Category: 科学易学

  • 审计工具该信谁:把”主张方”判定从文件名猜改成目录角色定(R14 正向化)

    Evidence level: C

    一句话结论:本库的跨文件一致性审计工具,判断”哪个文件在主张一个数值、哪个文件只是在引用”时,原来靠文件名正则猜(文件名含 audit/governance/backlog 才算”引用方”)。这个办法有一个结构性后门:审计文、实验设计、台账只要文件名不像审计文,就被当成”主张方”——于是工具会指控”忠实记录错误”的台账和”忠实引用正本”的实验设计,而不是指控真正写错的正本。这一轮把它改成按目录角色定:03-sites/ 是正本(网站材料落点),08-ops/ 是内部工作区(实验设计/治理卡/台账,一律引用方)。改完,跨文件假阳性少 1 条,剩下的 5 条违反全部指向两个已知伤口。

    一、这个工具在干什么

    池④ 有一张”趋肤深度—衰减”对照表(δ 表),被十几个下游课题当共同基准。为了防这张表继续带错数值传播,我们写了一个跨文件扫描器(delta-table-scan.py),它做三件事:

    • R1/R2/R3:把任意 md 里的 δ / A(dB) 数值断言抽出来,用”该断言自己声明的 σ/μr”独立复算,判定是否自洽。
    • R5(同文件内):同一个文件里,同一材料同一频点下若出现两组”厚度—衰减”,两者之比必须相等(A ∝ d,零参数)。
    • R5X(跨文件):两个文件对同一材料同一频点给出互不相容的数字 → 报违反。

    R5X 是 R5 的跨文件版,也是它第一次真正抓到本库的主体缺陷形态——双语镜像 + 审计文互引的库里,同一个数字经常同时出现在正本、审计文、backlog 三处。

    二、问题:按文件名猜”谁在主张”会冤枉谁

    R5X 一开始遇到 32 对假阳性,收敛到 6 对靠的是两条规则,其中一条叫 R14「引用的不是主张」:审计文/台账里并排写的是”原文值 vs 复算值”,两数本就该不同;把它们当主张方,等于把”有人在引用错值”读成”这个文件也错了”。

    但 R14 当时是按文件名正则实现的:

    文件名含 audit / governance / backlog / report / registry / check / integrity / decision / queue / ledger → 不作主张方
    

    这个办法漏了两类文件,而且漏的都是”本应被信任”的文件:

    1. 审计文文件名不像审计文(第 17 轮撞上):pool4-cross-file-a-proportionality-contradiction.md 名字里没有 audit/backlog,但它确实是审计文——它把”散文 0.27 dB / 表 27 dB”这一对已知矛盾并排写进正文当例子,被工具当成主张方,一次报出 16 条假阳性。
    2. 实验设计、治理卡、台账(本轮撞上):08-ops/experiments/pool4-mountain-elf-typical-granite-transparent.md 是配套实验设计,它在引用正本的 0.27 dB,名字里也没有任何审计词 → 被当成主张方,与正本互比,凭空多出 1 条违反。

    这两类文件的共同点:它们不是网站正本,是内部工作区。文件名正则永远猜不全”谁是工作区”——靠猜,就会有一个漏网之鱼,而且漏的都是最该被保护的”忠实记录者”。台账越忠实记录错误,越会被自己的工具判成有错——这是一个鼓励”少记录”的反向激励,必须用结构消灭。

    三、修法:目录角色说了算,文件名不再参与

    把”谁在主张”从”按文件名猜”改成”按目录角色定”:

    def _is_quote_doc(rel):
        # 真实路径不在 03-sites/ 下(即 08-ops/ 内部目录)→ 内部工作区,一律不作主张方
        if os.sep in rel and not rel.replace(os.sep, "/").startswith("03-sites/"):
            return True
        return bool(QUOTE_DOC.search(os.path.basename(rel)))  # 裸文件名(夹具)仍按正则
    

    语义很直白:

    • 03-sites/ = 网站材料正本落点 → 主张方(该信它说的数,也该被审计)。
    • 08-ops/(pools / experiments / decisions / 台账) = 内部工作区 → 引用方(它写的数是在引用正本,不是新主张,不参与互比)。
    • 裸文件名(selftest 夹具、临时文件)→ 退回文件名正则(保留”审计文不算主张方”的测试语义)。

    为什么不是”看 frontmatter 有没有 pool 字段”:08-ops/experiments/ 下所有实验设计都没有 pool frontmatter(本轮已核),两种口径结果一致;但目录角色不依赖每个文件的 frontmatter 是否写全,也就不怕”某个实验设计以后补了 pool 字段”而悄悄变成主张方。目录是结构事实,frontmatter 是书写习惯——结构事实更稳。

    四、实测:违反 6→5,剩下的全是正本互比

    同一工具、同一扫描根(--root . 全库)重跑:

    指标 v1.6(上一轮) v1.7(本轮) 变化
    命中文件 / 断言总数 52 / 553 52 / 553 不变
    MISMATCH(OK/param/unit) 297(208/240/33) 297(208/240/33) 不变
    RANGE-unjudged 48 48 不变
    表格覆盖率 63/83 = 75.9% 63/83 = 75.9% 不变
    R5 违反 1 1 不变
    R5X 违反 6 5 −1
    R5X 配对 14 10 −4
    R5X 排除点 7 8 +1
    selftest 92/92 92/92 不变
    • 消除的 1 条:granite @ 7.83 Hz,A 端是实验设计 08-ops/experiments/pool4-mountain-elf-typical-granite-transparent.md L19(引用正本 0.27 dB)。这正是 #66 要修的后门。
    • 保留的 5 条,全部是正本之间互比:skin-depth-basement(L81/L82/L85)× skin-depth-reference-table(L126/L130/L131/L132)+ mountain-elf-typical-granite-transparent L81。与既有 D18/D21/D22 完全对应——basement §2 表花岗岩 δ 32 m → 17.99 km(差 562×),reference-table §三 6 行参数错配,同一根错的两个症状。
    • 配对 14→10 而违反只 −1:移出实验设计那 1 点后,与它相关的组合消失。剩下的 10 对全是正本间配对,5 条违反——违反密度不降反升(5/10 vs 6/14),正向化后的信号更干净。

    判定逻辑零改动的硬证据:MISMATCH 五行(OK / MISMATCH / param / unit / RANGE)数字全部未变——只改了 _is_quote_doc() 一个函数,R1/R2/R3/R5/R5X 的判读规则、阈值、排除规则一概未动。

    五、必须说清的三条边界

    第一,违反 6→5 不是”库内矛盾变少”。 是”工具对主张方的判定规则”修正了。库里的错还在原处(5 条违反指向的两处伤口,D18/D21/D22 已登记待 JOE 批)。

    第二,本检验只可证伪,不可验证。 比例对得上 ≠ 数值对。R5X 的”通过数”不得进任何质量指标(R11)。剩下 10 对配对、5 条违反——样本量仍不足以推出”全库矛盾率”(R12:分母必须报,且小到不能谈率)。

    第三,目录角色规则有一个新的已知盲区。 03-sites/ 内如果有一天出现”审计文”(文件名没命中正则),仍会被当主张方——但这条盲区比”08-ops 实验设计被当主张方”小得多:03-sites/ 是正本落点,按定义该信它、也该审计它。真正要防的是工作区冒充正本,现在由目录结构保证,不再靠文件名猜。

    六、证据与局限

    外部文献支撑的是公式前提,不是本库数字。

    • A(dB) ≈ 8.686·d/δ 与 δ = √(2/(ωμσ)) 是均匀导电半空间下的标准结果;其在岩石上的适用性受电导率频散限制:Revil (2013, Water Resour. Res., doi:10.1029/2012WR012700) 给出非饱和多孔介质有效电导率与介电常数在宽频段的模型。地物期刊不在 PubMed 收录范围 → 无 PMID → 如实标 E3。

    局限(必须与结论同读):

    1. 本检验判”不相容”仅指跨文件数字不满足 A ∝ d,不指真实地下 A 为另一值,也不指谁的 σ 引用错了。
    2. A ≈ 8.7d/δ 假设均匀半空间、无位移电流、μr/σ 为常数。真实山体分层、各向异性、σ 随频率弥散——公式本身在宽频段只是量级估算。
    3. 有效配对仅 10 组、正本仅 6 个文件 27 点,不足以外推任何”全库矛盾率”。
    4. 只覆盖等式断言;不等式断言与文字区间不参与。
    5. 03-sites/ 内若出现文件名不像审计文的审计文,仍是已知盲区(见第五节)。

