证」的独立条数就已经跌破 2。

    这件事的后果不是修辞问题。 一个具体例子:三条判据各自的输入集合是 12、14、12 个变量,交集 10 个——J(合成指数, 逐变量诊断) = 10/14 ≈ 0.714。按上表,这两条的一致率下界是 81%,有效独立条数约 1.4。此时若真观察到 92% 的一致率,多出来的 11 个百分点不能算作”稳健性证据”,它只说明两条判据的裁决规则形式(阈值设在哪、权重怎么给)也高度雷同——那已经不是输入冗余,是规则冗余。

    真正值得追查的,恰恰是重叠 >0.9 却仍出现 >15% 不一致的情形。 这时输入几乎完全一样,分歧只能来自规则形式:一条用了标准化均值差 0.1/0.25 双阈值,另一条用了 20%/30% 的结论成立比例。这种不一致才是信息——它暴露的是阈值选择而非数据,应当被单独写成敏感性分析,而不是被”三口径”的多数票抹平。

    方法论上还能再退一步。 与其先算三条判据再比较,不如先把输入集合的冗余结构清出来:算相关矩阵的特征值谱,看”12 个候选变量实际能撑起几个自由度”。若谱显示只有一个大特征值(λ₁ 占了 70% 以上方差),那么不论后面写几条判据,它们读的都是那一个方向;此时正确的报告方式不是”三口径一致”,而是”该批数据仅含约 1.4 个独立信息方向,结论的不确定度未因多口径而收窄“。

    要注意这里有个不能越的界。 上面的换算把裁决分数当成了等方差正态、把共享份额当成相关系数——这是最粗的一阶近似,真实规则里的阈值截断、加权方式、非线性变换都会让它偏。所以表里的数字应当读成量级判据(”八成重叠 → 独立条数不到两条”),而不是精确预测(”一致率一定等于 80.8%”)。这也正是本文只能标 C 的原因:换算链条是自算的解析结果,但整条”判据独立性”的边在 L1 对应网络里根本没有登记位。

    最后一句直白的话:三条判据都同意,只有在你能证明它们读的不是同一份输入时才算证据。证明方式很简单——把三条判据实际用到的变量列出来,算 Jaccard,算特征值。如果 J 过 0.7,就把”三口径互证”这四个字删掉,改成报告”约 1.4 条独立信息下的单一结论”。前者是修辞,后者是可以被检验的陈述。

    证据与局限

    已验证: Jaccard 系数作为集合重叠度量、有效独立检验数(特征值法)作为相关多重检验的校正量,两者在生物信息与统计遗传学里都是成熟工具(PMID 11678987、PMID 16077740、PMID 33375940)。多重共线性会膨胀系数方差、使推断不可靠,也是既成结论(PMID 31304696);把 VIF 直接接入变量选择流程已有实践(PMID 34896682)。

    部分验证: “输入重叠 → 裁决相关 → 一致率”这条换算链在本文是按等方差正态、二值裁决的最简模型自算的;共享份额当作相关系数这一步只在一阶近似下成立,未在真实判据规则上做过仿真标定。Jaccard 本身对集合大小的偏倚是已知问题,本文未做大小校正(如相对加权一致性)。

    待验证: ① 三条判据在真实丹田数据集上的输入集合到底是什么,交集有几个变量——从未清点过;② 按 30%/60%/90% 三档人为构造重叠后,一致率曲线的实测斜率与本文解析下界差多少;③ 输入重叠 >0.9 却仍出现 >15% 不一致时,分歧是否确实由阈值设定而非数据驱动。

    相关阅读

    证据等级: C
    参考文献: PMID 11678987, PMID 16077740, PMID 31304696, PMID 34896682, PMID 33375940, PMID 30788363, PMID 15066682, PMID 41626969

    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「呼吸率 / 呼吸模式 | 可穿戴 | 与节律耦合最直接」),全表检索「判据 / 裁决 / 一致率 / Jaccard / 重叠 / 有效独立」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 概念计数无「判据独立性 / 信息重叠 / 有效检验数」条目。故本文证据等级封顶 C,不得读作已验证机制。

    本文自算说明:文中所有 Jaccard→一致率换算(f = 2J/(1+J);一致率 = 1 − arccos(ρ)/π)、有效独立条数(Li–Ji 型 M_eff = m − Σ(λᵢ−1);特征值熵 exp(−Σpᵢln pᵢ))、三判据特征值谱(1+2ρ, 1−ρ, 1−ρ)均为按定义直接推导所得,非引自文献;文献仅用于支撑方法本身的合法性与已有实践。


    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。

    Do the three criteria agree: calculating input overlap leaves only one degree of freedom for consistency.

    Evidence level: C

    “Three criteria agreeing” likely means the same input was read three times, not three independent pieces of evidence pointing to the same conclusion.

    Body

    In methodological writing, there is almost a reflexive phrase: “Synthetic index, variable-by-variable diagnosis, and proportion caliber—all three criteria point to the same conclusion, so the conclusion is robust.” This sounds like evidence stacking, but in most practical cases it is not. The three criteria use the same candidate variables, the same subjects, and data collected by the same equipment; if their input sets overlap by 80%, then “three-way cross-verification” is not three witnesses testifying separately, but one person changing three different narratives.

    First, how to calculate overlap. The most direct is the Jaccard coefficient: the intersection of the actual input variable sets used by two criteria divided by their union. This is the standard measure of set similarity and serves as a proxy for reproducibility in feature selection stability evaluation—commonly used in literature to compare the overlap of feature sets selected under different resampling schemes, where a distribution centered near 1 indicates stability (PMID 33375940 uses such set similarity to assess the non-redundancy of selected markers in multi-view feature selection). However, Jaccard has a known pitfall: it is sensitive to set size—a point explicitly noted in feature selection literature as “list-size bias,” meaning the closer two lists are to the full feature set, the easier it is to achieve high overlap (PMID 16077740’s concept of valid test numbers for the same family precisely addresses this “appears numerous but actually highly correlated” redundancy). Therefore, Jaccard should only serve as a lower-bound reference for overlap, not as sole evidence of independence.

    Next, can overlap directly convert to agreement rate? One step of self-calculation suffices, requiring nothing beyond the stated assumptions. Let f be the shared proportion of inputs between two criteria, treated as the fraction of variance in decision scores attributable to shared components (i.e., correlation coefficient ρ = f). When decisions are reduced to binary (retain/degrade), the probability that two independent normal scores with correlation ρ have the same sign is:

    Lower bound of agreement rate = 1 − arccos(ρ)/π
    

    Converting Jaccard to f requires one step: for two sets of equal size, the shared fraction f = 2J/(1+J). Substituting yields this table:

    Input Jaccard Shared variance f = 2J/(1+J) Lower bound of binary decision agreement rate
    0.20 0.333 60.8%
    0.40 0.571 69.4%
    0.60 0.750 77.0%
    0.70 0.824 80.8%
    0.80 0.889 84.9%
    0.90 0.947 89.6%

    Working backwards is even more striking: to achieve 90% agreement among three criteria, input Jaccard must reach 0.907; for 95%, J must be 0.976. That is, if two criteria’ input sets differ by just 10%, their structural agreement rate cannot exceed 80.8%. Conversely, observing an agreement rate above 95% should not be celebrated—it almost certainly means the input sets are identical.

    The second quantity is the number of effective independent information streams. This is the direct answer to “are there three pieces of evidence?” All three algorithms point in the same direction:

    • Cheverud-type (using eigenvalue variance of the correlation matrix): the more dispersed the eigenvalues, the fewer independent test numbers (PMID 11678987 proposes this approach for interval-based genome-wide association studies);
    • Li–Ji-type (subtracting eigenvalues exceeding 1): M_eff = m − Σ(λᵢ−1), provided λᵢ>1 (PMID 16077740 notes Cheverud’s estimate is biased upward and conservative, offering a more accurate version);
    • Eigenvalue entropy (Hill number): exp(−Σpᵢln pᵢ), where pᵢ are normalized eigenvalues.

    When all three criteria have equal correlation ρ, the eigenvalues are 1+2ρ, 1−ρ, and 1−ρ. Substituting yields:

    Correlation ρ between criteria Li–Ji effective count Eigenvalue entropy effective count
    0.00 3.000 3.000
    0.30 2.400 2.758
    0.50 2.000 2.381
    0.70 1.600 1.895
    0.90 1.200 1.338
    0.99 1.020 1.046

    At ρ = 0.75, Li–Ji gives 1.50—three criteria together provide less than two independent pieces of information; at ρ = 0.8, it drops to 1.4. Meanwhile, “input overlap at 80%” corresponds to Jaccard = 0.7, which translates to ρ = 0.82, placing it right on this line. In short: once input Jaccard exceeds 0.7, the independent count for “three-way cross-verification” falls below 2.

    The consequence is not merely rhetorical. A concrete example: three criteria have input sets of 12, 14, and 12 variables respectively, with an intersection of 10—yielding J(synthetic index, variable-by-variable diagnosis) = 10/14 ≈ 0.714. From the table above, the lower bound of agreement between these two is 81%, and the effective independent count is approximately 1.4. If we truly observe 92% agreement, the extra 11 percentage points cannot be counted as “robustness evidence”; it merely indicates that the decision rules themselves (where thresholds are set, how weights are assigned) are also highly similar—this is no longer input redundancy, but rule redundancy.

    What truly deserves investigation is the case where overlap >0.9 yet disagreement remains >15%. Here, inputs are nearly identical, so divergence must stem from rule structure: one uses dual thresholds of standardized mean difference 0.1/0.25, while another employs a conclusion-valid proportion of 20%/30%. This disagreement is the actual information—it exposes threshold selection rather than data, and should be reported separately as a sensitivity analysis, not smoothed out by “three-way majority voting.”

    Methodologically, we can go further. Instead of computing three criteria first and then comparing, clear the redundant structure of input sets upfront: calculate the eigenvalue spectrum of the correlation matrix to see “how many degrees of freedom can 12 candidate variables actually support.” If the spectrum shows only one dominant eigenvalue (λ₁ accounting for over 70% of total variance), then regardless of how many criteria are written, they all read the same direction; the correct reporting style is not “three-way agreement,” but “this dataset contains only about 1.4 independent information directions, and uncertainty did not shrink due to multiple calibers.”

    Note there is a boundary that cannot be crossed. The above conversions treat decision scores as equal-variance normal distributions and shared fraction as the correlation coefficient—the coarsest first-order approximation. Real-world rules involving threshold truncation, weighting schemes, and nonlinear transformations will introduce bias. Thus, numbers in the table should be read as order-of-magnitude indicators (“80% overlap → fewer than two independent streams”), not precise predictions (“agreement rate exactly equals 80.8%”). This is precisely why this paper can only be labeled C: the conversion chain derives analytically, but the entire “criteria independence” edge lacks registration in the L1 corresponding network.

    The final plain statement: three criteria agreeing counts as evidence only if you can prove they did not read the same input. The proof is simple—list the variables actually used by each criterion, compute Jaccard, and calculate eigenvalues. If J exceeds 0.7, delete “three-way cross-verification” and report instead “a single conclusion under approximately 1.4 independent information streams.” The former is rhetoric; the latter is a testable claim.

    Evidence and Limitations

    Verified: Jaccard coefficient as a set overlap measure, and effective independent test counts (eigenvalue methods) as corrections for correlated multiple testing, are both mature tools in bioinformatics and statistical genetics (PMID 11678987, PMID 16077740, PMID 33375940). Multicollinearity inflates coefficient variance and makes inference unreliable, which is an established conclusion (PMID 31304696); integrating VIF directly into variable selection workflows has practical precedent (PMID 34896682).

    Partially verified: The conversion chain “input overlap → decision correlation → agreement rate” in this paper is derived via the simplest model assuming equal-variance normal distributions and binary decisions; treating shared fraction as the correlation coefficient holds only under first-order approximation, without simulation calibration against real-world criterion rules. Jaccard’s inherent sensitivity to set size is a known issue, and this paper does not perform size correction (e.g., relative-weighted agreement).

    To be verified: ① what exactly are the input sets of three criteria on real Dantian datasets, and how many variables lie in their intersection—never counted; ② after artificially constructing overlap at 30%/60%/90%, how much does the measured slope of the agreement rate curve deviate from this paper’s analytical lower bound; ③ when input overlap exceeds 0.9 yet disagreement remains above 15%, is the divergence truly driven by threshold settings rather than data.

    Related Reading

    Evidence level: C
    References: PMID 11678987, PMID 16077740, PMID 31304696, PMID 34896682, PMID 33375940, PMID 30788363, PMID 15066682, PMID 41626969

    L1 Network Status: This topic has not yet entered the L1 corresponding network. The three-domain node table in CORRESPONDENCE_NETWORK.md registers only single-point signals (physical side “phase/rhythm | — | periodic coupling key”, physiological side “breathing rate / breathing pattern | wearable | most directly coupled with rhythm”), and full-table search for “criteria / decision / agreement rate / Jaccard / overlap / effective independent” 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”); EDGE_TABLE concept count lacks entries for “criteria independence / information overlap / effective test numbers.” Hence this paper’s evidence level caps at C and cannot be read as verified mechanism.

    Self-calculation explanation: All Jaccard→agreement rate conversions (f = 2J/(1+J); agreement rate = 1 − arccos(ρ)/π), effective independent counts (Li–Ji type M_eff = m − Σ(λᵢ−1); eigenvalue entropy exp(−Σpᵢln pᵢ)), and three-criteria eigenvalue spectra (1+2ρ, 1−ρ, 1−ρ) are derived directly by definition, not cited from literature; literature serves only to support the legitimacy of the methods and existing practice.


    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.