    免责声明:本文是库内文本一致性的方法学说明,不是地球物理测量结论,不构成任何”靠山能否屏蔽 ELF”的实证判断,更不构成健康或医疗建议。任何据此做的采样决策,须先经 #14 阶段3 实测与 779 终版审核。

    七、下一步

    1. D18 + D1′ + D20 + D21/D22/D23 合并提请 JOE:全部同源(δ 数值错配 → 误判普通花岗岩”值得测” / 高 σ 岩体”无屏蔽”),一次批完,避免分散。
    2. #68:basement §2 表花岗岩 δ 格 32 m → 17.99 km(5 频点连带)——需与 reference-table 双改(R23:跨文件违反的两端都要查全)。
    3. 不要再做:R5X 扩成跨材料/跨 μr(物理不成立);词元重叠检测器(第 9 轮证伪);表头正则重写(第 12 轮证伪该归因);区间端点判读(R13);”通过数”当质量指标(R11)。

    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。

    Who to Trust in Audit Tools: Switching from Filename Guessing to Directory Role Determination (R14 Forwarding)

    Evidence level: C

    One-sentence conclusion: This repository’s cross-file consistency audit tool, which determines “which file asserts a value and which merely cites it,” previously relied on filename regex (files containing audit/governance/backlog were deemed “citing parties”). This approach contained a structural loophole: audit documents, experimental designs, and ledgers were incorrectly classified as “asserting parties” if their filenames did not resemble audit terms—causing the tool to falsely accuse “faithfully recording errors” in ledgers and “faithfully citing originals” in experimental designs, rather than the actual original document with errors. This revision shifts the decision to directory roles: 03-sites/ is designated as the original (website material drop-off point), while 08-ops/ serves as the internal workspace (experimental designs, governance cards, ledgers—all citing parties). After this change, cross-file false positives decreased by 1; the remaining 5 violations all point to two known wounds.

    I. What This Tool Does

    Pool ④ contains a “skin-depth vs attenuation” comparison table (δ-table), which serves as a common baseline for over a dozen downstream studies. To prevent this table from propagating incorrect values, we developed a cross-file scanner (delta-table-scan.py) that performs three functions:

    • R1/R2/R3: Extracts δ / A(dB) value assertions from any markdown file and independently recalculates them using the σ/μr values declared within each assertion, determining self-consistency.
    • R5 (within-file): If two sets of “thickness vs attenuation” data appear for the same material at the same frequency within one file, their ratio must be equal (A ∝ d, zero-parameter).
    • R5X (cross-file): If two files provide mutually incompatible numbers for the same material at the same frequency, report a violation.

    R5X is the cross-file version of R5 and represents its first true capture of the repository’s core defect pattern—bilingual mirroring + cross-referenced audit documents, where the same number frequently appears in three places: the original, the audit document, and the backlog.

    II. Problem: Filename-Based “Who Asserts” Misidentifies Trusted Documents

    R5X initially encountered 32 false positives, reduced to 6 using two rules, one of which is R14 “Citing ≠ Asserting”: audit documents and ledgers list “original value vs recalculated value” side-by-side; these two numbers should differ. Treating them as asserting parties means interpreting “someone cited an incorrect value” as “this file is also wrong.”

    However, R14 was originally implemented via filename regex:

    Filename contains audit / governance / backlog / report / registry / check / integrity / decision / queue / ledger → not an asserting party
    

    This approach missed two categories of files, both of which should be trusted:

    1. Audit documents with non-audit filenames (hit in round 17): pool4-cross-file-a-proportionality-contradiction.md lacks audit/backlog in its name, yet it is indeed an audit document—it places the known contradiction “prose 0.27 dB / table 27 dB” side-by-side in the body as an example. The tool treated it as an asserting party, generating 16 false positives.
    2. Experimental designs, governance cards, ledgers (hit in this round): 08-ops/experiments/pool4-mountain-elf-typical-granite-transparent.md is the accompanying experimental design, which cites the original’s 0.27 dB. Its name contains no audit terms → treated as an asserting party, creating 1 additional violation against the original.

    The commonality of these two file types: they are not website originals; they are internal workspaces. Filename regex can never fully guess “which is the workspace”—guessing creates a loophole, and the loophole always catches the most trusted “faithful recorders.” The more faithfully a ledger records errors, the more likely it is to be flagged as erroneous by its own tool—a reverse incentive discouraging thorough documentation, which must be eliminated structurally.

    III. Fix: Directory Roles Decide, Filenames No Longer Matter

    Shift the “who asserts” decision from “guess via filename” to “determine via directory role”:

    def _is_quote_doc(rel):
        # True if the real path is not under 03-sites/ (i.e., inside 08-ops/ internal directory) → internal workspace, never an asserting party
        if os.sep in rel and not rel.replace(os.sep, "/").startswith("03-sites/"):
            return True
        return bool(QUOTE_DOC.search(os.path.basename(rel)))  # Bare filenames (selftest fixtures) still use regex
    

    The semantics are straightforward:

    • 03-sites/ = website material original drop-off point → asserting party (trust its stated numbers; it must be audited).
    • 08-ops/ (pools / experiments / decisions / ledgers) = internal workspace → citing party (numbers written here cite the original; they are not new assertions, so they do not participate in cross-comparison).
    • Bare filenames (selftest fixtures, temporary files) → fallback to filename regex (preserve “audit documents are not asserting parties” test semantics).

    Why not check frontmatter for pool field: All experimental designs under 08-ops/experiments/ lack pool frontmatter (verified in this round); both approaches yield consistent results. However, directory roles do not depend on whether each file’s frontmatter is complete, avoiding the risk of an experimental design later adding a pool field and silently becoming an asserting party. Directory structure is factual; frontmatter is stylistic—factual structure is more stable.

    IV. Test Results: Violations Drop 6→5, Remaining All Original-Original Comparisons

    Re-run the same tool on the full repository (--root .):

    Metric v1.6 (previous round) v1.7 (this round) Change
    Files hit / total assertions 52 / 553 52 / 553 unchanged
    MISMATCH (OK/param/unit) 297 (208/240/33) 297 (208/240/33) unchanged
    RANGE-unjudged 48 48 unchanged
    Table coverage 63/83 = 75.9% 63/83 = 75.9% unchanged
    R5 violations 1 1 unchanged
    R5X violations 6 5 −1
    R5X pairs 14 10 −4
    R5X excluded points 7 8 +1
    selftest 92/92 92/92 unchanged
    • Eliminated 1 violation: granite @ 7.83 Hz, where the A-end is experimental design 08-ops/experiments/pool4-mountain-elf-typical-granite-transparent.md L19 (citing the original’s 0.27 dB). This is exactly the #66 loophole to fix.
    • Remaining 5 violations, all between originals: skin-depth-basement (L81/L82/L85) × skin-depth-reference-table (L126/L130/L131/L132) + mountain-elf-typical-granite-transparent L81. These correspond exactly to existing D18/D21/D22—basement §2 table granite δ 32 m → 17.99 km (off by 562×), reference-table §three 6 rows parameter mismatch, same root error manifesting in two symptoms.
    • Pairs drop 14→10 while violations only −1: removing the experimental design point eliminates its associated combinations. The remaining 10 pairs are all original-original, yielding 5 violations—violation density actually increases (5/10 vs 6/14), making the positively framed signal cleaner.

    Zero-change judgment logic evidence: All five MISMATCH categories (OK / MISMATCH / param / unit / RANGE) remain numerically unchanged—only _is_quote_doc() was modified. R1/R2/R3/R5/R5X judgment rules, thresholds, and exclusion criteria remained untouched.

    V. Three Boundaries That Must Be Clarified

    First, violations dropping 6→5 is not “fewer contradictions within the repository.” It reflects correction of the tool’s asserting-party decision rule. The errors in the repository remain in place (the 5 violations point to two wounds, D18/D21/D22 already logged pending JOE approval).

    Second, this test can only falsify, not verify. Proportional matching ≠ numerical agreement. R5X’s “pass count” cannot enter any quality metrics (R11). The remaining 10 pairs and 5 violations—sample size is still insufficient to infer “full repository contradiction rate” (R12: denominator must be reported, and it is too small to discuss rates).

    Third, the directory role rule has a new known blind spot. If an audit document appears inside 03-sites/ but its filename does not match the regex, it will still be treated as an asserting party—but this blind spot is far smaller than “08-ops experimental designs being treated as asserting parties”: 03-sites/ is the original drop-off point; by definition, it should be trusted and audited. What truly needs guarding is workspaces masquerading as originals, now ensured by directory structure rather than filename guessing.