  • 节律到底是大家都在抖,还是各抖各的:把群体均值被相位差冲淡多少算清楚,顺手废掉一条恒真的判据

    Evidence level: C

    扣不扣群体节律曲线,取决于一件从未被测过的事:人和人之间差的是”同一时刻抖同样的幅度”,还是”同一幅度抖在不同的时刻”。前一种情况扣掉等于降噪,后一种情况扣掉等于把签名削掉。

    正文

    前几篇把整条校正逻辑的命门压在一个从未验证的前提上:节律被当作”全民共享的共同噪声”。如果它真是共同噪声,扣掉它等于纯降噪,个体识别力必涨;如果每个人自己的相位(早起型、夜猫型)和幅度(季节反应强弱)差异很大,扣群体平均曲线就等于把个体签名的一部分一起削了。两种情况都能让重测一致性指标 ICC 上涨,所以光看指标永远分不出来——必须做方差成分分解。

    一、先把”个体成分占比”这个提法换掉

    直觉上想说:把每个人 24 小时曲线的方差拆成”群体共享部分”和”个体部分”。但直接这么算会踩空——单个个体的曲线方差不随相位偏移改变。一个人凌晨 3 点血压最低、一个人 5 点最低,两条曲线的方差是一样大的。真正的差异只出现在一个地方:把一群人平均起来的时候。

    所以正确的统计量不是”个体方差里有多少是我的”,而是”群体平均曲线还剩多少振幅”。设第 k 阶谐波、个体相位偏移标准差 σ(小时)、角频率 ω=2π/24,平均之后的振幅保留比例是

    S_k = exp( −(k·ω·σ)² )
    

    这是相位去相干(phase decoherence)的直接后果,形式上是特征函数的模。指数里的 k 是关键:阶次越高,同样的相位差被放大得越厉害。k=1 差 1 小时,相当于 k=6 差 6 小时。十万次蒙特卡洛与闭式的偏差在 0.003 以内,这条式子可以放心用。

    σ=1.43 小时时,逐阶的群体共享占比是:k=1 剩 87%,k=2 剩 57%,k=3 剩 28%,k=4 剩 11%,k=5 剩 3%,k=6 剩 0.6%。也就是说,到第三阶,群体曲线已经冲淡了七成;到第五阶,冲淡到几乎没有。

    二、σ 到底多大:从已经测出来的数据反解

    σ 不是拍出来的。褪黑素起始时间(DLMO)是昼夜相位最有共识的标记物,一项跨约 100 天三次测量的研究给出组内 ICC=0.85、个体内标准差约 0.6 小时(PMID 34396981 的系统综述也确认相位个体差异大且模型普遍低估)。用 ICC=σ²真/(σ²真+σ²误差) 反解:

    研究给出的 ICC 反解出的真 σ k=1 共享 k=3 共享 k=6 共享
    0.85 1.43 h 87% 28% 0.6%
    0.75 1.04 h 93% 51% 7%
    0.67 0.85 h 95% 64% 17%

    即便取最保守的 ICC=0.67,k=4 以上也已经跌破一半。“节律是共同噪声”这个前提,在低阶勉强站得住,在高阶站不住。

    三、顺手废掉一条恒真的判据

    原定的判据里有一条:若个体成分占比随阶次单调上升,就”独立印证了低阶身份层、高阶状态层的分层”。看到上面的式子就明白——S_k 是 k² 的指数衰减,单调性是这个函数形式的几何必然,跟数据没关系。任何有相位差的群体,逐阶共享占比都会单调下降。这条判据无论数据长什么样都会通过,它不能当作任何证据使用。

    这不是小修小补。它意味着”低阶身份层/高阶状态层”这个分层,至今没有被独立证据支撑过,只是被一个数学恒等式解释掉了。要真正检验分层,得换判据——比如看个体相位偏移本身是否随阶次变化(相位谱是否线性),而不是看共享占比的单调性。

    四、还有一道更硬的墙:采样点

    真实数据里没人能在 24 小时内采上百个点。按奈奎斯特,每周期 n 个点最多分辨到 k=(n−1)/2 阶:n=6 点只能到 k=2,n=8 到 k=3,n=13 才到 k=6。而 σ=1.43 小时时 k=6 的共享成分只有 0.6%——低于任何实测精度。换句话说,即使判据说”高阶该个体化”,多数研究的采样密度也测不到那个阶次。结论应该收敛为:在常规采样(每日 6–13 点)下,可测阶次里只有 k=1 到 k=3 需要认真讨论群体校正与个体化校正的选择,k≥4 的问题是工程问题不是科学问题。

    综合起来,这条边目前只能给出方向,给不出定论:低阶共享成分占优、扣群体曲线大体是降噪;中高阶个体成分占优、群体校正会削掉真信号。但具体切在第几阶,取决于 σ 的真实值、也取决于采样密度,而这两件事在脉搏谐波数据上都还没有实测过。

    证据与局限

    已验证:24 小时生理节律存在明确的个体相位差异,DLMO 组内 ICC 在 0.67–0.85 区间(PMID 34396981、PMID 28452285);晨型与晚型的核心钟基因表达相位差约 2.6–3.1 小时,可直接测出(PMID 27103930);个体相位可由光暴露模式预测,DLMO 预测标准差约 1.1 小时(PMID 28452285)。

    部分验证:相位个体差异的幅度可被光敏感度与内禀周期共同解释,模型参数化后误差从 1.02 小时降到 0.28 小时(PMID 33063595);睡眠时相与褪黑素相位的相位角关系在人群内保持约 2 小时但方差随作息延后而扩大(PMID 21188265)。这些支持”σ 是真实、可解释的个体参数”,但都在昼夜相位标记物上,不在脉搏谐波上。

    待验证:① σ 在脉搏谐波各阶上是否与 DLMO 的 σ 同量级,完全未测;② S_k=exp(−(kωσ)²) 的相位去相干模型假定相位偏差服从正态且与振幅独立,若存在晨晚型双峰分布,尾部会改变衰减速度;③ 群体均值的方差低估会连带影响所有基于它的标准化,本文只讨论校正取舍,未讨论标准化的连带误差。

    L1 网络状态:在 CORRESPONDENCE_NETWORK.md 中,相位 / 节律 仅作为「物理侧节点」列出(备注”周期性耦合的关键”),边表中无任何一条以节律、昼夜、相位变异或个体成分占比为源的边;EDGE_TABLE_science-yixue.md 概念表中 无”节律””昼夜””生物钟”条目(最接近的 时辰 有 42 篇、最高 A 级 9 篇,但指的是时辰关联而非节律方差结构),且该表未包含任何 L2/L3 层级边。故本主题尚未进入 L1 对应网络,证据等级封顶 C。

    相关阅读

    证据等级: C
    参考文献: PMID 34396981, PMID 28452285, PMID 27103930, PMID 33063595, PMID 21188265


    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。

    Rhythm is either collective shaking or individual shaking: quantify how much the group mean is diluted by phase differences and simultaneously invalidate a constant-true criterion.

    Evidence level: C

    Whether to subtract the group rhythm curve depends on one thing never yet tested: whether people differ in “shaking with the same amplitude at the same moment” or “shaking with the same amplitude but at different moments.” Subtracting in the first case equals noise reduction; subtracting in the second case equals shaving off the signature.

    Main Text

    Previous articles pinned the crux of the entire correction logic on an unverified premise: rhythm is treated as “shared common noise.” If it truly is shared noise, subtracting it equals pure denoising, and individual discriminability must rise; if each person’s own phase (early riser vs. night owl) and amplitude (strength of seasonal response) differ greatly, subtracting the group average curve equals shaving off part of the individual signature. Both scenarios can cause the intraclass correlation coefficient (ICC) for retest consistency to rise, so looking at the metric alone never distinguishes them—variance component decomposition is required.

    I. First, Replace the Term “Individual Component Proportion”

    Intuitively, one wants to decompose the variance of each person’s 24-hour curve into “group-shared part” and “individual part.” But calculating it directly leads to a gap—the variance of a single individual’s curve does not change with phase shift. If one person’s blood pressure is lowest at 3 AM and another’s at 5 AM, the variance of both curves is identical. The true difference appears only in one place: when averaging a group.

    Therefore, the correct statistic is not “what fraction of individual variance is mine,” but rather “how much amplitude remains in the group average curve.” Let $k$ be the harmonic order, $\sigma$ (hours) the standard deviation of individual phase shifts, and angular frequency $\omega = 2\pi/24$. The proportion of amplitude retained after averaging is

    S_k = exp( −(k·ω·σ)² )
    

    This is the direct consequence of phase decoherence, formally the modulus of the characteristic function. The key in the exponent is $k$: the higher the order, the more any given phase difference is amplified. A 1-hour difference at $k=1$ is equivalent to a 6-hour difference at $k=6$. Monte Carlo simulations of 100,000 runs and closed-form deviations stay within 0.003, so this formula can be used with confidence.

    When $\sigma = 1.43$ hours, the proportion of group-shared content for each harmonic is: $k=1$ retains 87%, $k=2$ retains 57%, $k=3$ retains 28%, $k=4$ retains 11%, $k=5$ retains 3%, and $k=6$ retains 0.6%. In other words, by the third harmonic, the group curve has washed out 70%; by the fifth harmonic, it is almost completely washed out.

    II. How Large is $\sigma$? Inverting from Already Measured Data

    $\sigma$ is not guessed. Dim light melatonin onset (DLMO)—the most traditional consensus marker of circadian phase—was measured in a cross-study of about 100 days with three measurements, yielding an intragroup ICC = 0.85 and an individual standard deviation of approximately 0.6 hours (PMID 34396981’s systematic review also confirms large individual phase differences and that models generally underestimate them). Inverting using ICC = $\sigma^2_{\text{true}} / (\sigma^2_{\text{true}} + \sigma^2_{\text{error}})$:

    Study ICC Inverted True σ k=1 Shared k=3 Shared k=6 Shared
    0.85 1.43 h 87% 28% 0.6%
    0.75 1.04 h 93% 51% 7%
    0.67 0.85 h 95% 64% 17%

    Even taking the most conservative ICC = 0.67, harmonics $k \geq 4$ already drop below half. “Rhythm is shared noise” holds barely at low orders but fails at high orders.

    III. Incidentally Invalidating a “Always True” Criterion

    Among the original criteria was this one: if the individual component proportion rises monotonically with harmonic order, it “independently validates stratification of low-order identity layers and high-order state layers.” Looking at the formula above makes this clear—$S_k$ is an exponential decay of $k^2$, so monotonicity is a geometric inevitability of this function form, unrelated to data. For any group with phase differences, the shared proportion will monotonically decrease with each harmonic. This criterion passes regardless of what the data looks like; it cannot serve as evidence.

    This is not a minor fix. It means the stratification of “low-order identity layer / high-order state layer” has, to date, never been independently supported by evidence, merely explained away by a mathematical identity. To truly test stratification, one must change the criterion—e.g., examine whether individual phase shift itself changes with harmonic order (is the phase spectrum linear), rather than whether shared proportion is monotonic.

    IV. One More Hard Wall: Sampling Points

    In real data, no one can collect hundreds of points within 24 hours. By Nyquist, with $n$ points per cycle, the maximum resolvable harmonic is $k = (n-1)/2$: $n=6$ points reach only $k=2$, $n=8$ reaches $k=3$, and $n=13$ reaches $k=6$. But when $\sigma = 1.43$ hours, the shared component at $k=6$ is only 0.6%—below any measurable precision. In other words, even if the criterion says “higher orders should be individualized,” most studies’ sampling density cannot detect that harmonic order. The conclusion must converge to: under conventional sampling (6–13 points per day), only harmonics $k=1$ to $k=3$ require serious discussion of group vs. individual correction; the issue for $k \geq 4$ is an engineering problem, not a scientific one.

    In summary, this edge currently provides only direction, not conclusion: low-order shared components dominate, so subtracting the group curve is largely denoising; mid-to-high order individual components dominate, so group correction will shave off true signal. But exactly where to cut depends on the true value of $\sigma$ and sampling density, both of which have not yet been measured in pulse harmonic data.

    Evidence and Limitations

    Verified: 24-hour physiological rhythms exhibit clear individual phase differences; DLMO intragroup ICC falls in the 0.67–0.85 range (PMID 34396981, PMID 28452285); core clock gene expression phase differences between early and late types are about 2.6–3.1 hours, directly measurable (PMID 27103930); individual phase can be predicted by light exposure patterns, with DLMO prediction standard deviation around 1.1 hours (PMID 28452285).

    Partially Verified: The magnitude of individual phase differences can be explained jointly by light sensitivity and intrinsic period; after model parameterization, error drops from 1.02 hours to 0.28 hours (PMID 33063595); the phase angle relationship between sleep phase and melatonin phase remains about 2 hours within populations but variance expands with delayed schedules (PMID 21188265). These support “σ is a real, interpretable individual parameter,” but they apply to circadian phase markers, not to pulse harmonics.

    To Be Verified: ① Whether σ at each harmonic order of pulse harmonics matches the DLMO σ in magnitude, completely untested; ② Whether the phase decoherence model $S_k = \exp(−(k\omega\sigma)^2)$ assumes phase deviation follows a normal distribution and is independent of amplitude—if bimodal distributions exist (early/late types), tails will alter decay speed; ③ Underestimation of group mean variance will propagate to all standardizations based on it; this paper discusses only correction trade-offs, not the compounded error from standardization.

    L1 Network Status: In CORRESPONDENCE_NETWORK.md, phase / rhythm is listed only as a “physical-side node” (note: “key for periodic coupling”), and no edge in the table originates from rhythm, circadian, phase variation, or individual component proportion; the EDGE_TABLE_science-yixue.md concept table lacks entries for “rhythm,” “circadian,” or “biological clock” (the closest 时辰 has 42 papers, highest A-level 9, but refers to time-point association, not rhythm variance structure), and the table contains no L2/L3 level edges. Thus this topic has not yet entered the L1 corresponding network, and evidence level caps at C.

    Further Reading

    Evidence Level: C
    References: PMID 34396981, PMID 28452285, PMID 27103930, PMID 33063595, PMID 21188265


    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.

  • 「固定离散度、只换形状」这个实验设计本身是矛盾的

    Evidence level: C

    离散度不是形状旁边的另一个旋钮,它是形状的积分结果。所以「固定 SD、只换形状」不是控制变量,是在不同的可行域边界上做比较——而这两件事在数学上根本不是一回事。

    正文

    先说结论:「形状影响高阶谐波」这个命题,从来没能被干净地检验过,因为过去号称检验它的设计都是无效的。 但命题本身没死——它只要求换一套坐标系。

    一、为什么「固定 σ」这个操作是自相矛盾的

    一天里相位落在哪个时刻是一个分布,两个常见描述:离散度 σ(大家几点醒的散得多开)与形状(单峰还是早晚双峰)。直觉上这是两个独立旋钮:σ 拧到 1.43 小时不动,再换形状试试。问题是 σ 是形状的函数,不是形状的同伴。

    拿最典型的双峰来算。两峰放在 ±d、权重各 0.5,则 σ = d。要固定 σ = 1.43 h,峰距就被唯一锁死在 2.86 h,一个数都不能动。想让峰距变到 3.5 h,σ 必须跟着变成 1.75 h。想保持 σ 却拉大峰距,唯一办法是把权重改成不对称(如 0.9 / 0.1),此时 σ = d/0.6,峰距可放到 4.77 h——但偏度同时飙到 +2.67,它已经不是「双峰」而是「主峰加一小撮离群」。

    所以固定 σ 的切片里,形状只剩一维自由度,且这一维还被 σ 卡着。这不是控制变量,是在一条被约束死的曲线上滑动。

    二、真算一遍:形状的净效应到底多大

    我按六族(正态、拉普拉斯、均匀、等权双峰、90-10 双峰、三峰 60-20-20)在 σ 全部等于 1.43 h 的条件下逐阶算了共享份额 S_k(相位合向量长度,取值 0 到 1):

    族 σ 超峰度 S₁ S₂ S₃ S₄ S₅ S₆
    正态 1.430 0.00 93.23 75.55 53.22 32.59 17.34 8.02
    拉普拉斯 1.430 +2.96 93.45 78.11 61.32 47.14 36.34 28.39
    均匀 1.430 −1.20 93.14 74.23 47.84 20.07 3.10 17.51
    三峰 60-20-20 1.430 −0.50 93.19 75.09 51.86 31.39 20.66 23.32
    等权双峰 1.430 −2.00 93.07 73.25 43.29 7.32 29.65 62.52
    90-10 双峰 1.430 +5.11 93.65 82.23 81.95 93.25 99.99 94.05

    (单位:%,σ 单位:小时;相位以 24 h 为周期,k = 1…6)

    看跨度:k=1 时六族只差 0.58 个百分点,k=2 差 8.98,k=3 差 38.67,k=4 差 85.93,k=5 差 96.89,k=6 差 86.02。

    这一组数字把一条老命题彻底打掉:「S_k 只由 σ 决定、形状效应可忽略」是错的,且在 k≥4 错得离谱——同样 1.43 小时的离散度,S₄ 可以是 7.3%(等权双峰),也可以是 93.3%(90-10 双峰),相差 12.7 倍。任何「按 σ 对齐」的规则会把这两个世界判成同一个人群。

    三、但分层是真的,只是分界线画错了

    把族和 σ 一起做二维方差分解(6 族 × 5 个 σ 档),得到:

    阶 族主效应 σ 主效应 该阶总跨度
    k=1 0.8% 98.6% 10.1 pp
    k=2 14.6% 76.0% 38.8 pp
    k=3 43.4% 38.2% 95.4 pp
    k=4 69.9% 13.7% 97.6 pp
    k=5 72.0% 2.6% 100.0 pp
    k=6 69.4% 2.2% 100.0 pp

    分层确实存在——低阶看离散度、高阶看形状。但交叉点在 k=3,不在 k=4。

    再细一层:k=1 到 k=4,六个族的排序一字不差都是「90-10 双峰 > 拉普拉斯 > 正态 > 三峰 > 均匀 > 等权双峰」,与超峰度排序完全相同;k=5 才第一次翻转,k=6 再翻一次。

    这说明两件事。第一,中低阶(k≤4)的形状效应不是一团乱麻,它就是一个数:峰度。 峰度大(尾巴重、有离群)→ S_k 高;峰度小(平顶、双峰)→ S_k 低。所以「低阶看形状」可精确化为「低阶看峰度」。第二,k≥5 后峰度就不够用了——那时起作用的是「尾巴具体长在哪几个位置」这种结构信息,压成一个数就丢了。

    四、一个反直觉的发现:σ 的效应方向不恒定

    通常说法是「离散度越大、谐波衰减越快」,实测只在最低两阶成立。六族在 σ 从 1.0 走到 2.0 时 S_k 的变化量:k=1、k=2 六族全部为负(方向一致);k=3、k=4 时 90-10 双峰反转为正(+13.9、+17.2 个百分点);k=5、k=6 时等权双峰的 S₅ 涨 60.7、S₆ 涨 100.0,三峰族的 S₆ 涨 41.7。

    原因是几何的:σ 变大时双峰峰距 d 跟着变大,当 2d 恰好凑到某阶周期的整数倍附近,该阶合向量就重新对齐。凑上与否纯看 σ 与阶次的比例,与「离散」本身没有单调关系。所以「σ 越大越不共享」只是一条对单峰族、且只对 k≤2 成立的局部规则,不能当通则引用。

    五、好消息:σ 其实可以干净地拿回来

    既然形状会污染高阶,σ 能不能单独估出来?能。从一个数字——S₁——按正态反解:σ_implied = √(−2 ln S₁) / ω,ω = 2π/24。六族在真实 σ=1.43 下的偏差分别是:正态 −0.0%、拉普拉斯 −1.7%、均匀 +0.7%、三峰 +0.3%、等权双峰 +1.2%、90-10 双峰 −3.3%。

    最大偏差 3.3%。 S₁ 是相当忠实的 σ 估计器,形状对它的污染很小——因为 S₁ 只看一阶,形状细节在这个尺度上还没展开。这与第二节不矛盾,只是分工:σ 从 S₁ 拿,形状从高阶拿。

    六、正确的坐标系:形状残差谱

    于是「固定 σ 换形状」应被替换成:先从 S₁ 反解 σ,再用该 σ 下的正态分布预测各阶,最后把实测各阶除以预测值,得到无量纲的残差谱 R_k:

    族 σ_implied R₂ R₃ R₄ R₅ R₆
    正态 1.430 1.00 1.00 1.00 1.00 1.00
    拉普拉斯 1.406 1.02 1.13 1.39 1.98 3.25
    均匀 1.440 0.99 0.91 0.63 0.18 2.26
    三峰 60-20-20 1.434 1.00 0.98 0.97 1.20 2.95
    等权双峰 1.447 0.98 0.83 0.23 1.78 8.28
    90-10 双峰 1.383 1.07 1.48 2.66 5.15 9.96

    正态族恒等于 1(分母就是它自己),所有偏离都归给形状。判别力在蒙特卡洛下相当强:σ 固定 1.43、n=200 时用 R₆ 区分正态与等权双峰,正态族 1.15 ± 0.52、双峰族 8.26 ± 0.11,几乎不重叠。但有个必须写下来的陷阱:n 小时 R_k 会被系统性抬高。n=60 时正态族的 R₆ 中位数已是 1.49(四分位距 0.97–2.20),n=200 回落到 1.15,n=480 为 1.10。因为分母(正态第 6 阶预测值约 8%)本身很小,估计误差在除法里被放大。「R₆ > 1.5 就判有结构」这条线在 n=60 时会把约两成正态人群误判**,安全阈值应设在 R₆ > 4 并同时报告 n。

    七、这件事在真人身上意味着什么

    2026 年一项基于居家唾液褪黑素起始时刻(DLMO)的临床队列研究(PMID 42392022)对 253 人分型,「不规则 / 多峰」占 43 例(17.0%);第二个 300 人队列中占 42 例(14.0%)。真实人群里约 14%–17% 的人,其昼夜相位分布本身不是单峰;而睡眠时型在人群中又是近高斯(PMID 17936039,MCTQ 五万余人)——一个近乎正态的主群,加一个不规则的尾。

    把这个比例放进模型:主群正态 σ=1.30 h,掺入 15% 的等权双峰亚群(峰距 3.0 h),总 σ 只从 1.300 h 升到 1.332 h——升高 2.5%,任何以 σ 为输入的判据都感觉不到。但同一个混合分布的谱已经彻底变了:对正态拟合的最大残差达 11.1 个百分点,S₆ 从 12.4% 掉到 0 附近(20% 掺入时回到 4.2%,30% 时 12.5%——这个非单调本身就是形状效应在起作用)。

    这就是本文最要紧的一句话:离散度是近视的,形状才听得见。 一个亚群可以小到不改变你的 σ,却足以抹掉你第 6 阶的全部信号。

    证据与局限

    已验证(本文自算,可复现):
    – 在 σ 固定为 1.43 h 的条件下,六族 S_k 的跨度从 k=1 的 0.58 pp 单调升到 k=5 的 96.89 pp,「形状效应可忽略」不成立。
    – 二维方差分解给出交叉点在 k=3(族 43.4% / σ 38.2%),不是 k=4。
    – k=1 至 k=4 的族秩与超峰度秩完全一致;k=5 起首次偏离。
    – S₁ 反解 σ 的六族偏差在 −3.3% 至 +1.2% 之间。
    – σ 的效应方向在 k=3、k=4 时对 90-10 双峰族反转为正,k≥5 时对等权双峰族反转为正。

    部分验证:
    – 「真实人群中约 14%–17% 的相位分布非单峰」有直接实测支撑(PMID 42392022 的两个队列:17.0% 与 14.0%)。但该研究用的是 DLMO 分型(不规则 / 多峰),不是本文意义上的「等权双峰」,两者只是同族不同种;掺入比例与峰距是我设定的场景,不是实测参数。
    – 「人昼夜时型在人群中近高斯」(PMID 17936039,MCTQ 五万余人样本)支持主群用正态建模。注意这条说的是睡眠时型,不是本文的相位谐波合向量,不能直接等同。

    待验证:
    – 全部结论建立在 24 h 周期的 δ 峰族上;换成有宽度的真实核(峰宽 s₀ 与峰距 d 之比超过 0.7)时,双峰在 σ=1.43 h 下会熔成单峰,此时 S₄ 从 7.3% 升到 24.1%、S₆ 从 62.5% 降到 9.6%,可用信息大幅流失。真实核宽到底落在哪个区间,没有实测。
    – R_k 的抽样分布只用了 Rayleigh 估计的近似标准差,n 小时的系统性抬高在 n<60 区间未刻画;R₆>4 这条阈值需要独立数据校准。
    – 方差分解里存在极强的族×σ 交互,主效应百分比在此情形下不应单独解读(这一点我在文中用「跨度」作交叉印证,但完整的交互结构还需要一个明确的交互项模型)。

    相关阅读

    证据等级: C
    参考文献: PMID 42392022, PMID 17936039, PMID 31391040, PMID 32728310, PMID 27103930

    L1 网络状态:该主题尚未进入 L1 对应网络。CORRESPONDENCE_NETWORK.md 全表 grep「形状 / 离散 / 谐波 / 衰减 / 方差 / 分布」零命中,物理侧节点表仅有「相位 / 节律 | — | 周期性耦合的关键」一行(限定两振荡锁相,未及相位分布形态);EDGE_TABLE_science-yixue.md 中「谐波」仅为概念计数条目(147 篇 / A 级 20),无对应边。故本文封顶 C。

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


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

    Fixed dispersion, changing shape only is a contradictory experimental design

    Evidence level: C

    Dispersion is not another knob beside the shape; it is the integral of the shape. Therefore, “fixing SD while only changing shape” is not a control variable; it is comparing across different feasible domain boundaries—and these two things are mathematically not the same.

    Body

    The conclusion is: the proposition “shape affects higher-order harmonics” has never been cleanly tested, because past designs claiming to test it were invalid. But the proposition itself is not dead—it only requires changing the coordinate system.

    I. Why “Fixing σ” is Self-Contradictory

    The phase within a day falls at which moment is distributed, described in two common ways: dispersion σ (how scattered everyone wakes up) and shape (unimodal or bimodal). Intuitively these seem like two independent knobs: fix σ at 1.43 hours and try different shapes. The problem is that σ is a function of the shape, not a companion of the shape.

    Take the classic bimodal case. Two peaks placed at ±d with weights 0.5 each, then σ = d. To fix σ = 1.43 h, the peak distance is uniquely locked at 2.86 h; not a single number can move. To make the peak distance go to 3.5 h, σ must become 1.75 h. To keep σ while widening the peak distance, the only way is to change weights to asymmetric (e.g., 0.9 / 0.1), where σ = d/0.6, and the peak distance can go to 4.77 h—but skewness simultaneously jumps to +2.67, so it is no longer “bimodal” but “main peak plus a small outlier cluster.”

    Thus, within the slice of fixed σ, shape has only one degree of freedom, and that degree is constrained by σ. This is not a control variable; it is sliding along a curve that is rigidly constrained.

    II. True Calculation: How Large Is the Net Effect of Shape?

    I computed shared fractions S_k (phase vector length, ranging from 0 to 1) for six families (normal, Laplace, uniform, equal-weight bimodal, 90-10 bimodal, three-peaked 60-20-20) under the condition that σ equals 1.43 h for all:

    Family σ Excess Kurtosis S₁ S₂ S₃ S₄ S₅ S₆
    Normal 1.430 0.00 93.23 75.55 53.22 32.59 17.34 8.02
    Laplace 1.430 +2.96 93.45 78.11 61.32 47.14 36.34 28.39
    Uniform 1.430 −1.20 93.14 74.23 47.84 20.07 3.10 17.51
    Three-peaked 60-20-20 1.430 −0.50 93.19 75.09 51.86 31.39 20.66 23.32
    Equal-weight bimodal 1.430 −2.00 93.07 73.25 43.29 7.32 29.65 62.52
    90-10 bimodal 1.430 +5.11 93.65 82.23 81.95 93.25 99.99 94.05

    (Unit: %, σ unit: hours; phase cycles on 24 h, k = 1…6)

    Looking at the span: at k=1 the six families differ by 0.58 percentage points, at k=2 by 8.98, at k=3 by 38.67, at k=4 by 85.93, at k=5 by 96.89, and at k=6 by 86.02.

    This set of numbers completely overturns an old proposition: “S_k depends only on σ, and shape effects are negligible” is wrong, and at k≥4 it is terribly wrong—same 1.43-hour dispersion, S₄ can be 7.3% (equal-weight bimodal) or 93.3% (90-10 bimodal), differing by 12.7 times. Any rule “aligning by σ” would judge these two worlds as the same population.

    III. But Stratification Is Real, Only the Boundary Was Drawn Wrong

    Performing a two-dimensional variance decomposition of family and σ (6 families × 5 σ levels) yields:

    Order Family Main Effect σ Main Effect Total Span for That Order
    k=1 0.8% 98.6% 10.1 pp
    k=2 14.6% 76.0% 38.8 pp
    k=3 43.4% 38.2% 95.4 pp
    k=4 69.9% 13.7% 97.6 pp
    k=5 72.0% 2.6% 100.0 pp
    k=6 69.4% 2.2% 100.0 pp

    Stratification indeed exists—low orders reflect dispersion, high orders reflect shape. But the crossover point is at k=3, not k=4.

    Digging deeper: from k=1 to k=4, the ranking of the six families is exactly identical: “90-10 bimodal > Laplace > normal > three-peaked > uniform > equal-weight bimodal,” matching the excess kurtosis ranking completely; at k=5 it flips for the first time, and at k=6 it flips again.

    This indicates two things. First, for mid-low orders (k≤4) the shape effect is not a mess; it is a single number: kurtosis. High kurtosis (heavy tails, outliers) → higher S_k; low kurtosis (flat top, bimodal) → lower S_k. Thus “low orders reflect shape” can be precisely refined to “low orders reflect kurtosis.” Second, after k≥5 kurtosis is no longer sufficient—what matters then is structural information like “where exactly the tails lie,” which gets lost when compressed into a single number.

    IV. An Counterintuitive Finding: the Direction of σ’s Effect Is Not Constant

    The usual claim is “the larger the dispersion, the faster harmonic decay,” which holds only for the lowest two orders. For the six families as σ moves from 1.0 to 2.0, the change in S_k: at k=1 and k=2 all six families are negative (consistent direction); at k=3 and k=4 the 90-10 bimodal family reverses to positive (+13.9, +17.2 percentage points); at k=5 and k=6 the equal-weight bimodal family’s S₅ rises by 60.7 and S₆ rises by 100.0, while the three-peaked family’s S₆ rises by 41.7.

    The reason is geometric: as σ increases, the peak distance d of the bimodal family also increases; when 2d happens to align near an integer multiple of some order’s period, that order’s phase vector realigns. Whether this alignment occurs depends purely on the ratio of σ to the order, not on “dispersion” itself. Thus “the larger σ, the less shared” is only a local rule for unimodal families and valid only for k≤2, not a general law.

    V. Good News: σ Can Still Be Cleanly Recovered

    Since shape contaminates higher orders, can σ be estimated alone? Yes. From a single number—S₁—inverted via the normal distribution: σ_implied = √(−2 ln S₁) / ω, where ω = 2π/24. The deviations of the six families from the true σ=1.43 are: normal −0.0%, Laplace −1.7%, uniform +0.7%, three-peaked +0.3%, equal-weight bimodal +1.2%, 90-10 bimodal −3.3%.

    Maximum deviation is 3.3%. S₁ is a quite faithful estimator of σ, and shape’s contamination on it is small—because S₁ only looks at the first order, and shape details have not yet unfolded at this scale. This does not contradict Section II; it simply divides labor: σ is recovered from S₁, shape is recovered from higher orders.

    VI. The Correct Coordinate System: Shape Residual Spectrum

    Therefore, “fix σ and change shape” should be replaced by: first invert σ from S₁, then predict each order under that σ using the normal distribution, and finally divide the observed orders by the predicted values to obtain dimensionless residual spectra R_k:

    Family σ_implied R₂ R₃ R₄ R₅ R₆
    Normal 1.430 1.00 1.00 1.00 1.00 1.00
    Laplace 1.406 1.02 1.13 1.39 1.98 3.25
    Uniform 1.440 0.99 0.91 0.63 0.18 2.26
    Three-peaked 60-20-20 1.434 1.00 0.98 0.97 1.20 2.95
    Equal-weight bimodal 1.447 0.98 0.83 0.23 1.78 8.28
    90-10 bimodal 1.383 1.07 1.48 2.66 5.15 9.96

    The normal family is always 1 (the denominator is itself), so all deviations are attributed to shape. Discriminative power under Monte Carlo is quite strong: with σ fixed at 1.43 and n=200, using R₆ to distinguish normal from equal-weight bimodal, the normal family yields 1.15 ± 0.52 and the bimodal family yields 8.26 ± 0.11, with almost no overlap. But there is a trap that must be written down: when n is small, R_k is systematically elevated. At n=60 the median of the normal family’s R₆ is already 1.49 (interquartile range 0.97–2.20); at n=200 it falls back to 1.15, and at n=480 it is 1.10. Because the denominator (the normal family’s predicted sixth order, about 8%) is itself small, estimation error is amplified in the division. “R₆ > 1.5 indicates structure” would misclassify about two out of every ten normal populations at n=60; the safe threshold should be set at R₆ > 4 and also report n.