    VI. Evidence and Limitations

    External literature supports formula premises, not repository numbers.

    • A(dB) ≈ 8.686·d/δ and δ = √(2/(ωμσ)) are standard results for uniform conducting half-space; their applicability to rock is limited by conductivity dispersion: Revil (2013, Water Resour. Res., doi:10.1029/2012WR012700) provides a model for effective conductivity and dielectric constant of unsaturated porous media across wide frequency bands. Since Earth Science journals are not indexed in PubMed, there is no PMID; thus, label E3 accurately.

    Limitations (must be read alongside conclusions):

    1. This test defines “incompatible” only as cross-file numbers failing A ∝ d, not implying the true subsurface A is another value, nor that whose σ was cited incorrectly.
    2. A ≈ 8.7d/δ assumes uniform half-space, no displacement current, constant μr/σ. Real mountain stratification, anisotropy, and σ dispersion with frequency mean the formula itself is only order-of-magnitude estimation across wide bands.
    3. Valid pairs only 10 sets, originals only 6 files with 27 points—insufficient to extrapolate any “full repository contradiction rate.”
    4. Covers only equality assertions; inequality assertions and textual intervals are excluded.
    5. 03-sites/ may contain audit documents with non-audit filenames (known blind spot, see Section V).

    Disclaimer: This document describes the methodology for textual consistency within the repository, not geophysical measurement conclusions. It does not constitute any empirical judgment on whether “basement can shield ELF,” nor does it provide health or medical advice. Any sampling decisions based on this must first undergo #14 stage 3 measurement and the final review of 779.

    VII. Next Steps

    1. Merge D18 + D1′ + D20 + D21/D22/D23 for JOE submission: all share the same root cause (δ value mismatch → misjudging ordinary granite as “worth measuring” / high σ rock as “no shielding”), so batch them together to avoid fragmentation.
    2. #68: basement §2 table granite δ grid 32 m → 17.99 km (5 frequencies involved)—requires dual revision of reference-table (R23: both ends of cross-file violations must be fully checked).
    3. Stop doing: expanding R5X to cross-material/cross-μr (physically invalid); token overlap detector (falsified in round 9); rewriting header regex (round 12 falsified this attribution); interval endpoint interpretation (R13); using “pass count” as quality metric (R11).
  • 排名翻转的实测版:把 ≥90 人的真实呼吸数据交给两套算法,看组间次序到底翻不翻

    Evidence level: C

    合成数据里发现的问题,必须回到真实数据里验证:同一批人的胸腹相位离散度,用两套算法各排一次名,组间次序究竟翻不翻、翻在哪一类人身上。

    为什么合成判据必须经过真实回算

    上一篇文章用合成数据标定了两套算法在跨零呼吸上的偏差方向:φ 是角度量,越接近 0° 越容易跨零,而跨零会把算术标准差放大、把圆形离散度(1−R)的期望值压低。但合成数据只能证明「偏差可能存在」,不能证明「偏差足以改变真实人群的排名」——真实数据上次序究竟翻不翻,至今没有任何实测。

    这正是本篇要补的洞。三种结局指向三种完全不同的处置:若整体次序高度一致(Kendall τ > 0.95),担忧只是「报告规范问题」,既往按算术 SD 报的结论可原样保留;若整体就不一致(τ < 0.8),所有按算术 SD 报的组间结论必须重算;若整体一致但分层后局部翻转,只需对翻转亚组局部重算。目前没有任何真实数字告诉我们该走哪条路。

    两条文献边界:偏差很大,但可能只伤特定亚组

    关于 φ 的算法偏差,文献给出两条边界。算法层面:同一批 φ 数据用六种时间域方法测出的结果可差 30° 以上,而交叉相关与最大线性相关能压到 5° 以内(PMID 12422344)——算法选型本身就能吃掉整个效应量。人群层面:胸腹异步角在 N3 睡眠与 REM 之间差 2.4 倍(PMID 29991414),真实人群亚组间差异往往远大于算法偏差——只要两套算法没把本就很大的组间差距排反,组间结论依然稳。

    两条边界给出可检验的预测:偏差集中在「同步最好(μ < 15°)、呼吸最慢(n < 12)」的人群里——这两类人最容易跨零、最容易被算术 SD 放大。真实回算就是检验这个预测:若翻转只发生在这两个亚组,既往文献只需局部重算;若翻转到处都有,才是真正的全盘性问题。

    一次纯二次分析:18 个格子逐一检验

    方法上不新招人:取已有成人静息 RIP 库(同体位、≥10 分钟、带呼吸次数),共 ≥90 人,覆盖修炼者与新手两组。对每人稳定段,算术口径取逐拍 φ 的标准差,圆形口径取 1−R/圆方差/von Mises κ 之一(固定单一口径)。按 φ 均值 μ 三层(0°–15°/15°–30°/>30°)× 呼吸次数三层(<12/12–18/>18)× 组别二分,切出 18 格。每格算两套算法的 Kendall τ 与翻转对数,再摊开 18 格的 τ 分布看翻转是否集中。

    为什么按 μ 和 n 切而不是只比整体?因为 1−R 的期望值随 n 变化、须重归一化(PMID 37756609),偏差的方向与量级都随 μ 和 n 变化——只看整体 τ,一个「整体很一致、局部全翻转」的问题会被完美隐藏。

    三档判据:把担忧升级为处置规则

    第一档:整体 τ > 0.9,但分层后出现 ≥1 对组次序翻转,且全部集中在 μ < 15° 或 n < 12 层。 这是最优结局——偏差只在同步最好、呼吸最慢的人群中生效,既往文献只需对这两个亚组重算,其余保留,同时验证了「偏差经跨零进入」的机制。

    第二档:整体 τ < 0.8。 这是最坏结局——两套算法排不出同一套次序,测的已接近两个不同构念。处置:所有按算术 SD 报的组间结论一律不可用,须全部按圆形口径重算。

    第三档:τ > 0.95 且各层均无翻转。 这是平安结局——偏差存在但不足以改排名,担忧降级为「报告规范问题」。同类先例:RIP 与光电体积描记测 TAA 高度一致(Bland-Altman 在界内,PMID 28767680)。

    别忘了噪声底:翻的是算法还是翻的是人

    φ 的个体内变异横跨 11%–123%(PMID 24716708),且主要来自生理不稳而非测量方法。这意味着任何「排名翻转」的解读都必须先与噪声底对比——若某格子的双算法排名差本就在个体内 CV 范围内,翻转只是噪声、不是算法偏差;只有超过个体内 CV 的置信区间,才能把翻判定给算法口径。合成数据没有「同一人隔天再测」的抖动,真实数据有——这正是判断「翻的是算法还是翻的是人」的试金石。

    证据与局限

    已验证:两套 φ 算法在合成数据上存在方向可预测的跨零偏差;六种 φ 测法间可差 30° 以上、交叉相关法误差 <5°(PMID 12422344);真实人群亚组间 φ 差异可大于算法偏差(N3 与 REM 差 2.4 倍,PMID 29991414);RIP 与 OEP 测 TAA 高度一致(PMID 28767680)。

    部分验证:1−R 期望值随 n 变化须重归一化(PMID 37756609);φ 个体内 CV 11%–123% 源于生理不稳(PMID 24716708)——共同提示「翻转须先排除噪声底」,但尚无人做过真实 RIP 库双算法排名检验。

    待验证:真实 RIP 的双算法 Kendall τ 落在哪一档;翻转是否集中在 μ < 15° 或 n < 12 亚组;处置规则是否可执行。

    相关阅读

    证据等级: C(传统文化共识/初步探索——方法学二次分析提案,尚无真实数据回算结果)
    参考文献: PMID 12422344, PMID 29991414, PMID 37756609, PMID 24716708, PMID 28767680

    L1 网络状态:该主题(呼吸相位离散度的「算法口径 → 排名一致性」)尚未进入 L1 对应网络——CORRESPONDENCE_NETWORK 节点表仅有「相位 / 节律」「呼吸率 / 呼吸模式」两个节点,无任何以「算法口径 / 排名一致性 / 亚组翻转」为源或目标的边;已登记 science-yixue-l1-gaps 第 71 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。

    Real-world test of rank reversal: feeding breathing data from ≥90 people to two algorithms to see if the group ordering actually flips

    Evidence level: C

    Issues identified in synthetic data must be validated against real-world data: the dispersion of thoraco-abdominal phase angles in the same cohort, ranked separately by each algorithm, determines whether—and for which subgroups—rank reversals occur.

    Why Synthetic Criteria Must Be Back-Calculated Against Real Data

    The previous article used synthetic data to calibrate the direction of algorithmic bias across zero-crossings: φ is an angular measure, and values approaching 0° are more prone to crossing zero, which amplifies the arithmetic standard deviation and depresses the expected value of circular dispersion (1−R). However, synthetic data can only demonstrate that “bias may exist,” not that “bias is sufficient to alter rankings in real populations.” To date, no empirical measurement has confirmed whether real data actually reverses order.

    This gap is precisely what this article addresses. Three possible outcomes correspond to three distinct courses of action: if overall ordering remains highly consistent (Kendall τ > 0.95), concerns are merely “reporting convention issues,” and prior conclusions based on arithmetic SD can be retained as-is; if overall ordering is inconsistent (τ < 0.8), all inter-group conclusions based on arithmetic SD must be recalculated; if overall ordering is consistent but local reversals occur after stratification, only the reversed subgroups need recalculation. Currently, no real numerical evidence indicates which path to follow.

    Two Literature Boundaries: Large Bias May Only Affect Specific Subgroups

    Regarding algorithmic bias in φ, literature provides two boundaries. At the algorithm level: results from six time-domain methods applied to the same φ dataset can differ by over 30°, whereas cross-correlation and maximum linear correlation reduce error to under 5° (PMID 12422344)—algorithm selection alone can consume the entire effect size. At the population level: the thoraco-abdominal asynchrony angle differs by 2.4× between N3 sleep and REM stages (PMID 29991414); real-world subgroup differences often far exceed algorithmic bias—provided neither algorithm reverses the already large inter-group gap, inter-group conclusions remain stable.

    These two boundaries yield testable predictions: bias concentrates in individuals with the best synchrony (μ < 15°) and slowest breathing (n < 12)—these groups are most prone to zero-crossing and most susceptible to arithmetic SD amplification. Real back-calculation tests this prediction: if reversal occurs only in these two subgroups, prior literature requires local recalculation; if reversal occurs universally, it represents a true systemic issue.

    One Pure Secondary Analysis: Testing All 18 Grids Sequentially

    Methodologically, no new participants are recruited: we use an existing adult resting RIP database (same posture, ≥10 minutes, with breath count), totaling ≥90 individuals covering both practitioners and novices. For each individual’s stable segment, the arithmetic definition uses the standard deviation of successive φ values, while the circular definition uses 1−R/圆方差/von Mises κ之一(fixed to a single definition). Stratification occurs across three levels of φ mean μ (0°–15° / 15°–30° / >30°) × three levels of breath count n (<12 / 12–18 / >18) × two groups, yielding 18 cells. For each cell, we compute Kendall τ and reversal count for both algorithms, then examine the τ distribution across all 18 cells to assess whether reversals cluster.

    Why stratify by μ and n rather than comparing overall? Because the expected value of 1−R varies with n and requires re-normalization (PMID 37756609); both the direction and magnitude of bias depend on μ and n—relying solely on overall τ would perfectly mask a scenario where “overall consistency hides universal reversal.”

    Three-Tier Criteria: Upgrading Concerns into Action Rules

    Tier 1: Overall τ > 0.9, but ≥1 pair of group orders reverses after stratification, all concentrated in μ < 15° or n < 12 layers. This is the optimal outcome—bias affects only those with best synchrony and slowest breathing; prior literature requires recalculation only for these two subgroups, while others remain valid, simultaneously confirming the mechanism that “bias enters via zero-crossing.”

    Tier 2: Overall τ < 0.8. This is the worst outcome—neither algorithm produces the same ranking, indicating they measure nearly different constructs. Action: all inter-group conclusions based on arithmetic SD are unusable; everything must be recalculated using the circular definition.

    Tier 3: τ > 0.95 with no reversals in any layer. This is the peaceful outcome—bias exists but is insufficient to alter rankings; concerns downgrade to “reporting convention issues.” Precedent: RIP and OEP show high agreement in measuring TAA (Bland-Altman limits of agreement, PMID 28767680).

    Don’t Forget the Noise Floor: Is the Reversal Algorithmic or Individual?

    Intra-individual variation of φ spans 11%–123% (PMID 24716708), primarily due to physiological instability rather than measurement error. This means any interpretation of “rank reversal” must first be compared against the noise floor—if the rank difference between two algorithms for a given cell falls within the intra-individual CV range, the reversal is merely noise, not algorithmic bias; only if it exceeds the intra-individual CV confidence interval can we attribute reversal to the algorithmic definition. Synthetic data lacks the day-to-day jitter of real measurements—real data has it, making this the gold standard for distinguishing “algorithm reversal” from “individual reversal.”

    Evidence and Limitations

    Verified: Two φ algorithms exhibit predictable zero-crossing bias in synthetic data; six φ measurement methods can differ by >30° while cross-correlation error <5° (PMID 12422344); real-world subgroup φ differences exceed algorithmic bias (N3 vs REM differ by 2.4×, PMID 29991414); RIP and OEP agree highly on TAA measurement (PMID 28767680).

    Partially Verified: Expected value of 1−R requires re-normalization as n changes (PMID 37756609); intra-individual CV of φ ranges 11%–123% due to physiological instability (PMID 24716708)—both suggest “reversal must first exclude noise floor,” yet no one has performed dual-algorithm ranking verification on real RIP databases.

    Pending Verification: Which tier does the real RIP dual-algorithm Kendall τ fall into; whether reversals concentrate in μ < 15° or n < 12 subgroups; whether action rules are executable.

    Related Reading

    Evidence Level: C (traditional consensus / preliminary exploration—methodological secondary analysis proposal; no real back-calculation results yet)
    References: PMID 12422344, PMID 29991414, PMID 37756609, PMID 24716708, PMID 28767680

    L1 Network Status: This topic (respiratory phase dispersion’s “algorithm definition → ranking consistency”) has not yet entered the L1 correspondence network—the CORRESPONDENCE_NETWORK node table contains only “phase / rhythm” and “breath rate / breath pattern” nodes, with no edges sourced or targeted by “algorithm definition / ranking consistency / subgroup reversal”; it has been registered as item 71 in science-yixue-l1-gaps.


    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.

  • 谁的尺子歪了,这个问题数学上没有答案:一个恒亏一秩的模型,加多少判读者都补不上

    Evidence level: C

    上一篇证明训练只练尺度、不练零点。这一篇往更前提退一步:「零点歪了多少」这个量,在只有打分、没有刻度的条件下,根本不是估不准,而是估不出来。但同一份数据里,两人零点之差是能估的——这两件事常被混为一谈,代价是把一整类研究判成不可做。

    正文

    上一篇(#1402)的结论是:两位判读者的差异有两个自由度,尺度 a 和位置 b,而训练材料只压得住尺度、压不住零点。

    顺着往下,会撞上一个更基础的问题:那我们把 b 测出来不就行了?

    答案分两半,而且这两半方向相反。b 的绝对值测不出来,b 的差测得很准。 现有方法学讨论经常把这两半当成一件事,于是要么过度悲观(认为整条偏倚归因不可做),要么过度乐观(认为多加几位读者就解决了)。

    一、加读者不降秩:这是恒亏一秩,不是样本量问题

    读者模型是 s = a·x + b + ε。只看均值层面,R 位读者给出 R 个可观测的均值,未知量是 R+1 个:这批人的真实均值 μ,加上 R 个读者各自的零点 b₁…b_R。

    R 位读者 → R 个观测,R+1 个未知
    设计矩阵秩 = R,零空间维数 = 1(对任意 R)
    

    实测 R 从 2 加到 8,亏的秩一直是 1,一点也不缩小:

    设计 观测数 未知数 设计矩阵秩 亏秩
    2 位读者 2 3 2 1
    3 位读者 3 4 3 1
    5 位读者 5 6 5 1
    8 位读者 8 9 8 1
    2 位读者 + 1 位金标准判读者 3 4 3 1
    4 位读者 + 2 位金标准判读者 6 7 6 1

    最后一行的意思是:再加一位「资深专家」当金标准也没用。资深专家也是一个读者,也带一个任意零点,他只是把 R 变成 R+1,亏的秩还是 1。这一步很容易被想岔——文献里常见的做法是「找一位权威来定标准」,而权威的标准答案本身也是从某个未标定的零点出发的。

    要真正降秩,必须有一个有刻度的外部锚:不是「某位专家怎么判」,而是一个自带单位的物理量(比如皮温变化的摄氏度数、握力的牛顿值)。锚一进来,μ 被钉住,R+1 个未知里少一个,秩才补齐。