    VII. What This Means for Real People

    A 2026 clinical cohort study based on home saliva melatonin onset time (DLMO) classified 253 individuals, with “irregular/multimodal” accounting for 43 cases (17.0%); in a second cohort of 300 individuals it accounted for 42 cases (14.0%). In the real population, about 14%–17% of people have a circadian phase distribution that is itself not unimodal; meanwhile, sleep chronotypes in the population are nearly Gaussian (PMID 17936039, MCTQ over 50,000 people)—a nearly normal main group plus an irregular tail.

    Plugging this proportion into the model: the main group is normal with σ=1.30 h, mixed with a 15% equal-weight bimodal subgroup (peak distance 3.0 h), the total σ rises only from 1.300 h to 1.332 h—an increase of 2.5%, so any criterion based on σ would not feel it. But the spectrum of this mixed distribution has completely changed: the maximum residual from fitting normal reaches 11.1 percentage points, S₆ drops from 12.4% to near zero (at 20% mixing it returns to 4.2%, at 30% it is 12.5%—this non-monotonicity itself reflects the shape effect).

    This is the most important sentence of this paper: dispersion is nearsighted; only shape can be heard. A subgroup can be so small that it does not change your σ, yet it is enough to erase all signal at your sixth order.

    Evidence and Limitations

    Verified (self-calculated in this paper, reproducible):
    – Under the condition that σ is fixed at 1.43 h, the span of S_k across the six families monotonically rises from 0.58 pp at k=1 to 96.89 pp at k=5, so “shape effects are negligible” does not hold.
    – Two-dimensional variance decomposition shows the crossover point is at k=3 (family 43.4% / σ 38.2%), not at k=4.
    – From k=1 to k=4, the family rank and excess kurtosis rank are completely identical; at k=5 they first diverge.
    – The six-family deviations of inverting σ from S₁ lie between −3.3% and +1.2%.
    – The direction of σ’s effect reverses to positive for the 90-10 bimodal family at k=3 and k=4, and for the equal-weight bimodal family at k≥5.

    Partially Verified:
    – “In the real population, about 14%–17% of phase distributions are not unimodal” has direct empirical support (PMID 42392022’s two cohorts: 17.0% and 14.0%). However, that study used DLMO classification (irregular/multimodal), not the “equal-weight bimodal” sense of this paper; they are the same family but different species. The mixing proportion and peak distance are scenario settings, not measured parameters.
    – “Human circadian chronotypes in the population are nearly Gaussian” (PMID 17936039, MCTQ over 50,000 subjects) supports using the normal model for the main group. Note this refers to sleep chronotypes, not the phase harmonic vectors of this paper, so they cannot be directly equated.

    To Be Verified:
    – All conclusions are built on δ-peaked families with a 24 h cycle; when replacing them with real kernels of width (peak width s₀ and peak distance d ratio exceeding 0.7), bimodal distributions at σ=1.43 h would melt into unimodal, causing S₄ to rise from 7.3% to 24.1% and S₆ to fall from 62.5% to 9.6%, losing much usable information. The exact interval of real kernel width has no measurement yet.
    – The sampling distribution of R_k used only the Rayleigh-estimated approximate standard deviation; the systematic elevation for n<60 is not characterized; the threshold R₆>4 needs independent data calibration.
    – The variance decomposition contains strong family×σ interaction; the main effect percentages should not be interpreted alone in this situation (this point is cross-validated by “span” in the text, but the full interaction structure requires an explicit interaction term model).

    Further Reading

    Evidence Level: C
    References: PMID 42392022, PMID 17936039, PMID 31391040, PMID 32728310, PMID 27103930

    L1 Network Status: This topic has not yet entered the corresponding L1 network. CORRESPONDENCE_NETWORK.md grep for “shape / dispersion / harmonic / decay / variance / distribution” yields zero hits; the physical-side node table contains only “phase / rhythm | — | key for periodic coupling” (limited to two-oscillator phase locking, not reaching phase distribution shape); EDGE_TABLE_science-yixue.md lists “harmonic” only as a concept counting entry (147 papers / A-level 20), with no corresponding edge. Hence this paper is capped at C.

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


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

  • K1–K4 判档体系的分辨率下限:当档位比尺子还细

    Evidence level: B

    一句话结论

    池④ 用 K1/K2/K3/K4 四档来决定”这个山头值不值得花钱去测”。本文逐格复算后给出的结论是:这套判档体系的分辨率(±3 dB 带宽)比它能拿到的参数精度(σ 只有一个数量级)细了一个量级,因此三档的实际有效域是空集,K4 不是”特殊情况兜底”,而是唯一诚实的默认档。

    这个结论不来自任何新数据,只来自对文件自身三个数字的对照。三个数字都在库里,只是此前没有任何一轮把它们并排放过。


    一、判档体系是什么,为什么它重要

    high-sigma-lithology-elf-shielding-path.md §6 与配套实验卡 P4-EXP-004 §4 定义了一套先验判档:

    档 原文条件 原文含义
    K1 可测 σ ≥ 3×10⁻² 且 连续厚度 ≥ 1.2 km 且 频段 ≥ 14 Hz 预期 A ≥ 10 dB,超出本底 → 值得投入实测
    K2 边缘 落在 10 dB 线 ±3 dB 内 预期效应与未建模项同量级 → 强制双站差分
    K3 透明 σ ≤ 10⁻³,任意 1–3 km 厚度,任意 1–30 Hz A < 2 dB → 不得作屏蔽论据
    K4 不可判 仅有区域地质图 σ 估计(σ 不确定度 ≥ 1 个数量级) 三元表任何一格都跨 ±10 dB → 只能写”不可判定”

    这套体系在库里承担的实际职能很重:它是 #14(靠山屏蔽系数实测)、#22(山体样本优先级)、#24/#25(Schumann 站选址)三个课题的共同前置闸门。判档判错,预算就投错样本。上一轮(第 22 轮)在池③ 立的「测量链分辨率下限」讲的正是这件事的另一面——先测尺子,再量东西。本文把同一把尺子转向池④ 自己的判档表。

    复算所用公式就是文件自己写的那个:

    A(dB) = 8.686 · d / δ,    δ = √(2/(ω μ₀ μᵣ σ))
    

    代入后可以合并成一条更好用的形式(d 以米计):

    A(dB) = 0.0172584 · d · √(μᵣ · σ · f)
    

    下面全部数字都由这一条式子和文件自述的 σ / μᵣ 算出,不引入任何外部参数。工具 skin-depth-recompute.py judge 已独立复现。


    二、第一处:K1 一档三定义

    同一档 K1,在同一个文件家族里有三个互不等价的定义:

    出处 K1 的条件 用 A(dB) 表达
    文章 §6 表 σ ≥ 3×10⁻² 且 t ≥ 1.2 km 且 f ≥ 14 Hz 一个三维 AND 盒(不是 dB 门槛)
    文章 §6 同表「含义」列 —— 预期 A ≥ 10 dB
    工具 judge_tier() —— A ≥ 13 dB(= 10 dB 线 + 3 dB 带)

    三者可以同时”正确”,因为它们在回答不同问题;但它们被并列写进同一张表,读者会当成一件事。差异有多大?把 AND 盒的角点代进去:

    σ=3×10⁻², t=1.2 km, f=14 Hz
    A = 0.0172584 × 1200 × √(3×10⁻² × 14) = 13.42 dB
    

    角点值是 13.42 dB,不是 10 dB。「预期 A ≥ 10 dB」这句话在角点上偏低了 3.4 dB;若要让角点正好落在 10 dB,σ 得是 1.665×10⁻² 而不是 3×10⁻²。

    方向更值得注意:AND 盒与 dB 门槛不是包含关系,而是交叉关系。 用工具逐个跑:

    σ (S/m) t f (Hz) A (dB) 在 AND 盒内? 工具判档
    1×10⁻¹ 1.0 km 7.83 15.27 ❌ 否(t<1.2km 且 f<14) K1
    1×10⁻¹ 1.2 km 7.83 18.33 ❌ 否(f<14) K1
    1×10⁻¹ 0.8 km 14 16.34 ❌ 否(t<1.2km) K1
    3×10⁻² 1.2 km 7.83 10.04 ❌ 否(f<14) K2

    第一行是最扎眼的:σ=10⁻¹、1 km、7.83 Hz 给出 15.27 dB——比 K1 门槛 13 dB 高出 2.3 dB,却因为”厚度差 200 m、频率差 6 Hz”被 AND 盒判成不合格。

    物理上说不通:A 是 σ、t、f 的连续单调函数,任何”必须同时满足三个各自门槛”的盒子,都会在盒面上造出不连续跳变。一个场地的 t 从 1.19 km 改成 1.21 km,判档从”不值得测”跳到”值得测”,而 A 只变了 0.2 dB。

    这不是判据严不严的问题,是判据的形状不对。 该判的是 A,不该判它的三个自变量。AND 盒唯一合理的用途是”预筛候选清单”(快速剔除明显不可能的组合),不能当成”值得测”的充分条件。


    三、第二处:K3「任意厚度」是假的

    K3 的原文写「σ ≤ 10⁻³,任意 1–3 km 厚度,任意 1–30 Hz → A < 2 dB」。

    固定 σ=10⁻³,逐格算:

    t f=7.83 Hz f=14 Hz f=30 Hz
    1 km 1.53 dB 2.04 dB 2.99 dB
    2 km 3.05 dB 4.08 dB 5.98 dB
    3 km 4.58 dB 6.13 dB 8.97 dB

    九格里只有一格满足”A < 2 dB”。最不利的一格 8.97 dB 是 2 dB 的 4.5 倍,并且已经越过 K2 下界(7 dB)——按工具自己的判档,σ=10⁻³, t=3 km, f=30 Hz 得到 K2-边缘,会触发”强制双站差分”规则。也就是说:K3 说”终止,不作屏蔽论据”,K2 说”进 Step 3,必须双站差分”,同一组参数被两条规则同时判给两个不同的处置。

    反解真实边界:A<2 dB 允许的最大厚度是

    t_max @30 Hz  = 669 m
    t_max @7.83 Hz = 1 310 m
    

    所以 K3 的正确表述不是”任意厚度”,而是 “σ ≤ 10⁻³ 且 t ≤ 0.67 km(按 30 Hz 最不利端)”。原文把 t_max 写成了”任意 1–3 km”,在 1–3 km 段高估了 1.5–4.5 倍。

    这条错误的性质与第 20/21 轮修的那批不同:那批是数值错(δ 格算错倍率),这条是量词错——用”任意”把一个有限区间说成了全域。量词错的危害在于它不可通过复算单格发现:你随便抽一格可能都对,只有把边界推到端点才看得见。


    四、第三处:K4 自己否证了 K1/K2/K3

    这是本文最重要的一处,也是唯一不需要任何新计算的一处。

    原文给 K4 的定义里写了这么一句:

    仅有区域地质图估计(σ 不确定度 ≥ 1 个数量级)→ 三元表任何一格都跨 ±10 dB → 结论只能写「不可判定」。

    这句话是对的。σ 差一个数量级(10×),而 A ∝ √σ,所以 A 摆动 √10 = 3.16 倍。在 A=10 dB 处,3.16 倍就是 10 dB 的摆动。原文的”±10 dB”准确。

    问题在于:K2 的带宽是 ±3 dB。

    一个 ±10 dB 的不确定度,套不进一个 ±3 dB 的判据带。这不是”精度差一点”,是参数误差比判据带宽大 3.3 倍。用一句更直白的话:你手上那把尺子的刻度抖动,比你要分辨的两个刻度之间的距离还宽。

    把这个要求反解成对 σ 精度的要求(求 A 判到 ±3 dB 时,σ 允许的区间跨度):

    目标 A 允许的 σ 区间 σ 需准到
    5 dB ×0.16 – ×2.56 ±60%
    7 dB(K3/K2 界) ×0.33 – ×2.04 ±41%
    10 dB(K2 中心) ×0.49 – ×1.69 ±30%
    13 dB(K2/K1 界) ×0.59 – ×1.51 ±23%
    20 dB ×0.72 – ×1.32 ±15%

    结论:要判 K1/K2/K3 任何一个边界,σ 必须准到 ±23%–±41%。

    而现实中能拿到什么精度?

    σ 精度 A 能可靠判到的上限
    ±2 倍(区间 4×) 3.00 dB
    ±3.16 倍(区间 10×,即 K4 定义) 1.39 dB
    ±10 倍(区间 100×,多数文献区间) 0.33 dB

    K4 定义里那个”1 个数量级”,对应 A 只能判到 1.39 dB。而 K3 的上界是 7 dB。1.39 << 7,差 5 倍。

    于是整件事就清楚了:

    K4 的触发条件(σ 不确定度 ≥ 1 数量级)
      ⇒ A 只能判到 1.39 dB
      ⇒ 远小于 K3 的 7 dB 上界
      ⇒ K3/K2/K1 三档全部不可判
      ⇒ 任何满足 K4 触发条件的场地,必然全部落进 K4
    

    K4 不是例外档,是默认档。它在表里被摆在第四行、写成”特殊情况”,实际覆盖面是判档体系的全部有效域——只要场地没有 ±40% 以内的实测 σ。

    反过来说,K1/K2/K3 三档要成立,需要一个隐含前提:场地已有一份精度优于 ±40% 的实测 σ。这个前提在文件里从没写出来,也从没被检查过。而 #26(Schumann 站 σ 数据库)第 31 轮的自评是”6 站,a 级 0 站“——按该库自己的分级,一个 a 级都没有。


    五、三处错误其实是同一个错误

    把三处并排放:

    # 表面现象 共同结构
    1 K1 用 AND 盒判,不判 A 判据的维度选错——判了自变量,没判结果量
    2 K3 用”任意”覆盖有限区间 判据的量词过宽——把 t_max 说成 ∞
    3 ±3 dB 带宽 配 ±10 dB 输入精度 判据的分辨率高于输入分辨率

    三者都指向同一件事:这套判档表是在”参数已知”的假设下设计的,而它的适用场景恰恰是”参数未知”。

    再往下一层:为什么这个错误能存活这么多轮?因为它不在任何单格里。第 20 轮修的 6 行 δ 表是逐格复算能抓的;第 22 轮修的 SHAP φ₀ 是加性分解能抓的。本文这三处,逐格抽检任意一格都可能”对”,只有把判据的边界和参数的边界并排放才看得见。

    这给出一条可迁移的方法学条目:

    R26「判据分辨率必须与被判量的输入分辨率匹配」:任何分档/分级/阈值体系,除了检查”每一档的数值算对没有”,还必须检查”判据带宽 ÷ 输入不确定度”。比值 < 1 时,档位越多越有害——它给出的是伪精度。

    R26 与已有几条的关系:

    • R13(区间不可单点判):R13 说”区间不能当单点”。R26 更进一步:即使你拿到了单点,也要问这个单点准到多少,判据带够不够宽。
    • R19(区间不可判 ≠ 区间没问题):R19 管”数值是否自洽”。R26 管”自洽的数值够不够用来分档”。两件事。
    • R22(canonical 表优先复算):R22 管”哪些表要复算”。R26 管”复算之后这张表能不能用”。
    • 池③ 第 22 轮「测量链分辨率下限」:那是一台设备换另一台设备引起的摆动。R26 是参数本身的不确定度。同族,但测量链分辨率下限能通过重复测量降低,σ 的不确定度在没做 MT 实测前降不下来——所以 R26 更硬:它不是”先测一下尺子”,是”这把尺子的刻度根本不够画这么多格”。