    二、三个「基准」给出的不是三个答案,是同一个答案的三种写法

    上一轮预设的判据是:三种基准算出的 τ̂ 若符号相反或幅度差超过 0.15 个标准差,就判「不可识别」。

    实测发现这个判据测的不是识别性。同一份数据(真值 b₁=0.30、b₂=0.05),三种基准拟合出的残差平方和逐位相同,差值是 0.0:

    基准假设 b̂₁ b̂₂ τ̂ = b̂₂−b̂₁ 残差平方和
    认定 μ=0.00 +0.300 +0.050 −0.2500 330.806977
    认定 μ=0.35 −0.050 −0.300 −0.2500 330.806977
    认定 μ=−0.20 +0.500 +0.250 −0.2500 330.806977

    三行的 τ̂ 一模一样。所谓「符号相反」,是读者一为基准与读者二为基准的差别(τ₁₂ = −τ₂₁),这是定义的方向在这个翻转,不是估计值在打架;所谓「幅度差 50%」,是「均值去中心」把差取了一半(报的是半差),也是定义。

    所以这三个数不是三个互相矛盾的估计,是同一个可识别量在三种约定下的三种写法。危险不在数学,在报告环节:同一份数据可以报出 +0.25、−0.25、−0.125 三个数,读者无从判断哪个是「真的」。规范该管的是「必须声明 τ 是谁减谁、有没有除 2」,而不是「τ 不可报」。

    三、错的是绝对值,对的永远是差

    绝对值那一侧确实彻底不可识别,而且无界。把 μ 分别当成 −5、0、+5,观测到的读者一均值是 0.285 时,b̂₁ 依次是 +5.285、+0.285、−4.715——μ 挪多少,b̂₁ 就朝反方向挪多少,一个都不少。

    真实研究里 μ 没人知道,习惯做法是默认 μ=0(把样本均值当零点)。这个默认的代价可以量出来:真 μ=0.35、真 b₁=0.30、真 b₂=0.05 时,按 μ=0 解释得到 b̂₁=0.647(高估 116%)、b̂₂=0.395(高估 690%)。绝对偏移量错得一塌糊涂。

    但 δ̂ = b̂₂ − b̂₁ = −0.252,真值 −0.25,n=120 与 n=480 都一样准。这就是本篇最实用的一句:偏倚的绝对值不可答,偏倚之差可答,而报告和研究需要的恰恰是后者。

    四、n=120 的区间半宽刚好等于一个整 5 个百分点

    差可估,但要算多准。真 τ = 0.25 个标准差时,τ̂ 的抽样精度(800 次重抽):

    n SD(τ̂) 95% 区间半宽 折成「5pp 当量」的倍数
    60 0.089 0.175 1.38
    120 0.066 0.128 1.02
    240 0.045 0.089 0.71
    480 0.033 0.065 0.51
    960 0.023 0.045 0.36

    「5pp 当量」指位置差达到 0.126 个总标准差时,边际差正好 5 个百分点(实测 5.00 pp)。上一轮设计的 ≥120 例,其 95% 区间半宽是 0.128——整整一个 5pp 当量。也就是说 n=120 时,「这个位置差超过 5pp 了吗」这个问题,区间的下半段说没超、上半段说超了,判决力接近于零。要到 n=240 才降到 0.71 个当量、n=480 才 0.51。结论:要判 5pp 级的位置差,n≥240 是底线,480 才算稳。

    顺带一个坑:那个 0.126 的门槛没有声明分母是哪个 σ。用总标准差算,b=0.141 给 5.00 pp;用噪声标准差算,b=0.063 只给 2.26 pp。差 2.2 倍,全看分母怎么取——凡引用这个门槛的地方,必须写明「相对总离散度」还是「相对测量噪声」。

    五、不可识别不等于无所谓:校错方向比不校更糟

    「谁偏了」答不出来,很容易被理解成「那就算了,别管」。这个理解是错的。三个动作的实测后果(真差 b₁=0.141、b₂=0):

    • 不校正、只如实报告两人不一致:δ = 5.02 pp
    • 校正对方向(把读者一减去 0.141):δ = 0.01 pp
    • 校正错方向(把读者二再加上 0.141):δ = 10.01 pp

    校错方向把偏倚放大到两倍,比什么都不做还糟一倍。而可识别性告诉我们:在没有外部锚的情况下,「往哪个方向校」这个选择没有任何数据依据——它只能靠约定(比如「一律以资深者为准」)来定。这就把风险从「可能偏」变成了「有一半概率把偏倚翻倍」。

    这才是可识别性缺失的真正代价:它不是让研究变得不可做,而是让任何主动校正动作都带一个 50% 的翻倍风险。对应的规范应当是:无外部锚时,只报差值、不报归属、不做方向性校正。

    证据与局限

    已验证: 秩亏恒为 1,与读者数无关(R=2…8 及含金标准的 3、6 观测设计,设计矩阵秩恒等于观测数,零空间维数恒为 1);三基准下残差平方和精确相等(差 0.0)、τ̂ 数值恒定;b̂ 绝对值随 μ 任意平移而无界(μ 取 −5 至 +5,b̂₁ 从 +5.285 到 −4.715);δ̂ = b̂₂−b̂₁ 在 μ 误设下基本无偏(−0.252 对真值 −0.25);τ̂ 的 95% 区间半宽 n=60/120/240/480/960 分别为 0.175/0.128/0.089/0.065/0.045 个标准差;校错方向使 δ 从 5.02 pp 变为 10.01 pp。

    部分验证: 「加外部锚可降秩」是线性代数的直接推论,但真实研究中「皮温变化幅度」这类锚本身也带测量误差,锚不精确时秩是补上了、估计量的方差却会变大,这个折中本文没有算。0.126 这个 5pp 当量门槛来自上一篇的数值解,本文只复核了它在两种 σ 定义下的取值,未重新推导。

    待验证: ① 三位以上读者 + 外部锚的完整设计中,锚的误差如何传播到 τ̂ 的方差,从未算过;② 真实判读数据里「一律以资深者为准」这条约定造成的方向错误率是多少,没有实测(本文只算了「错一次」的后果,没算「多久错一次」);③ 当 μ 本身随批次变化(不同批受试者人群构成不同)时,跨批次合并 τ̂ 是否仍然无偏,未检验。

    相关阅读

    证据等级: C
    参考文献: PMID 2868172, PMID 8501467, PMID 22129533, PMID 39854081, PMID 31444807
    L1 网络状态:CORRESPONDENCE_NETWORK.md 全表 grep「判读 / kappa / 一致性 / 信度 / 评分 / 读者 / 校准 / 偏倚 / 金标准 / 参照 / 基准 / 可识别 / 归属 / 真值」仅命中第 192 行「生克一致性 | 相生/相克方向能否复现 | 未测」(指五行生克,与判读信度无关);物理侧最近位为第 57 行「相位 / 节律 | — | 周期性耦合的关键」,语义限于两振荡锁相;生理/体验侧九项节点无「判读信度」「位置偏倚」「可识别性」任何节点;边表止于 E-050「六十甲子纳音 | 对应 | 60 个频率」(L1 / open,现代证据栏原文「尚无任何可测来源」);网络对齐指标行原文「L2 以上边数 | 0」。该主题尚未进入 L1 对应网络,已登记至 science-yixue-l1-gaps.md(「测量一致性」簇第 16 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。

    Whose ruler is crooked, a question with no mathematical answer: adding any number of classifiers to a model that is one rank deficient cannot compensate

    Evidence level: C

    The previous proof showed that training only calibrates the scale, not the zero point. This paper steps back further: the magnitude of “how much the zero is bent” cannot be estimated at all under conditions where only scores are given, not scales. Yet within the same dataset, the difference between two zero points can be estimated. These two facts are often conflated, with the cost of declaring an entire class of studies infeasible.

    Body

    The conclusion of the previous paper (#1402) is that the difference between two raters involves two degrees of freedom: scale $a$ and position $b$, while the training material constrains only the scale, not the zero point.

    Following this, we encounter a more fundamental question: Can’t we just measure $b$?

    The answer splits into two halves, which point in opposite directions. The absolute value of $b$ cannot be measured, but the difference between $b$ values can be measured very accurately. Current methodological discussions often treat these two halves as one thing, leading either to excessive pessimism (that entire classes of bias attribution are infeasible) or excessive optimism (that adding more raters solves the problem).

    1. Adding Raters Does Not Reduce Rank: This Is a Permanent Rank Deficiency of One, Not a Sample Size Issue

    The rater model is $s = a \cdot x + b + \varepsilon$. Looking only at the mean level, $R$ raters provide $R$ observable means. The unknown quantities are $R+1$: the true mean $\mu$ of these people, plus the $R$ individual zero points $b_1 \dots b_R$.

    R raters → R observations, R+1 unknowns
    Design matrix rank = R, null space dimension = 1 (for any R)
    

    Empirically, increasing $R$ from 2 to 8 leaves the missing rank always at 1, never shrinking:

    Design Observations Unknowns Design Matrix Rank Deficient Rank
    2 raters 2 3 2 1
    3 raters 3 4 3 1
    5 raters 5 6 5 1
    8 raters 8 9 8 1
    2 raters + 1 gold-standard rater 3 4 3 1
    4 raters + 2 gold-standard raters 6 7 6 1

    The last row means: adding another “senior expert” as a gold standard is useless. A senior expert is also just a rater, carrying an arbitrary zero point; they merely increase $R$ to $R+1$, leaving the deficient rank still at 1. This step is easily misunderstood—common literature practice is “find an authority to set the standard,” yet the authority’s standard answer itself originates from some unstabilized zero point.

    To truly reduce the rank, we need an external anchor with units: not “how a certain expert judges,” but a physical quantity carrying units (e.g., skin temperature change in degrees Celsius, grip strength in Newtons). Once this anchor enters, $\mu$ is pinned down, reducing the $R+1$ unknowns by one, and the rank is completed.

    2. Three “Baselines” Do Not Give Three Answers, They Give Three Ways of Writing the Same Answer

    The previous assumption stated that if the estimated $\hat{\tau}$ from three baselines have opposite signs or a magnitude difference exceeding 0.15 standard deviations, it should be judged “unidentifiable.”

    Empirical testing shows this criterion does not measure identifiability. For the same dataset (true values $b_1=0.30, b_2=0.05$), the residual sum of squares fitted by three baselines is identical at each position, with a difference of 0.0:

    Baseline Assumption $\hat{b}_1$ $\hat{b}_2$ $\hat{\tau} = \hat{b}_2 – \hat{b}_1$ Residual Sum of Squares
    Assume $\mu=0.00$ +0.300 +0.050 −0.2500 330.806977
    Assume $\mu=0.35$ −0.050 −0.300 −0.2500 330.806977
    Assume $\mu=-0.20$ +0.500 +0.250 −0.2500 330.806977

    The $\hat{\tau}$ values are identical across all three rows. The so-called “opposite signs” refers to the difference between Rater 1 as baseline and Rater 2 as baseline ($\tau_{12} = -\tau_{21}$); this is a reversal of definition direction, not a conflict in estimates. The so-called “50% magnitude difference” refers to “mean-centering” halving the difference (reporting half-difference), which is also a matter of definition.

    Thus, these three numbers are not three contradictory estimates; they are three ways of writing the same identifiable quantity under different conventions. The danger lies not in mathematics, but in reporting: the same dataset can report +0.25, −0.25, and −0.125, leaving readers unable to judge which is “true.” Standards should govern “must declare who minus whom, whether divided by 2,” not “τ cannot be reported.”

    3. The Absolute Value Is Wrong, the Difference Is Always Right

    The absolute value side is indeed completely unidentifiable and unbounded. Setting $\mu$ to −5, 0, or +5 respectively, when the observed mean of Rater 1 is 0.285, yields $\hat{b}_1$ values of +5.285, +0.285, and −4.715 respectively—as $\mu$ shifts, $\hat{b}_1$ shifts in the opposite direction by exactly the same amount.

    In real research, $\mu$ is unknown; the habitual practice defaults to $\mu=0$ (treating the sample mean as the zero point). The cost of this default can be quantified: when true $\mu=0.35$, true $b_1=0.30$, and true $b_2=0.05$, interpreting with $\mu=0$ yields $\hat{b}_1=0.647$ (overestimated by 116%) and $\hat{b}_2=0.395$ (overestimated by 690%). The absolute offset is utterly wrong.

    However, $\hat{\delta} = \hat{b}_2 – \hat{b}_1 = -0.252$, matching the true value of −0.25 with $n=120$ and $n=480$ equally accurate. This is the most practical sentence of this paper: the absolute value of bias cannot be answered, but the difference of biases can be answered, and what reporting and research need is precisely the latter.

    4. The Half-Width of the Interval for $n=120$ Equals Exactly One Full 5 Percentage Points

    The difference can be estimated, but we must calculate how accurately. When the true $\tau = 0.25$ standard deviations, the sampling precision of $\hat{\tau}$ (800 resamples):

    n SD($\hat{\tau}$) 95% Interval Half-Width Multiplier of “5pp Equivalent”
    60 0.089 0.175 1.38
    120 0.066 0.128 1.02
    240 0.045 0.089 0.71
    480 0.033 0.065 0.51
    960 0.023 0.045 0.36

    “5pp equivalent” refers to when the positional difference reaches 0.126 of the total standard deviation, the marginal difference is exactly 5 percentage points (measured as 5.00 pp). The previous design’s $\ge 120$ cases have a 95% interval half-width of 0.128—exactly one 5pp equivalent. That means at $n=120$, the question “does this positional difference exceed 5pp?” has the lower half of the interval saying no, and the upper half saying yes, making the decision power nearly zero. Only at $n=240$ does it drop to 0.71 equivalents, and at $n=480$ to 0.51. Conclusion: to judge positional differences at the 5pp level, $n \ge 240$ is the baseline, and $n=480$ is considered stable.

    A side pitfall: the 0.126 threshold did not declare which $\sigma$ is in the denominator. Using total standard deviation, $b=0.141$ gives 5.00 pp; using noise standard deviation, $b=0.063$ gives only 2.26 pp. A difference of 2.2 times, entirely depending on how the denominator is chosen—any reference to this threshold must specify “relative to total dispersion” or “relative to measurement noise.”

    5. Unidentifiable Does Not Mean Irrelevant: Correcting in the Wrong Direction Is Worse Than Not Correcting

    That “who is biased” cannot be answered is easily understood as “then forget it, don’t worry.” This understanding is wrong. The empirical consequences of three actions (true difference $b_1=0.141, b_2=0$):

    • No correction, only reporting the inconsistency between two raters: $\delta = \mathbf{5.02\ pp}$
    • Correcting in the right direction (subtracting 0.141 from Rater 1): $\delta = \mathbf{0.01\ pp}$
    • Correcting in the wrong direction (adding 0.141 to Rater 2): $\delta = \mathbf{10.01\ pp}$

    Correcting in the wrong direction doubles the bias, making it twice as bad as doing nothing. Identifiability tells us: without an external anchor, the choice of “which direction to correct” has no data basis—it can only rely on convention (e.g., “always use the senior expert”). This turns the risk from “possibly biased” into “50% probability of doubling the bias.”

    This is the true cost of missing identifiability: it does not make research infeasible, but makes any active correction carry a 50% doubling risk. The corresponding standard should be: without an external anchor, report only differences, do not report attribution, and do not perform directional correction.

    Evidence and Limitations

    Verified: Rank deficiency is permanently 1, independent of the number of raters ($R=2 \dots 8$ and gold-standard designs with 3, 6 observations; the design matrix rank always equals the number of observations, and the null space dimension is always 1); residual sum of squares under three baselines are exactly equal (difference 0.0), and $\hat{\tau}$ values remain constant; absolute values of $\hat{b}$ shift unboundedly as $\mu$ arbitrarily translates (with $\mu$ ranging from −5 to +5, $\hat{b}_1$ ranges from +5.285 to −4.715); $\hat{\delta} = \hat{b}_2 – \hat{b}_1$ is essentially unbiased under mis-specification of $\mu$ (−0.252 vs true −0.25); 95% interval half-widths for $\hat{\tau}$ at $n=60/120/240/480/960$ are 0.175/0.128/0.089/0.065/0.045 standard deviations respectively; correcting in the wrong direction changes $\delta$ from 5.02 pp to 10.01 pp.

    Partially Verified: “Adding an external anchor reduces rank” is a direct consequence of linear algebra, but in real research, anchors like “skin temperature change amplitude” themselves carry measurement error; when the anchor is imprecise, the rank is filled but the variance of the estimator increases, a trade-off not calculated in this paper. The 0.126 threshold for the 5pp equivalent comes from numerical solutions in the previous paper; this paper only verified its values under two $\sigma$ definitions, without re-deriving it.