    六、该怎么改(建议,须 JOE 批)

    本文不修改任何已发布正文(见 §8 边界)。以下四条是给 D 清单的新增建议,不是已执行动作。

    建议 1(P1,改动最小):K1 的 AND 盒降级为”候选预筛”,不再是”值得测”的判据;判档一律以 A(dB) 为准。同时把「预期 A ≥ 10 dB」与工具的「≥13 dB」统一成一个数——推荐直接用两条 dB 边界:A < 7 dB → K3、7 ≤ A < 13 → K2、A ≥ 13 → K1,删掉三个自变量的门槛句。

    建议 2(P1,纯数值):K3 的量词从”任意 1–3 km 厚度”改为”t ≤ 0.67 km(30 Hz 最不利端)”,并在表下加一行 t_max 速查(30 Hz: 669 m / 7.83 Hz: 1 310 m)。

    建议 3(P1,最重要):K4 从”第四档”提升为前置门禁——判 K1/K2/K3 之前先判 σ 精度,σ 不确定度 > ±41% 直接停,不进三档。这等价于承认”没有实测 σ 就没有档位”,与 #46a(补 f_meas 列)、#52(f_meas 原书溯源)是同一条因果链的下游。

    建议 4(P2,工具):skin-depth-recompute.py judge 增加 --sigma-uncertainty <factor> 参数,输出”该 σ 精度下本判档是否可信”一行。这比再加判档功能有用——现有工具报 A 报得很准,但不报”这个 A 配不配分档”。


    七、证据与局限

    已有证据(可逐条复现):

    • 三处不一致全部由文件自述的公式与参数算出,skin-depth-recompute.py judge 独立复现了 §2 的四个反例与 §3 的 K2 误判(σ=10⁻³, t=3 km, f=30 Hz → A=8.97 dB → K2-边缘)。
    • 电磁屏蔽的工程背景有文献支撑(建筑地磁屏蔽实测:PMID 20000094;ELF 暴露测量:PMID 24131394;低磁环境生物效应综述:PMID 28654641)。
    • 岩石电导率的频率依赖(本判档”σ 当常数”这一近似的来源)有明确文献:Jonscher 通用介电响应(DOI 10.1038/267673a0)、Revil 非饱和多孔介质频段电导/介电模型(DOI 10.1029/2012WR012700)、Beskardes 低频电导率幂律分布(DOI 10.1093/gji/ggw375)。
    • 环境 ELF 与昼夜节律的关联有综述(PMID 36681118),仅作背景,不支撑任何屏蔽结论。

    局限(必须与结论同读):

    1. 判”错”仅指”与文件自身声明的公式 / 参数区间不自洽”,不指地下真实 δ 或 A 为另一值。 若某档本意是某个未写出的等效 σ,问题从”算错”变成”没写清 σ”——但无论哪种,档位都不能按现在这样用。
    2. 全部为解析复算 + 文本解析,无任何实地测量。δ 公式本身假设均匀半空间、无位移电流、μᵣ 常数;真实山体分层、各向异性,且 σ 随频率弥散。
    3. 本文没有回答”靠山能不能屏蔽 ELF”这个物理问题,只回答”这套判档表能不能用来回答它”。目前的答案是:不能,除非先有 ±40% 以内的 σ。
    4. §4 的”K4 覆盖全部有效域”是一个逻辑推论,前提是接受 K4 自己的”≥1 数量级”定义。若将来 σ 实测精度提高到 ±40%,该推论失效——这正是希望它失效的方向。
    5. 本文未审计 skin-depth-reference-table.md §三 的 9 行(第 20 轮已单独审计过,6 行不自洽),也未重开 D18/D21/D22 的 δ 数值错配——那批是另一类错误(数值错),本文是判据结构错,两者不应合并处理。
    6. §5 的 R26 是从本例抽象出的方法学条目,只经本库这一处实例检验,尚未在池①②③⑤交叉验证。按 R11「零参数判据只可证伪不可验证」,R26 目前只承担”发现伪精度”的职能,不能反过来当”判据合格”的证明。

    八、本轮未做(守边界)

    • 未修改任何已发布正文:high-sigma-lithology-elf-shielding-path.md、skin-depth-reference-table.md、skin-depth-basement-building-mu-metal-table.md、pool4-high-sigma-lithology-elf-shielding-path.md(P4-EXP-004)全部 mtime 未动。
    • 未修改 skin-depth-recompute.py(建议 4 留待 JOE 批后单独做,避免”改了哪一处导致判档变化”无法归因)。
    • 未做 σ 实测;未新增”靠山→健康”通路(净增通路 0);未标 accepted;未直写 779/782;未新增或修改任何 OC cron 与心跳配置。
    • §6 四条建议全部停留在建议层,需 JOE 批后才可执行正文更正。

    九、相关阅读

    • 《高电导岩性 ELF 屏蔽路径》(#42)——K1–K4 判档表的原始出处
    • 《趋肤深度对照表》(#16)——δ 数值层的对照表,第 20 轮已审出 6 行不自洽
    • 《靠山地形 ELF 屏蔽系数实测》(#14)——判档表的直接消费者
    • 《山体样本 σ 预筛 SOP》(#22)——被本判档体系修正过的上一版门槛
    • 《区间断言逐条判定》(第 52 轮)——区间层的处置;本文处理的是判据层

    本文为 draft。结论为”文件内部判据结构不自洽”的方法学审计,不构成任何关于山体电磁屏蔽、建筑屏蔽或人体健康的实测声明与健康建议。
    进入 779 终版前须走 intake;进入 782 前须去 AI 化。
    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。文中涉及的衰减量、电导率区间与判档门槛均为工程公式在文献参数下的推算,非实测值;真实地下介质存在分层、各向异性与频率弥散,实际结果可能显著偏离。任何”居住环境电磁条件与健康”的推论均停留在假说层,须经独立实验与同行评议后方可对外表述;如有身体不适或健康疑虑,请就医并由专业医师判断。

    Resolution floor of the K1–K4 grading system: when the grade scale is finer than the ruler

    Evidence level: B

    One-Sentence Conclusion

    Chi④ uses the four tiers K1/K2/K3/K4 to decide “whether this hilltop is worth spending money on to measure.” After recalculating each tier individually, the conclusion is: the resolution of this judgment system (±3 dB bandwidth) is one order of magnitude finer than the parameter precision it can obtain (σ spans only one order of magnitude); therefore, the actual effective domain of the three tiers is empty, and K4 is not a “special-case fallback” but the only honest default tier.

    This conclusion does not come from any new data, but from cross-referencing three numbers already present in the file. All three numbers exist in the repository; they were simply never placed side-by-side in any prior round of analysis.

    I. What is the Tiering System and Why Is It Important

    high-sigma-lithology-elf-shielding-path.md §6 and the accompanying experimental card P4-EXP-004 §4 define a priori tiering:

    Tier Original Condition Original Meaning
    K1 Measurable σ ≥ 3×10⁻² and continuous thickness ≥ 1.2 km and frequency band ≥ 14 Hz Expected attenuation A ≥ 10 dB, exceeding background → warrants field measurement
    K2 Edge Falls within 10 dB line ±3 dB Expected effect and unmodeled terms are of the same order → forces dual-site differencing
    K3 Transparent σ ≤ 10⁻³, any 1–3 km thickness, any 1–30 Hz A < 2 dB → cannot serve as shielding evidence
    K4 Untierable Only regional geological map σ estimate (σ uncertainty ≥ 1 order of magnitude) Ternary table any cell spans ±10 dB → can only write “undetermined”

    This system bears heavy practical responsibility within the repository: it serves as the common upstream gate for projects #14 (back-shield coefficient measurement), #22 (mountain sample priority), and #24/#25 (Schumann station siting). Mis-tiering results in misallocated budgets. The previous round (Round 22) established the “measurement chain resolution floor” in pool ③, which addresses another facet of this issue—measure the ruler first, then measure things. This paper turns the same ruler toward pool ④’s own tiering table.

    The formulas used for recalculation are exactly those written in the file itself:

    A(dB) = 8.686 · d / δ,    δ = √(2/(ω μ₀ μᵣ σ))
    

    Substituting yields a more convenient combined form (with d in meters):

    A(dB) = 0.0172584 · d · √(μᵣ · σ · f)
    

    All subsequent numbers are derived solely from this equation and the file’s stated σ / μᵣ values, introducing no external parameters. The tool skin-depth-recompute.py judge independently reproduced these results.

    II. First Instance: Three Non-Equivalent Definitions for Tier K1

    The same tier, K1, possesses three mutually non-equivalent definitions within the same document family:

    Source Condition for K1 Expressed via A(dB)
    Article §6 Table σ ≥ 3×10⁻² and t ≥ 1.2 km and f ≥ 14 Hz A three-dimensional AND box (not an A threshold)
    Article §6 same table “Meaning” column — Expected A ≥ 10 dB
    Tool judge_tier() — A ≥ 13 dB (= 10 dB line + 3 dB band)

    All three can simultaneously be “correct” because they answer different questions; yet they are listed side-by-side in the same table, leading readers to treat them as a single criterion. How significant is the discrepancy? Substituting the corner points of the AND box:

    σ=3×10⁻², t=1.2 km, f=14 Hz
    A = 0.0172584 × 1200 × √(3×10⁻² × 14) = 13.42 dB
    

    The corner value is 13.42 dB, not 10 dB. The statement “Expected A ≥ 10 dB” underestimates the corner by 3.4 dB; for the corner to fall exactly at 10 dB, σ would need to be 1.665×10⁻² rather than 3×10⁻².

    The direction of the discrepancy is more noteworthy: the AND box and the A threshold are not nested but intersecting. Running the tool iteratively:

    σ (S/m) t f (Hz) A (dB) Inside AND box? Tool tier
    1×10⁻¹ 1.0 km 7.83 15.27 ❌ No (t<1.2km and f<14) K1
    1×10⁻¹ 1.2 km 7.83 18.33 ❌ No (f<14) K1
    1×10⁻¹ 0.8 km 14 16.34 ❌ No (t<1.2km) K1
    3×10⁻² 1.2 km 7.83 10.04 ❌ No (f<14) K2

    The first row is the most striking: σ=10⁻¹, 1 km, 7.83 Hz yields 15.27 dB—2.3 dB above the K1 threshold of 13 dB—yet the AND box deems it ineligible due to “thickness difference of 200 m and frequency difference of 6 Hz.”

    This is physically inconsistent: A is a continuous monotonic function of σ, t, and f; any “box requiring simultaneous satisfaction of three individual thresholds” creates discontinuous jumps on the box faces. Changing a site’s t from 1.19 km to 1.21 km shifts the tier judgment from “not worth measuring” to “worth measuring,” while A changes only by 0.2 dB.

    This is not a matter of strictness but of shape. The criterion should judge A, not its three independent variables. The AND box’s only rational use is as a “pre-screening candidate list” (rapidly eliminating obviously impossible combinations), not as a sufficient condition for “worth measuring.”

    III. Second Point: K3’s “Arbitrary Thickness” is False

    The original text of K3 states “σ ≤ 10⁻³, arbitrary thickness 1–3 km, arbitrary frequency 1–30 Hz → A < 2 dB”.

    Fixing σ = 10⁻³ and calculating cell-by-cell:

    t f = 7.83 Hz f = 14 Hz f = 30 Hz
    1 km 1.53 dB 2.04 dB 2.99 dB
    2 km 3.05 dB 4.08 dB 5.98 dB
    3 km 4.58 dB 6.13 dB 8.97 dB

    Only one of the nine cells satisfies “A < 2 dB”. The most unfavorable cell, 8.97 dB, is 4.5 times the 2 dB threshold and already exceeds the K2 lower bound (7 dB)—according to the tool’s own classification, σ = 10⁻³, t = 3 km, f = 30 Hz yields K2-edge, triggering the “mandatory dual-station differencing” rule. In other words: K3 says “terminate, do not use as shielding evidence,” while K2 says “proceed to Step 3, must apply dual-station differencing.” The same set of parameters is assigned two different dispositions by two conflicting rules.

    Reversing to find the true boundary: the maximum thickness allowed under A < 2 dB is

    t_max @30 Hz  = 669 m
    t_max @7.83 Hz = 1 310 m
    

    Thus the correct statement of K3 is not “arbitrary thickness,” but “σ ≤ 10⁻³ and t ≤ 0.67 km (at the most unfavorable end, 30 Hz).” The original text erroneously writes t_max as “arbitrary 1–3 km,” overestimating the range by 1.5–4.5 times within the 1–3 km segment.

    The nature of this error differs from those corrected in rounds 20/21: those were numerical errors (incorrect multiplication factors in the δ-grid calculation), whereas this is a quantifier error—using “arbitrary” to describe a finite interval as if it were universal. The danger of such quantifier errors lies in their undetectability via single-cell recalculation: randomly selecting one cell may yield a correct result; only pushing the boundary to its extreme reveals the discrepancy.

    IV. Third Instance: K4 Self-Nullifies K1/K2/K3

    This is the most critical section of this paper, and the only one requiring no new calculations.

    The definition provided for K4 states:

    Regional geological map estimates alone (σ uncertainty ≥ 1 order of magnitude) → any cell in the ternary table spans ±10 dB → conclusion must be “undecidable.”

    This statement is correct. A σ difference of one order of magnitude (10×), with A ∝ √σ, results in an A fluctuation of √10 = 3.16 times. At A = 10 dB, this 3.16-fold variation equals a ±10 dB swing. The original “±10 dB” is accurate.

    The problem lies in: K2’s bandwidth is ±3 dB.

    A ±10 dB uncertainty cannot fit within a ±3 dB decision band. This is not “slightly imprecise”; it is a parameter error 3.3 times larger than the decision bandwidth. In plainer terms: the jitter of the scale on your ruler is wider than the distance between the two marks you are trying to distinguish.

    Reversing this requirement to determine the necessary σ precision (finding the allowable σ interval span when A is judged to ±3 dB):

    Target A Allowed σ Interval Required σ Accuracy
    5 dB ×0.16 – ×2.56 ±60%
    7 dB (K3/K2 boundary) ×0.33 – ×2.04 ±41%
    10 dB (K2 center) ×0.49 – ×1.69 ±30%
    13 dB (K2/K1 boundary) ×0.59 – ×1.51 ±23%
    20 dB ×0.72 – ×1.32 ±15%

    Conclusion: To judge any boundary of K1/K2/K3, σ must be accurate to ±23%–±41%.

    What precision is realistically achievable?

    σ Accuracy Maximum Reliable A Judgment
    ±2× (interval 4×) 3.00 dB
    ±3.16× (interval 10×, i.e., K4 definition) 1.39 dB
    ±10× (interval 100×, common in literature) 0.33 dB

    The “1 order of magnitude” in the K4 definition corresponds to an A judgment limit of only 1.39 dB. Since the upper bound for K3 is 7 dB, and 1.39 << 7 (a factor of 5 difference), the entire scenario becomes clear:

    K4 Trigger Condition (σ uncertainty ≥ 1 order of magnitude)
      ⇒ A can only be judged up to 1.39 dB
      ⇒ Far below K3's 7 dB upper bound
      ⇒ All three tiers (K3/K2/K1) become undecidable
      ⇒ Any site meeting K4 trigger conditions inevitably falls entirely into K4
    

    K4 is not an exception tier; it is the default tier. Although placed in the fourth row of the table and labeled “special case,” its actual coverage encompasses the entire valid domain of the decision tier system—provided the site lacks measured σ within ±40%.