    To Be Verified: ① In complete designs with three or more raters plus an external anchor, how the anchor’s error propagates to the variance of $\hat{\tau}$ has never been calculated; ② In real rating data, what is the rate of directional error caused by the convention “always use the senior expert,” has not been empirically tested (this paper only calculated the consequence of “one wrong correction,” not “how often it goes wrong”); ③ When $\mu$ itself varies across batches (different batches have different subject population compositions), whether merging $\hat{\tau}$ across batches remains unbiased has not been tested.

    Related Reading

    Evidence Level: C
    References: PMID 2868172, PMID 8501467, PMID 22129533, PMID 39854081, PMID 31444807
    L1 Network Status: CORRESPONDENCE_NETWORK.md full table grep for “rater / kappa / consistency / reliability / scoring / reader / calibration / bias / gold standard / reference / baseline / identifiable / attribution / true value” only hits row 192 “Sheng Ke Consistency | Can Sheng/Xiang directions be reproduced | Not tested” (referring to Five Elements Sheng/King, unrelated to rater reliability); the physical side’s latest position is row 57 “Phase / Rhythm | — | Key of periodic coupling,” with semantics limited to two oscillators phase-locked; physiological/experience side has nine nodes but none for “rater reliability,” “positional bias,” or “identifiability”; the edge table stops at E-050 “Sixty Jiazi Yin-Yin | Correspondence | 60 frequencies” (L1 / open, original text in modern evidence column states “no measurable source yet”); network alignment indicator row originally “L2 and above edges | 0”. This topic has not yet entered the L1 corresponding network, and is registered in science-yixue-l1-gaps.md (cluster “measurement consistency,” item 16).


    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.

  • 判读者培训一直在练尺子,没人练零点:一个把一致性练到 0.961 的训练方案,让偏倚从 7.93 涨到 8.85

    Evidence level: C

    上一篇证明「标准差之比 κ_s」对位置漂移是精确盲的。这一篇问一个更实际的问题:既然真实研究都在做判读者校准训练,那训练到底在校准哪个自由度?答案不好听——常规的训练材料只教尺度、不教零点,而位置恰恰是偏倚的主导项。

    正文

    上一轮 (#1392) 的结论是:两位判读者之间的差异只有两个自由度,尺度 a(尺子多长)和位置 b(零点在哪),而 κ_s = a₁/a₂ 只读第一个,对第二个精确盲——b 挪到 0.14 个标准差,δ 就到 5 个百分点,κ_s 仍是 1.000。

    顺着这条线往下,一个必须回答的现实问题是:现行的「统一评分标准」训练,究竟在校准哪个自由度?

    一、训练材料的两种设计,压的是两个不同的东西

    论文里最常见的做法是给判读者一批带标准答案的样例,看完再独立打分。但「一批样例」里装什么,结果完全不同:

    • 两端样例:只给明显改善和明显恶化的病例。这教会读者「多大的变化算大变化」,压的是尺度。
    • 边界样例:专门给「轻微改善 / 无变化 / 轻微加重」这种临界段。这压的是零点。

    现行为什么偏好两端?因为两端样例的标准答案好定,专家之间不会吵;边界样例的标准答案本身就难产。于是训练资源天然倾斜到尺度那一侧——而尺度恰好是 κ_s 会读、δ 不承重的那一侧。

    二、四臂数值模拟:原判据的方向整个反了

    本文做了一次四臂模拟(读者模型 s = a·x + b + ε,x~N(0,1),ε 为读者个体噪声;每组 20 万次抽样的解析解,无随机种子依赖)。基线设定读者 A 全程不动,读者 B 初始 a=1.20、b=0.26——即尺子偏长、零点也偏了:

    训练臂 κ_s δ(边际差) 相对无训练
    无训练 0.860 7.93 pp —
    只给两端样例(a: 1.20→1.05) 0.961(合格) 8.85 pp δ 恶化 +0.92 pp
    只给边界样例(b: 0.26→0.08) 0.828(不合格) 2.36 pp δ 改善 −5.56 pp
    两端+边界联合 0.917 2.62 pp δ 改善 −5.31 pp

    第一行就是坏消息。两端样例把 κ_s 从 0.860 拉到 0.961,跨过了 0.9–1.1 的合格线;与此同时 δ 从 7.93 涨到 8.85。 为什么?因为收窄尺度会同时压低 B 读者在判据线以上的概率——尺度修正是有效的,但它在边际上制造了一个与位置漂移同向的额外偏差。也就是说,κ_s 的改善与 δ 的恶化,可以是同一个动作的两个后果。

    第二行是更尖锐的反例:唯一真正把 δ 压下去的那个臂,κ_s 反而更差(0.828,被判不合格)。 原判据(「两端臂压 κ_s 但 δ 降 <1pp ⇒ 尺度锚定无效」)预设了 δ 只会”不降”,没设想 κ_s 变好时 δ 会变坏。

    三、合格区根本不是安全区:它连一个坏样本都筛不掉

    把参数空间(a_B∈[0.7,1.45]、b_B∈[−0.55,0.55]、读者噪声 0.3–0.8)扫 40 万点,看「κ_s 合格」这条线到底筛掉了什么:

    κ_s 区间 占样本 其中 δ>5pp 占比 全体 δ>5pp 占比
    [0.85, 1.15] 52.4% 75.2% 74.4%
    [0.9, 1.1] 35.1% 75.6% 74.4%
    [0.95, 1.05] 17.6% 75.9% 74.4%
    [0.99, 1.01] 3.6% 75.5% 74.4%

    把合格区从 ±15% 收到 ±1%,坏样本占比纹丝不动(74→75%,甚至略升),只是样本量从 52% 掉到 3.6%。 这不是筛选力弱,是筛选力为零——κ_s 与 δ 在参数空间里近乎正交,卡 κ_s 等于随机删数据。

    四、为什么两个自由度不对称:一个有天花板,一个没有

    这不是运气,是几何。尺度错单独存在时,它能造出的 δ 有天花板:a_B=1.05 时最多 0.95 pp,a_B=1.20 时 3.64 pp,即使极端到 a_B=1.45 也只有 8.37 pp;而且方向随切点翻转——切点选在两人中位数处(c=0),δ 精确等于 0.00 pp,切点挪到 c=±1.5,δ 才露出 3.4 pp。

    位置错单独存在时,没有天花板:b=0.3 是 10.6 pp,b=0.6 是 20.4 pp,b=1.5 是 41.0 pp,而 κ_s 全程恒等于 1.000。

    所以「κ_s 合格 ⇒ δ 可控」这句话的真实含义是:你在管一个有上限的次要项,放弃了一个无上限的主导项。这不是精度问题,是对象选错了。

    五、两个阈值都掉在噪声底下

    原判据用了两条线:δ̂ 降幅 <1 pp 判”无效”,>3 pp 判”有效”。在 n=120 例/臂、读者不一致率 0.25 的常见设定下,单臂 SE(δ̂)=4.51 pp,两臂差 SE=6.37 pp:

    • 1 pp 那条线:即使真的一点改善都没有,Δδ̂ 也有 87.5% 的概率超过 1 pp。也就是说「无变化」这个正确结论,只有 12.5% 的机会被说出来。
    • 3 pp 那条线:真无改善时,31.9% 的概率 Δδ̂>3 pp——三成”有效”是把噪声当疗效。
    • 要让 3 pp 达到 80% 功效,需要 4244 例/臂。原设计 120 例,功效 7.6%。

    两条线本身也不可分:SE 6.37 pp 的量级下,1 pp 与 3 pp 只隔 0.31 个 SE。

    六、第四条判据(相关系数 <0.3)在四个臂下无法判决

    原设计想用「Δδ 降幅与 Δκ_s 降幅的相关系数 <0.3」来判 κ_s 能不能当验收指标。四臂只有 3 个自由度,相关系数的 95% 置信区间在 atanh 尺度上宽达 ±1.96:真值 0.00 覆盖 (−0.96, +0.96),真值 0.90 也覆盖 (−0.45, +1.00)。任何真值都覆盖 0,这条判据永远判”不显著”,永远成立——包括在 κ_s 完全有效的世界里。 要 80% 功效检出 ρ=0.5,需要 30 个臂。

    证据与局限

    已验证(本文数值模拟,解析可复现):κ_s 与 δ 在参数空间近乎正交;尺度错有天花板(本参数域内最大 8.37 pp)、位置错无天花板(同域内 41.0 pp);κ_s 合格区对 δ>5pp 无筛选力;n=120 例/臂下 1 pp 与 3 pp 两条判据线均低于噪声底。