    Conversely, for the K1/K2/K3 tiers to hold, an implicit prerequisite exists: the site must possess measured σ with accuracy better than ±40%. This prerequisite is never stated in the document, nor ever checked. As noted in #26 (Schumann station σ database), the self-assessment of round 31 states “6 stations, zero a-grade stations“—according to its own classification, not a single a-grade station exists.

    V. Three Errors Are Actually One Error

    Placing the three errors side-by-side:

    # Surface Phenomenon Shared Structure
    1 K1 uses an AND gate for judgment but does not evaluate A Dimensional mismatch of the criterion—the criterion evaluates the independent variable, not the outcome quantity
    2 K3 uses “arbitrary” to cover a finite interval Overly broad quantifier in the criterion—stating t_max as ∞
    3 ±3 dB bandwidth paired with ±10 dB input precision Criterion resolution exceeds input resolution

    All three point to the same issue: this grading table was designed under the assumption of “known parameters,” yet its intended application scenario is precisely “unknown parameters.”

    Going one layer deeper: why has this error persisted through so many rounds? Because it does not reside in any single cell. The 6-row δ table revised in round 20 could be caught by recalculating cell-by-cell; the SHAP φ₀ revised in round 22 could be caught by additive decomposition. These three errors in this paper, however, can only be revealed by placing the boundaries of the criterion and the boundaries of the parameters side-by-side.

    This yields a transferable methodological entry:

    R26 “Criterion resolution must match the input resolution of the quantity being judged”: Any tiering/grading/threshold system, besides checking whether “the numerical value of each tier is calculated correctly,” must also check whether “criterion bandwidth ÷ input uncertainty.” When the ratio is < 1, more tiers are more harmful—they provide pseudo-precision.

    Relationships of R26 with existing entries:

    • R13 (interval cannot be judged as a single point): R13 states “an interval cannot be treated as a single point.” R26 goes further: even if you obtain a single point, you must ask how precise this point is and whether the criterion bandwidth is wide enough.
    • R19 (interval cannot be judged ≠ interval is fine): R19 manages “whether values are internally consistent.” R26 manages “whether internally consistent values are sufficient for tiering.” These are two distinct matters.
    • R22 (canonical tables take priority for recalculation): R22 manages “which tables require recalculation.” R26 manages “whether this table can be used after recalculation.”
    • Pool③ Round 22 “measurement chain resolution lower limit”: That refers to fluctuation caused by swapping one device for another. R26 refers to uncertainty inherent in the parameters themselves. They belong to the same family, but the measurement chain resolution lower limit can be reduced through repeated measurements, whereas σ uncertainty cannot be reduced until MT actual measurement is performed—thus R26 is more stringent: it is not “measure the ruler first,” but “the scale of this ruler is fundamentally insufficient to draw so many tiers.”

    VI. How to modify (Suggestions, subject to JOE’s review)

    This paper does not alter any published text (see §8 Boundary). The following four items are new additions to the D list, not actions already executed.

    Suggestion 1 (P1, minimal change): Downgrade the AND box of K1 to “candidate pre-screening,” no longer serving as the criterion for “worth testing.” All categorization shall be based solely on A(dB). Simultaneously, unify the “expected A ≥ 10 dB” and the tool’s “≥13 dB” into a single numerical scale. It is recommended to directly adopt two dB boundaries: A < 7 dB → K3, 7 ≤ A < 13 → K2, A ≥ 13 → K1, thereby removing the three threshold sentences involving independent variables.

    Suggestion 2 (P1, purely numerical): Change the quantifier for K3 from “any 1–3 km thickness” to “t ≤ 0.67 km (worst-case at 30 Hz),” and add a row for t_max quick reference beneath the table (30 Hz: 669 m / 7.83 Hz: 1,310 m).

    Suggestion 3 (P1, most important): Upgrade K4 from “fourth tier” to preliminary gate—assess σ accuracy before judging K1/K2/K3. If the uncertainty of σ exceeds ±41%, stop immediately and do not enter any tier. This is equivalent to acknowledging “no measured σ means no tiers,” which lies downstream of the same causal chain as #46a (adding f_meas column) and #52 (original source tracing for f_meas).

    Suggestion 4 (P2, tool): The skin-depth-recompute.py judge script should add the --sigma-uncertainty <factor> parameter, outputting a line stating “whether this σ accuracy makes the current categorization credible.” This is more useful than adding another categorization function—existing tools report A very accurately but do not indicate “whether this A matches the required tier.”

    VII. Evidence and Limitations

    Existing evidence (reproducible item by item):

    • All three inconsistencies were derived from the formulas and parameters explicitly stated in the document, independently reproduced by skin-depth-recompute.py judge. This re-computation confirmed the four counterexamples from §2 and the K2 misjudgment from §3 (with σ=10⁻³, t=3 km, f=30 Hz → A=8.97 dB → K2-edge).
    • The engineering background for electromagnetic shielding is supported by literature (measured geomagnetic shielding in buildings: PMID 20000094; ELF exposure measurement: PMID 24131394; review of biological effects in low-magnetic environments: PMID 28654641).
    • The frequency dependence of rock conductivity (the source of the approximation “σ treated as constant” in this grading scheme) is explicitly documented: Jonscher universal dielectric response (DOI 10.1038/267673a0), Revil non-saturated porous medium conductivity/dielectric models (DOI 10.1029/2012WR012700), Beskardes power-law distribution of low-frequency conductivity (DOI 10.1093/gji/ggw375).
    • The association between environmental ELF and circadian rhythms is reviewed (PMID 36681118), serving only as background and not supporting any shielding conclusion.

    Limitations (must be read alongside the conclusions):

    1. Judging “wrong” refers only to “inconsistency with the formulas and parameter ranges explicitly stated in the document itself,” not to the true subsurface δ or A having different values. If a grading scheme was intended for an unstated equivalent σ, the problem shifts from “calculated incorrectly” to “σ not clearly defined”—but either way, the grading scheme cannot be used as currently formulated.
    2. All calculations are analytical re-computations plus text parsing; no field measurements were conducted. The δ formula itself assumes a uniform half-space, no displacement current, and constant μᵣ; real mountain strata are layered and anisotropic, and σ disperses with frequency.
    3. This paper does not answer the physical question “Can a mountain backshield ELF?” It only answers “Can this grading scheme be used to answer that question.” The current answer is: no, unless σ is known within ±40%.
    4. The claim in §4 that “K4 covers the entire valid domain” is a logical inference, contingent on accepting K4’s own definition of “≥1 order of magnitude.” If future σ measurement precision improves to ±40%, this inference will fail—precisely the direction we hope it fails.
    5. This paper did not audit the 9 rows in §3 of skin-depth-reference-table.md (the 20th iteration was already audited separately, with 6 rows found inconsistent), nor did it re-open the δ value mismatches for D18/D21/D22—those represent a different class of error (numerical mismatch), whereas this paper addresses structural errors in the judgment criteria; the two should not be merged.
    6. The R26 in §5 is a methodological item abstracted from this example, validated only by this single instance in the current repository, not cross-validated across pools ①②③⑤. Per R11 “zero-parameter criteria can only be falsified, not verified,” R26 currently serves only the function of “detecting false precision” and cannot itself serve as proof that “the criteria are qualified.”

    VIII. This Cycle: Not Done (Boundary Preservation)

    • No modification of any published main text: high-sigma-lithology-elf-shielding-path.md, skin-depth-reference-table.md, skin-depth-basement-building-mu-metal-table.md, and pool4-high-sigma-lithology-elf-shielding-path.md (P4-EXP-004) all retain their original modification times (mtime).
    • skin-depth-recompute.py remains unmodified: (Recommended to reserve this for JOE’s post-approval separate execution, to avoid ambiguity regarding which specific alteration caused changes in the evidence-level categorization).
    • No σ measurements conducted; no new pathway established from “backrest” (a traditional consensus concept referring to a supportive spatial form) to “health” (net new pathways: 0); no items marked as accepted; no direct insertion of values 779/782; and no new or modified OC cron or heartbeat configurations.
    • The four recommendations in §6 remain strictly at the suggestion level and require JOE’s approval before any formal text corrections can be executed.

    9. Related Reading

    • High Conductivity Rock ELF Shielding Path (#42) — the original source of the K1–K4 classification table.
    • Skin Depth Reference Table (#16) — a reference table for δ numerical layers; the 20th review round identified six rows as internally inconsistent.
    • Backslope Terrain ELF Shielding Coefficient Field Test (#14) — the direct consumer of the classification table.
    • Mountain Sample σ Pre-screening SOP (#22) — the previous threshold version revised by this classification system.
    • Interval Statement Step-by-Step Evaluation (Round 52) — the interval layer processing; this paper addresses the criteria layer.

    This article is a draft. The conclusion that “the internal criteria structure of the document is self-inconsistent” represents a methodological audit and does not constitute any field-measured declaration or health recommendation regarding mountain electromagnetic shielding, building shielding, or human health.

    Before entering the final version 779, intake procedures must be completed; before entering version 782, AI-driven elements must be removed.

    Disclaimer: This article serves as an organized comparison 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 and treatment. The attenuation values, conductivity intervals, and classification thresholds mentioned herein are calculated via engineering formulas based on literature parameters, not field-measured values. Actual subsurface media exhibit layering, anisotropy, and frequency dispersion, meaning real-world results may deviate significantly. Any inference regarding “residential environment electromagnetic conditions and health” remains at the hypothesis level and must undergo independent experimentation and peer review before external publication. If you experience physical discomfort or health concerns, please consult a medical professional for diagnosis and treatment.

  • 合格区里的尺子本来就够不到那条线:κ_s 拼尽全力只造得出 2.5 个百分点,而它要拦的是 5 个

    Evidence level: C

    一个代理量能不能守住一条判据线,不取决于它测得准不准,取决于它自己的机制最多能造出多大的偏离——如果这个上界低于判据线,那它合格一万次也证明不了任何事。

    正文

    上一篇(#1386 / #1392)把判据从「两位判读者对主档占比的判断差 δ 是 1pp 还是 5pp」改成「先看两人打分的标准差之比 κ_s 在不在 0.90–1.10」。理由是 κ_s 测得更准、更省事。上一篇也指出 κ_s 对位置精确盲:纯平移时 κ_s 恒等于 1.000,而 δ 照样可能到 8pp。

    本篇要回答的是下一个问题,也是 backlog 里被写成一个假设的问题:κ_s 合格、但 δ 已经超过 5pp 的那个窗口,到底有多宽、偏向哪一边? backlog 当时的猜测是「窗口多半集中在 m>0.95 的高主档区,所以给高 m 区加一道附加门就行」。

    算完之后,这个猜测在两个方向上都是反的。原因是一个此前没人写下来的初等事实。

    一、δ 有一条与 κ_s 无关的天花板

    δ 的定义是「两位判读者把同一批受试者判进主档的人数比例之差」。那么它最多能是多少?把所有人推到主档,δ=(1−m)×100;把所有人在主档里全删掉,δ=−m×100。就这样,没有别的可能。

    于是 δ 的物理上界是 (1−m)×100 pp,与两位读者的尺度差、位置差、打分形状全都无关,只由基准比例 m 决定:

    m=0.60 时上界 40pp;m=0.75 时 25pp;m=0.85 时 15pp;m=0.90 时 10pp;m=0.95 时正好 5.00pp;m=0.99 时只剩 1.00pp。

    这条天花板把所有讨论改写了。当 m≥0.95 时,「δ>5pp」这件事在数学上就不可能发生——不是 κ_s 把它拦住了,是这个人群里根本没有 5 个百分点的空间可以挪。所以「m>0.95 是危险区」这个直觉,正好说反了:高 m 区是全表最安全的地方。

    二、反向门:m 越低越必须报 δ

    在 m<0.95 的每一格,κ_s 合格都不可能推出 δ<5pp。给定 m 与 κ_s,让读者整体上下挪一点(纯位置错位 b,以潜变量标准差为单位),δ 就会跨过 5pp,而 κ_s 纹丝不动:

    基准 m 该 m 下 δ 上界 b 需偏移多少 SD 才到 δ=+5pp(κ_s=1.00,完美合格)
    0.75 +25.0pp 0.167 SD
    0.85 +15.0pp 0.245 SD
    0.90 +10.0pp 0.363 SD
    0.93 +7.0pp 0.578 SD
    0.95 以上 ≤5.00pp 不可能

    也就是说:在最常见的低主档场景(m=0.75,即「约四分之一的人被判为有明显变化」),读者只要把整条尺子整体挪六分之一个标准差,δ 就到 5pp,而 κ_s 会诚实地报 1.000。这个错位量小到任何校准训练都不会当回事。

    三、真正致命的一条:κ_s 的探测上界低于判据线

    上面是「κ_s 看不见 b」。但更根本的问题是:κ_s 自己能动的那一部分,本来也造不出 5pp。

    在合格区 κ_s∈[0.90,1.10] 内、只让尺度变(b=0,纯尺度错位),δ 的最大绝对值是:

    m=0.75 时 2.32pp;m=0.85 时 2.53pp(全表最大);m=0.90 时 2.28pp;m=0.95 时 1.74pp;m=0.99 时 0.72pp。

    把这条曲线说白:κ_s 这个机制在整个合格区里,无论怎么扫,最多只能解释 2.5 个百分点的 δ。 而它被派去守的是一条 5pp 的线。探测上界不到判据阈值的一半——这不是精度问题、不是样本量问题、不是「多做几次重复判读」能解决的问题,是结构性无覆盖:κ_s 合格所覆盖的那部分 δ 空间,与「δ>5pp」那部分空间在合格边界上几乎不重叠(重叠只在 m 极靠近 0.95、δ 勉强够到 5.0pp 的一条缝里)。

    顺带一个可以核验的闭合性:κ_s 与 δ 之间不是独立的,存在一条「零脊线」b = c(1−a)(c 是该 m 的潜变量阈值)。在这条线上,无论 a 偏离 1 多远、κ_s 读到 0.80 还是 1.30,δ 恒等于 0.0000pp。也就是说 κ_s 的变化可以在 δ 完全不变的情况下任意大幅摆动——这条脊线就是「κ_s 合格」与「δ 合格」两张判据之间那条永远对不齐的缝。

    四、「那再补报 kappa 不就行了」——也不行

    常规的补救是「两个都报,看 κ_obs 有没有掉下来」。算一遍就知道这条也堵不住。

    在单调阈值模型下,δ=+10.53pp(m=0.75,κ_s=0.92、b=+0.30 SD)时,κ_obs 是 0.673;δ=+7.91pp 时 0.76;δ=+5.01pp 时 0.86。在全参数网格里,δ>5pp 的格子共有 618 个,其中 361 个格子的 κ_obs 高于 0.60(最高 0.891)。原因很清楚:κ_obs 在单调模型下只是 |δ| 的另一个刻度,它不提供关于「哪个自由度偏了」的任何新信息,所以它只能在 |δ| 很大(≈11pp 以上)时才报警。

    结论是:κ_s 守不住这条线,κ_obs 也守不住,两个一起报还是不构成拦截。

    五、这道门该往哪边设

    正确的附加门不是「高 m 区强制报 δ」,而是反过来:

    • m≥0.95:δ>5pp 数学上不可达,δ 可正式降为事后量,不必强制;
    • m<0.95:δ 一律必报,且报告时必须同时给 m——因为同一个 b 在不同 m 上造成的 δ 完全不同;
    • m<0.85:这是最危险的一段。此处 κ_s=1.00 只需 0.17–0.25 SD 的位置错位就能把 δ 推过线,而 δ 上界还有 15–25pp 的余量。凡落在这段的判读研究,κ_s 读数不得作为「不必报 δ」的依据。

    这一条对气功外气研究特别要紧:这类试验里「有响应者比例」通常就在 20%–50% 区间,m≈0.5–0.75——正好是全表 δ 上界最高、κ_s 最无力的那一段。换句话说,本领域最典型的样本构型,恰好撞在代理量最失效的位置上。

    证据与局限

    已验证(本文内部推导,可独立复算):δ 的物理上界 (1−m)×100;δ=+5pp 所需位置错位量随 m 单调上升(0.167→0.578 SD);κ_s 合格区内纯尺度 δ 上界 ≤2.53pp;零脊线 b=c(1−a) 上 δ≡0;δ>5pp 且 κ_obs>0.60 的格子占比 58.4%。

    部分验证:κ 类系数受边际分布(prevalence)与观察者偏倚(bias)双重影响、单报一个系数会误导,这一点有直接的文献支撑(PMID 8501467 明确提出应分列偏倚指数与流行度指数;PMID 25168681 在 2×2 表的系统比较中显示不同边际构型下各系数行为分化、并给出 B 统计量在两类悖论下更稳定的证据)。本文的贡献是把这条定性警告在「阈值模型 + 仿射错位」的具体参数化下算成了闭式数字。

    待验证:① 本文全程假定两位读者的判读是同一个潜变量上的单调阈值(无交互、无个体级差异),真实判读里读者与受试者可能有交互项,交互会改变 δ 的可达上界;②「b=0.167 SD 的位置错位在实践中算不算小」没有本领域实测——需要先测出判读者在重复判读之间的位置漂移量级;③ 低 m 高危这一结论依赖于 m 的准确已知,而 m 本身是从判读结果估的,估 m 的误差会传导到门槛上;④ 本文未处理「两位读者对同一批人判读顺序不同」造成的顺序效应。

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 25168681, PMID 40839487

    L1 网络状态:CORRESPONDENCE_NETWORK 边表全部边止于 L1 同构层、状态 open、现代证据栏多为「待补」;结构层统计「L2 以上边数 | 有「操作+观测量+阈值」的边 | 0」。全文件检索「判读 / 信度 / 一致率 / 尺度 / 位置 / 偏倚 / 代理 / 斜率 / 回归 / kappa / κ / 标准差 / 阈值 / 边际」共零命中,EDGE_TABLE_science-yixue 47 个概念条目中亦无「判读一致性」任何条目。该主题尚未进入 L1 对应网络,故本篇封顶 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 ruler in the qualified zone was already too short to reach that line: κ_s exerted its utmost effort yet could only produce 2.5 percentage points, whereas it needed to block 5.

    Evidence level: C

    Whether a proxy measure can hold a decision criterion line does not depend on how accurately it measures, but on the maximum deviation its mechanism can produce. If this upper bound lies below the criterion line, then qualifying a proxy measure one hundred thousand times proves nothing.

    Body

    The previous post (#1386 / #1392) revised the criterion from “the difference δ between two readers’ proportions of main cases (1pp vs. 5pp)” to “whether the ratio of standard deviations κ_s falls within 0.90–1.10”. The rationale was that κ_s measures more accurately and is more convenient. The previous post also noted that κ_s is blind to positional precision: under pure translation, κ_s remains constant at 1.000, yet δ can still reach 8pp.

    This post addresses the next question, which was formulated as a hypothesis in the backlog: If κ_s qualifies but δ exceeds 5pp, how wide is this window and which side does it favor? The backlog’s initial guess was “the window mostly concentrates in the high main-case region (m>0.95), so adding an extra gate for high m regions would suffice.”

    After calculation, this guess is reversed in both directions. The reason is an elementary fact that had not been written down before.

    I. δ Has a Ceiling Independent of κ_s

    δ is defined as “the difference in the proportion of subjects assigned to main cases by two readers.” What is its maximum possible value? Pushing everyone into main cases yields δ=(1−m)×100; removing everyone from main cases yields δ=−m×100. No other possibilities exist.

    Thus, the physical upper bound of δ is (1−m)×100 pp, determined solely by the baseline proportion m, regardless of scale differences, positional shifts, or scoring shapes:

    • m=0.60 → upper bound 40pp
    • m=0.75 → 25pp
    • m=0.85 → 15pp
    • m=0.90 → 10pp
    • m=0.95 → exactly 5.00pp
    • m=0.99 → only 1.00pp

    This ceiling fundamentally rewrites the discussion. When m≥0.95, the event “δ>5pp” is mathematically impossible—not because κ_s blocks it, but because there is simply no space of 5 percentage points available in that population. Therefore, the intuition “m>0.95 is the danger zone” is exactly backwards: the high m region is actually the safest part of the entire table.

    II. Reverse Gate: Lower m Requires Reporting δ

    In every cell where m<0.95, qualifying κ_s cannot imply δ<5pp. Given m and κ_s, shifting the readers’ scale up or down slightly (pure positional offset b, measured in latent standard deviations) causes δ to cross 5pp while κ_s remains unchanged:

    Baseline m Upper bound of δ under this m b offset required (in SD) to reach δ=+5pp (with κ_s=1.00, perfect qualification)
    0.75 +25.0pp 0.167 SD
    0.85 +15.0pp 0.245 SD
    0.90 +10.0pp 0.363 SD
    0.93 +7.0pp 0.578 SD
    >0.95 ≤5.00pp Impossible

    In other words, in the most common low-main-case scenario (m=0.75, meaning “about one-quarter of people are judged to have significant change”), readers need only shift their entire ruler by one-sixth of a standard deviation for δ to reach 5pp, while κ_s honestly reports 1.000. This positional shift is so small that any calibration training would overlook it.

    III. The Truly Fatal Issue: κ_s’s Detection Upper Bound Lies Below the Criterion Line

    The above shows that “κ_s cannot see b.” But the more fundamental problem is: the part of δ that κ_s can actually move never reaches 5pp.

    Within the qualifying region κ_s∈[0.90,1.10], considering only scale changes (b=0, pure scale offset), the maximum absolute value of δ is:

    • m=0.75 → 2.32pp
    • m=0.85 → 2.53pp (maximum across the entire table)
    • m=0.90 → 2.28pp
    • m=0.95 → 1.74pp
    • m=0.99 → 0.72pp

    Put plainly: κ_s’s mechanism, throughout the entire qualifying region, can at most explain 2.5 percentage points of δ. Yet it is tasked with guarding a 5pp line. The detection upper bound falls below half the criterion threshold—this is not an accuracy issue, nor a sample size issue, nor something solvable by “repeating readings multiple times.” It is structural non-coverage: the δ space covered by qualifying κ_s barely overlaps with the “δ>5pp” region (overlap exists only in a narrow seam where m is extremely close to 0.95 and δ barely reaches 5.0pp).

    As a verifiable closure: κ_s and δ are not independent; there exists a “zero ridge line” b = c(1−a) (where c is the latent threshold for this m). Along this line, regardless of how far a deviates from 1 (whether κ_s reads 0.80 or 1.30), δ remains exactly 0.0000pp. This means κ_s can swing wildly while δ stays completely unchanged—this ridge line represents the permanent misalignment between “κ_s qualifies” and “δ qualifies.”

    IV. “Then Just Report kappa Too”—That Also Fails

    The conventional remedy is “report both and check if κ_obs drops.” Calculation shows this also fails.

    Under the monotonic threshold model:
    – When δ=+10.53pp (m=0.75, κ_s=0.92, b=+0.30 SD), κ_obs = 0.673
    – When δ=+7.91pp, κ_obs = 0.76
    – When δ=+5.01pp, κ_obs = 0.86

    Across the full parameter grid, there are 618 cells where δ>5pp, of which 361 have κ_obs above 0.60 (maximum 0.891). The reason is clear: under the monotonic model, κ_obs is merely another scale of |δ|; it provides no new information about “which degree of freedom shifted.” Thus, κ_obs only triggers an alarm when |δ| is large (≈11pp or more).

    Conclusion: Neither κ_s nor κ_obs can guard this line; reporting both together does not constitute interception.

    V. Which Way Should This Gate Face?

    The correct supplementary gate is not “force δ reporting in high m regions,” but rather the reverse:

    • m≥0.95: δ>5pp is mathematically unreachable; δ can formally be reduced to a post-hoc measure and does not need mandatory reporting;
    • m<0.95: δ must always be reported, and the report must include m—because the same b produces completely different δ values at different m levels;
    • m<0.85: This is the most dangerous segment. Here, even κ_s=1.00 requires only 0.17–0.25 SD of positional shift to push δ over the line, while δ’s upper bound still leaves 15–25pp of margin. Any reading study falling in this segment must not use κ_s readings as grounds for “no need to report δ.”

    This rule is especially critical for qigong external qi studies: in such trials, the “proportion of responders” typically falls between 20%–50%, meaning m≈0.5–0.75—precisely the region where δ’s upper bound is highest and κ_s is weakest. In other words, the most typical sample configurations in this field exactly collide with the region where the proxy measure fails most severely.

    Evidence and Limitations

    Verified (derived internally in this paper, independently reproducible):
    – δ’s physical upper bound (1−m)×100;
    – the positional shift required for δ=+5pp increases monotonically with m (0.167→0.578 SD);
    – within κ_s’s qualifying region, pure-scale δ upper bound ≤2.53pp;
    – along the zero ridge line b=c(1−a), δ≡0;
    – 58.4% of cells have both δ>5pp and κ_obs>0.60.

    Partially verified:
    – κ-type coefficients are doubly affected by marginal distribution (prevalence) and observer bias; reporting only one coefficient is misleading. This has direct literature support: PMID 8501467 explicitly calls for separate reporting of bias index and prevalence index; PMID 25168681 shows in a systematic comparison of 2×2 tables that coefficient behaviors diverge under different marginal configurations and provides evidence that B statistics are more stable under both types of paradoxes. This paper’s contribution is converting this qualitative warning into closed-form numbers under the specific parameterization of “threshold model + affine offset.”

    To be verified:
    1. This paper assumes throughout that both readers’ judgments are monotonic thresholds on the same latent variable (no interaction, no individual-level differences); in real judgments, readers and subjects may have interaction terms, which would alter δ’s reachable upper bound;
    2. “Is b=0.167 SD considered small in practice?” lacks empirical measurement in this field—first, the magnitude of positional drift between repeated readings must be measured;
    3. The conclusion that low m is high-risk depends on accurate knowledge of m, yet m itself is estimated from reading results; error in estimating m propagates to the threshold;
    4. This paper does not address order effects caused by different reading orders of the same subjects by two readers.

    Related Reading

    Evidence level: C
    References: PMID 8501467, PMID 25168681, PMID 40839487

    L1 Network Status: All edges in the CORRESPONDENCE_NETWORK edge table terminate at the L1 isomorphic layer, state open, with most modern evidence columns marked “to be filled”; structural statistics show “number of edges above L2 | edges with ‘operation + observable + threshold’ | 0“. A full file search for “reading / reliability / consistency rate / scale / position / bias / proxy / slope / regression / kappa / κ / standard deviation / threshold / marginal” yields zero hits; among the 47 concept entries in EDGE_TABLE_science-yixue, there is no entry for “reading consistency.” This topic has not yet entered the L1 corresponding network, hence this paper 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.

  • 那条 ρ=0.5 的红线,画的不是危险,是收益:一条从未推导过的直觉,被算出来之后方向整个反了

    Evidence level: C

    1388 说「ρ>0.5 就该亮红牌」。把折扣的收益和代价放进同一个决策框架算下来,ρ 越高,打折越划算,不是越危险。那条红线不是错的线,是画错了变量——它真正在测的是「你的检验还有没有功效」,不是「你会不会判错」。

    正文

    一、上一轮那句话是我拍的,而且拍反了

    1388 在第三节末尾写了这么一句:ρ>0.5 时除了打折,还要额外声明「可比性存疑」。理由很顺:折价过半就值得警惕。

    但「折价过半」和「危险」是两件事。折价过半说明门槛降了一半,门槛降低既可能是「把该放行的放行了」(收益),也可能是「把不该放行的也放行了」(代价)。到底哪头大,取决于被比较的两份判读里有多大比例真有差——这个比例在一致性研究里没人报过,我也没算过就下了判断。

    现在把它算清。设定很干净:T = κ̂₁ − κ̂₂ 近似正态,真差为零时 T ~ N(0, 2(1−ρ)SE²)。两条规则:不打折用阈值 1.96·√2·SE,打折用 1.96·√(2−2ρ)·SE。两者差一个 √(1−ρ)。

    先看一个让人意外的事实:如果 ρ 是精确已知的,打折后的假阳性率恒等于 5.000%——不是 4.4%,不是 2.1%,是精确的 5.000%,ρ 从 0 走到 0.95 一动不动。因为打折用的本来就是正确的标准误,1.96 这个系数在正确标准误上就是 5%。

    也就是说,打折从来不会制造假阳性。真正在制造假阳性的是不打折——但方向是反的:不打折让阈值变得过严,假阳性率从 5% 掉到 ρ=0.5 时的 0.56%、ρ=0.9 时的 0.0002%。

    二、收益与代价的全表

    固定真差 = 1 个独立 SE 单位(下面再谈为什么这个点最关键):

    ρ 打折扣功效 不打折功效 收益 假阳性代价 临界先验 π*
    0.20 20.09% 14.20% +5.89pp +2.16pp 0.268
    0.30 22.30% 12.58% +9.72pp +3.08pp 0.241
    0.50 29.30% 8.73% +20.57pp +4.44pp 0.178
    0.70 44.67% 3.98% +40.69pp +4.97pp 0.109
    0.90 88.54% 0.12% +88.42pp +5.00pp 0.054

    代价那一列封顶在 5pp——因为它就是「从不打折的 5% 涨回正确的 5%」。收益那列没有上限,ρ 越大涨得越凶。

    π* 的定义是:被比较的判读里真有差的比例超过多少,打折才净赚。ρ=0.5 时这个门槛是 0.178。也就是说,只要「三个比较里有一个真的有差」,打折就划算。ρ=0.9 时门槛降到 0.054——十八个比较里有一个真有差就够本。

    所以 ρ 越高,打折越划算。红线该画在低 ρ,不在高 ρ。

    三、π* 的形状:为什么这不可能是单调的

    如果 π* 随 ρ 单调下降,那「ρ 越大越该打折」就是普遍规律。它不是——它在 z 上有个谷,然后反弹。

    真差 z ρ=0.3 ρ=0.5 ρ=0.7 ρ=0.9
    0.5 0.383 0.332 0.254 0.124
    1.0 0.241 0.178 0.109 0.054
    1.5 0.178 0.127 0.079 0.051
    2.0 0.173 0.135 0.104 0.100
    2.5 0.223 0.212 0.240 0.533
    3.0 0.360 0.428 0.635 0.990

    π* 的最低点落在 z≈1.3–1.8(ρ 越大越往前挪)。越过 z≈2.5 之后急升,ρ=0.9 在 z=3 时到 0.99——几乎没有任何先验能让打折划算。

    机制很直白:折扣把阈值从 1.96√2 拉到 1.96√(2−2ρ),这个平移量在高 ρ 时极大(ρ=0.9 时阈值只剩三分之一)。当真差本身已经大到能压过原阈值时,打折不再提供收益,只提供额外的假阳性。 收益是对「差一点就够」的人群起作用的——一旦真差本来就够,折扣就纯亏。

    这就给出一条可直接用的操作规则:先估 Δ/D_min 这个比值,比值在 0.7–2.5 之间才谈打折;超过 2.5 一律不打,打了纯亏。

    四、这条红线的真身:它是个功效告警

    把 ρ=0.5 这个点单独拎出来看,它到底在说什么?

    不打折的功效在 ρ=0.5 时是 8.73%。八个百分点的功效,意味着一份真有中等差值的比较,九成会被判成「不可辨」。ρ=0.7 时功效 3.98%,ρ=0.9 时 0.12%——这个检验已经死了。

    所以 ρ>0.5 真正标记的是「你的检验已经失去分辨能力」,不是「你的结果不可信」。这两个警告要求的动作相反:前者要求必须打折(或者干脆承认测不出来),后者要求不要打折。#1388 把前者写成了后者。

    而 ρ 在现实中集中在什么位置?#1388 那批仿真里,最常见的场景(同批受试者、换一组读者对)ρ 中位数只有 0.111,36 格里 5 格为负。ρ 天然就低,高 ρ 只出现在「同一对读者、隔一段时间再判一遍」这种特定设计里。 换句话说,这条红线在实际数据上几乎从不触发——它是一道针对罕见情形的、方向还写反了的门。

    五、还有个更实在的问题:ρ 根本估不准

    上面全部计算都假定 ρ 已知。它是估的,σ(ρ̂)≈(1−ρ²)/√n。

    用 1.645 个标准误的高估(双侧检验里足以把 FPR 推过 10% 的量级)反解所需精度:

    真 ρ 允许 σ(ρ̂) 自估所需 n
    0.30 ≤0.1258 ≥53
    0.50 ≤0.0898 ≥70
    0.70 ≤0.0539 ≥90
    0.90 ≤0.0180 ≥112

    n=60 时实测:ρ=0.5 的 FPR 均值 5.57%、95 分位 10.54%、6.2% 的批次会冲破 10%;ρ=0.9 时均值 5.98%、95 分位 13.31%、11.7% 的批次冲破 10%。#1388 表里那个「ρ=0.256 时 FPR 从 2.10% 涨到 4.40%」,落在 5.57% 均值附近的抽样波动范围内——它不是 ρ 造成的,是有限样本造成的。上一篇把抽样噪声读成了机制。

    六、没有先验怎么办:用「对半赌」这条退路

    π* 里那个 π 是真差先验,实务中确实不知道。退一步取 π=0.5(完全无知,值得与不值得打平),得到的边界极其宽松:

    ρ 打折划算的 z 区间(π=0.5)
    0.2 0.05 – 3.38
    0.5 0.05 – 3.11
    0.9 0.05 – 2.48

    ρ=0.5 时,只要 Δ < 3.11·D_min,打折就值得。 这是一条比 π* 粗糙但不用估先验的规则,覆盖面比绝大多数实际研究都宽。

    取 π=0.25(「四个里有一个真有差」,判读研究里偏保守的合理值)时,区间收窄成 ρ=0.5 的 [0.72, 2.62]。下界 0.72 这条线值得记住:当真差小到连 D_min 的七成都不到时,打折也救不回来——此时问题不是门槛,是效应本身就太细。

    证据与局限

    已验证:① 精确已知 ρ 时,折扣规则的假阳性率恒为 5.000%,与 ρ 无关(ρ∈{0,0.3,0.5,0.7,0.9,0.95} 逐点验算);② 不打折的假阳性率随 ρ 单调塌陷(5.00%→0.56%→0.0002%),即折扣在修的是「过严」而非「过松」;③ π* 在 z 上非单调,谷底落在 z≈1.3–1.8,z>2.5 后急升,这是折扣阈值平移量与真实效应竞争的必然结果,不是数值巧合;④ 所需 ρ̂ 精度与 n 的换算式由 (1−ρ²)/√n 直接反解。

    部分验证:① z 的定义用了「以独立 SE 为单位」,而真实研究的 SE 依赖 m、δ、n 一组参数,把 z 单一化是简化;② π* 曲面是正态近似下算的,小样本或高 ρ 时 T 的分布偏离正态(这正是第五节 FPR 分布右尾变厚的来源);③ ρ̂ 的方差近似式 (1−ρ²)/√n 假定双变量正态,实际 kappa 的联合分布未必满足。

    待验证:① 真差先验 π 在任何真实一致性研究里都没被估过,本文所有「划算/不划算」的判决都压在它上面,而它恰好是唯一没有数据支撑的量;② 「Δ/D_min 先算再决定折不打折」这条流程要求先知道 D_min,而 D_min 里的 δ、ρ 两个输入本身都带误差,误差传播从未算过;③ 本文假定两位读者的 kappa 是同一批受试者上的成对比较,若研究里比较的是「两组不同受试者、不同读者」的结构,ρ 的定义直接失效、整张表不适用;④ 折扣与不折扣的判决用的是同一个 5% 名义水平,若研究采用更严的水平(如 1%),π* 曲面整体要重算。

    L1 网络状态:CORRESPONDENCE_NETWORK.md 边表止于 E-050「六十甲子纳音 | 对应 | 60 个频率」(层级 L1、状态 open,现代证据栏原文「尚无任何可测来源」);全文件检索「判读 / 信度 / 一致率 / 相依 / 净收益 / 阈值 / 假阳性」仅命中第 79 行层级定义与第 159 行「L2 以上边数 | 0」。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 17099194, PMID 26810254, PMID 25155798, PMID 40405411, PMID 38841381, PMID 36872574, PMID 41239778, PMID 36123642, PMID 18430244, PMID 3337089

    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。

    That ρ=0.5 red line, which was drawn not as danger but as gain: an intuition never derived before, once calculated, completely reversed the direction

    Evidence level: C

    1388 states that “ρ > 0.5 should trigger a red flag.” When the discounted benefit and cost are calculated within the same decision framework, higher ρ makes discounting more advantageous, not more dangerous. That red line is not the wrong line; it misidentifies the variable—it actually measures “whether your test retains power,” not “whether you will make a false negative.”

    Body

    1. That previous statement was mine, and I got it backwards

    1388 wrote at the end of section three: “When ρ > 0.5, besides discounting, one must additionally declare ‘comparability is questionable’.” The reasoning seems straightforward: discounting beyond half is worth warning about.

    However, “discounting beyond half” and “danger” are two distinct matters. Discounting beyond half indicates the threshold has halved. A lowered threshold can either mean “allowing those who should be allowed” (gain) or “allowing those who should not be allowed” (cost). Which side dominates depends on the proportion of comparisons that truly differ—a ratio never reported in consistency studies, and which I did not calculate before issuing the judgment.

    Now let us calculate it precisely. The setup is clean: T = κ̂₁ − κ̂₂ approximates a normal distribution; when the true difference is zero, T ~ N(0, 2(1−ρ)SE²). Two rules: no discounting uses threshold 1.96·√2·SE, discounting uses 1.96·√(2−2ρ)·SE. The difference between them is √(1−ρ).

    First, consider a surprising fact: if ρ is precisely known, the false positive rate after discounting is constant at 5.000%—not 4.4%, not 2.1%, but exactly 5.000%, regardless of ρ moving from 0 to 0.95. This is because discounting uses the correct standard error, and the coefficient 1.96 applied to the correct standard error yields exactly 5%.

    Thus, discounting never creates false positives. What creates false positives is not discounting—but the direction is reversed: not discounting makes the threshold overly strict, driving the false positive rate from 5% down to 0.56% when ρ=0.5, and down to 0.0002% when ρ=0.9.

    2. Full table of gain and cost

    Fixed true difference = 1 independent SE unit (we discuss why this point is critical below):

    ρ Discounted power Undiscounted power Gain False positive cost Critical prior π*
    0.20 20.09% 14.20% +5.89pp +2.16pp 0.268
    0.30 22.30% 12.58% +9.72pp +3.08pp 0.241
    0.50 29.30% 8.73% +20.57pp +4.44pp 0.178
    0.70 44.67% 3.98% +40.69pp +4.97pp 0.109
    0.90 88.54% 0.12% +88.42pp +5.00pp 0.054

    The cost column caps at 5pp—because it represents “the 5% of never discounting returning to the correct 5%.” The gain column has no upper limit; the higher ρ, the more it rises.

    The definition of π* is: above what proportion of true differences must exist in the compared judgments for discounting to be net profitable. When ρ=0.5, this threshold is 0.178. That means as long as “one out of three comparisons truly differs,” discounting is worthwhile. When ρ=0.9, the threshold drops to 0.054—one out of eighteen comparisons truly differing is enough.

    Therefore, the higher ρ, the more advantageous discounting becomes. The red line should be drawn at low ρ, not high ρ.

    3. The shape of π*: why this cannot be monotonic

    If π* decreased monotonically with ρ, then “the higher ρ, the more one should discount” would be a universal rule. It is not—it has a valley in z, then rebounds.

    True difference z ρ=0.3 ρ=0.5 ρ=0.7 ρ=0.9
    0.5 0.383 0.332 0.254 0.124
    1.0 0.241 0.178 0.109 0.054
    1.5 0.178 0.127 0.079 0.051
    2.0 0.173 0.135 0.104 0.100
    2.5 0.223 0.212 0.240 0.533
    3.0 0.360 0.428 0.635 0.990

    The minimum of π* occurs at z≈1.3–1.8 (the valley shifts forward as ρ increases). Beyond z≈2.5 it rises sharply; at ρ=0.9 and z=3, π*=0.99—almost no prior can make discounting worthwhile.

    The mechanism is straightforward: discounting shifts the threshold from 1.96√2 to 1.96√(2−2ρ); this shift is large at high ρ (at ρ=0.9 the threshold shrinks to one-third). When the true difference itself is already large enough to surpass the original threshold, discounting no longer provides gain; it only adds extra false positives. Gain applies to the population “just barely sufficient”—once the true difference is already sufficient, discounting is purely a loss.

    This yields an immediately usable operational rule: first estimate the ratio Δ/D_min; discounting is discussed only if this ratio lies between 0.7 and 2.5; above 2.5, never discount, as discounting is purely a loss.

    4. The true identity of this red line: it is a power warning

    Isolating the point ρ=0.5, what does it signify?

    The power of not discounting at ρ=0.5 is 8.73%. Eight percent power means that a comparison with a medium true difference will be judged “indistinguishable” in ninety percent of cases. At ρ=0.7, power is 3.98%; at ρ=0.9, it is 0.12%—this test is dead.

    Therefore, ρ > 0.5 truly marks “your test has lost its ability to distinguish,” not “your results are untrustworthy.” These two warnings demand opposite actions: the former requires must discount (or simply admit it cannot be measured), while the latter requires do not discount. #1388 wrote the former as the latter.

    Where does ρ naturally lie in reality? In #1388’s simulations, the most common scenario (same subjects, different readers) has a median ρ of only 0.111, with 5 out of 36 cells negative. ρ is naturally low; high ρ appears only in specific designs such as “the same pair of readers, re-evaluated after some time.” In other words, this red line is almost never triggered by actual data—it is a door aimed at rare cases, with its direction written backwards.

    5. A more practical problem: ρ cannot be estimated accurately

    All calculations above assume ρ is known. It is estimated, with σ(ρ̂)≈(1−ρ²)/√n.

    Solving for required precision using 1.645 standard errors (enough to push the false positive rate over 10% in a two-sided test):

    True ρ Allowed σ(ρ̂) Required n for self-estimation
    0.30 ≤0.1258 ≥53
    0.50 ≤0.0898 ≥70
    0.70 ≤0.0539 ≥90
    0.90 ≤0.0180 ≥112

    With n=60, actual results show: for ρ=0.5, the mean false positive rate is 5.57%, the 95th percentile is 10.54%, and 6.2% of batches exceed 10%; for ρ=0.9, the mean is 5.98%, the 95th percentile is 13.31%, and 11.7% of batches exceed 10%. The entry in #1388’s table stating “when ρ=0.256, the false positive rate rises from 2.10% to 4.40%” falls within the sampling fluctuation around the 5.57% mean—it is not caused by ρ, but by finite sample size. The previous article misread sampling noise as mechanism.

    6. What if there is no prior: using the “half-or-all bet” fallback

    The π in π* is the true difference prior, which is indeed unknown in practice. Taking a step back and setting π=0.5 (complete ignorance, where worth discounting equals not worth discounting) yields an extremely loose boundary:

    ρ z interval where discounting is worthwhile (π=0.5)
    0.2 0.05 – 3.38
    0.5 0.05 – 3.11
    0.9 0.05 – 2.48

    When ρ=0.5, discounting is worthwhile as long as Δ < 3.11·D_min. This is a rule coarser than π* but does not require estimating the prior, and its coverage is wider than most actual studies.

    Taking π=0.25 (“one out of four truly differs,” a conservative reasonable value in judgment studies) narrows the interval to [0.72, 2.62] for ρ=0.5. The lower bound of 0.72 is worth remembering: when the true difference is less than seventy percent of D_min, discounting cannot save it—here the problem is not the threshold, but the effect itself is too fine.

    Evidence and Limitations

    Verified: ① When ρ is precisely known, the false positive rate of the discounting rule is constant at 5.000%, independent of ρ (verified point-by-point for ρ∈{0, 0.3, 0.5, 0.7, 0.9, 0.95}); ② the false positive rate of not discounting collapses monotonically with ρ (5.00% → 0.56% → 0.0002%), meaning discounting fixes “overly strict” rather than “overly loose”; ③ π* is non-monotonic in z, with the valley at z≈1.3–1.8 and a sharp rise beyond z>2.5, which is the inevitable result of competition between the discount threshold shift and true effect size, not numerical coincidence; ④ the conversion formula for required ρ̂ precision and n is directly derived from (1−ρ²)/√n.

    Partially verified: ① the definition of z uses “independent SE units,” while real-world study SE depends on parameters m, δ, n; simplifying z to a single unit is an approximation; ② the π* surface is calculated under normal approximation; in small samples or high ρ, the distribution of T deviates from normal (this is precisely the source of the thick right tail in section five’s FPR distribution); ③ the variance approximation for ρ̂, (1−ρ²)/√n, assumes bivariate normality; the joint distribution of actual kappa may not satisfy this.

    To be verified: ① the true difference prior π has never been estimated in any real consistency study; all judgments of “profitable/not profitable” in this paper rest on it, yet it is the only quantity without data support; ② the procedure “first compute Δ/D_min, then decide whether to discount” requires knowing D_min in advance, but the inputs δ and ρ inside D_min themselves carry error, and error propagation has not been calculated; ③ this paper assumes the comparison of two readers’ kappa is a paired comparison on the same subjects; if the study compares “two groups of different subjects and different readers,” the definition of ρ fails directly, rendering the entire table inapplicable; ④ the decision between discounting and not discounting uses the same nominal 5% level; if the study adopts a stricter level (e.g., 1%), the π* surface must be recalculated entirely.

    L1 Network Status: The edge list in CORRESPONDENCE_NETWORK.md stops at E-050 “Sixty Jiazi Na Yin | Correspondence | 60 frequencies” (level L1, status open, original text in modern evidence column states “no measurable source yet“); searching the entire file for “judgment / reliability / consistency / dependence / net gain / threshold / false positive” hits only row 79 (level definition) and row 159 (“L2 and above edge count | 0”). This topic has not yet entered the L1 correspondence network.

    Related Reading

    Evidence Level: C
    References: PMID 17099194, PMID 26810254, PMID 25155798, PMID 40405411, PMID 38841381, PMID 36872574, PMID 41239778, PMID 36123642, PMID 18430244, PMID 3337089

    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.