    部分验证:真实判读者训练确实能提高一致性,但方向集中于经验不足者——内镜萎缩性胃炎分级训练后 κ 从 0.18/0.32 双升并维持(PMID 29629520);BI-RADS 专门训练后新读者 κ 0.41→0.50(p=0.01),而有经验读者 0.48→0.46 无变化(PMID 21899969)。这说明训练主要修的是”校准不足”,不是”零点分歧”——与本文”两端样例只压尺度”的推断方向一致,但这些研究都没有同时报边际偏倚,无法直接验证 δ 是否也降。

    待验证:① 真实训练研究里 κ 改善与 δ 改善的相关性从未被并排报告——目前只能引用 Byrt 等(PMID 8501467)的立场:κ 同时受偏倚与流行度污染,只报 κ 会误导,必须并报偏倚与流行度指标;② 乳腺密度跨中心研究(PMID 37127220)用 STUART-MAXWELL 边际同质性检验发现四档分布显著不同(P<0.05),而加权一致性仍达 0.86——这是”一致性很高、边际仍分家”的现成实例,但它没有做训练干预,无法回答”训练能否修边际”;③ 本文四臂是模拟,真实读者在”只给两端样例”条件下 δ 是否真的上升,需要 ≥120 例 × ≥2 读者 × 四臂的实测;④ 若改用边界样例,δ̂ 的最佳可达精度受 n=120 限制在 ±8.8 pp,要看清 3 pp 需要 ≥4244 例/臂。

    方法学建议(可执行):任何以”统一评分标准”提高一致性的研究,须并报 δ̂ 及其 CI,并注明训练材料是否含边界样例;报”训练有效”须先说明 δ̂ 的可检出下限,低于该下限的改善不得称为有效。

    相关阅读

    证据等级: C
    参考文献: PMID 8501467, PMID 29629520, PMID 21899969, PMID 37127220
    L1 网络状态:CORRESPONDENCE_NETWORK.md 全表 grep「判读 / kappa / 一致性 / 信度 / 评分 / 读者 / 校准 / 偏倚」仅命中第 192 行「生克一致性 | 相生/相克方向能否复现 | 未测」(指五行生克,非判读信度);物理侧最近位为第 57 行「相位 / 节律 | — | 周期性耦合的关键」,生理/体验侧九项节点(HRV / 皮温皮电 / 呼吸率 / 肌电 / EEG / 胃肠蠕动 / 气感报告 / 体质分型)无「判读信度」「边际偏倚」任何节点;边表止于 E-050「六十甲子纳音 | 对应 | 60 个频率」(L1 / open,现代证据栏原文「尚无任何可测来源」);网络对齐指标行原文「L2 以上边数 | 0」。该主题尚未进入 L1 对应网络,已登记至 science-yixue-l1-gaps.md(「测量一致性」簇第 15 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。

    Judging Reader Training Always Practices Rulers, No One Practices Zero: A Training Scheme That Consistency to 0.961, Letting Bias Rise from 7.93 to 8.85

    Evidence level: C

    The previous article demonstrated that the ratio of standard deviations, κ_s, is precisely blind to position drift. This article addresses a more practical question: since real-world studies conduct reader calibration training, what degree of freedom is actually being calibrated? The answer is unpalatable—conventional training materials teach the scale but not the zero, and position is precisely the dominant term of bias.

    Main Text

    The conclusion from the previous round (#1392) is that the difference between two readers involves only two degrees of freedom: the scale $a$ (how long the ruler is) and the position $b$ (where the zero point is). Since $\kappa_s = a_1/a_2$ reads only the first, it is precisely blind to the second—even if $b$ shifts by 0.14 standard deviations, $\delta$ increases by 5 percentage points, yet $\kappa_s$ remains 1.000.

    Following this line, a necessary practical question arises: which degree of freedom does the current “unified scoring standard” training actually calibrate?

    I. Two Designs of Training Materials Address Two Different Things

    The most common approach in papers is to provide readers with a batch of cases with standard answers, which they then independently score. However, what is contained in “a batch of cases” yields completely different results:

    • End-point cases: Only cases with obvious improvement or obvious deterioration are provided. This teaches readers “what magnitude of change counts as significant,” addressing the scale.
    • Boundary cases: Specifically, cases in the critical zone of “slight improvement / no change / slight worsening.” This addresses the zero point.

    Why does current practice prefer end-points? Because standard answers for end-point cases are easy to determine, and experts do not argue; boundary case standard answers are inherently difficult to generate. Thus, training resources naturally tilt toward the scale side—while the scale is precisely the term $\kappa_s$ reads, but $\delta$ does not carry.

    II. Four-Arm Numerical Simulation: The Original Direction of the Evidence is Entirely Reversed

    This article conducts a four-arm simulation (reader model $s = a \cdot x + b + \varepsilon$, where $x \sim N(0,1)$ and $\varepsilon$ is reader individual noise; each arm has 200,000 analytical solutions based on sampling, with no random seed dependency). The baseline sets Reader A stationary throughout, while Reader B starts with $a=1.20$ and $b=0.26$—that is, the ruler is too long and the zero point is also offset:

    Training Arm $\kappa_s$ $\delta$ (marginal difference) vs. No Training
    No training 0.860 7.93 pp —
    Only end-point cases ($a$: 1.20→1.05) 0.961 (qualified) 8.85 pp $\delta$ worsens +0.92 pp
    Only boundary cases ($b$: 0.26→0.08) 0.828 (unqualified) 2.36 pp $\delta$ improves −5.56 pp
    Combined end-point + boundary 0.917 2.62 pp $\delta$ improves −5.31 pp

    The first row is the bad news. End-point cases pull $\kappa_s$ from 0.860 to 0.961, crossing the 0.9–1.1 qualification threshold; simultaneously $\delta$ rises from 7.93 to 8.85. Why? Because narrowing the scale simultaneously suppresses Reader B’s probability above the decision line—the scale correction is effective, but it marginally creates an additional bias in the same direction as position drift. That is, the improvement of $\kappa_s$ and the worsening of $\delta$ can be two consequences of the same action.

    The second row is a sharper counterexample: the only arm that truly pushes $\delta$ down has an even worse $\kappa_s$ (0.828, judged unqualified). The original evidence (“the end-point arm suppresses $\kappa_s$ but $\delta$ drops <1 pp ⇒ scale anchoring is ineffective”) presupposes that $\delta$ will only “not drop,” failing to envision that $\delta$ could worsen when $\kappa_s$ improves.

    III. The Qualified Zone is Not a Safe Zone: It Cannot Filter Out Even One Bad Sample

    Scanning 400,000 points in parameter space ($a_B \in [0.7,1.45]$, $b_B \in [-0.55,0.55]$, reader noise 0.3–0.8) to see what “qualified $\kappa_s$” actually filters out:

    $\kappa_s$ interval Proportion of samples Among them, proportion with $\delta > 5\text{pp}$ Overall proportion with $\delta > 5\text{pp}$
    [0.85, 1.15] 52.4% 75.2% 74.4%
    [0.9, 1.1] 35.1% 75.6% 74.4%
    [0.95, 1.05] 17.6% 75.9% 74.4%
    [0.99, 1.01] 3.6% 75.5% 74.4%

    Reducing the qualified zone from ±15% to ±1% leaves the proportion of bad samples unchanged (74→75%, even slightly rising); only the sample size drops from 52% to 3.6%. This is not weak filtering power—it is zero filtering power. $\kappa_s$ and $\delta$ are nearly orthogonal in parameter space; constraining $\kappa_s$ is equivalent to randomly deleting data.

    IV. Why the Two Degrees of Freedom Are Asymmetric: One Has a Ceiling, the Other Does Not

    This is not luck; it is geometry. When scale error exists alone, the $\delta$ it can produce has a ceiling: at $a_B=1.05$ the maximum is 0.95 pp, at $a_B=1.20$ it is 3.64 pp, and even at the extreme $a_B=1.45$ it reaches only 8.37 pp; moreover, the direction flips with the cut point—if the cut point is chosen at the median of the two readers ($c=0$), $\delta$ is precisely 0.00 pp, but shifting the cut point to $c=\pm1.5$ reveals only 3.4 pp.

    When position error exists alone, there is no ceiling: $b=0.3$ yields 10.6 pp, $b=0.6$ yields 20.4 pp, $b=1.5$ yields 41.0 pp, while $\kappa_s$ remains constantly 1.000 throughout.

    Thus the statement “$\kappa_s$ qualified ⇒ $\delta$ controllable” truly means: you are managing a secondary term with an upper limit, while abandoning the dominant term without an upper limit. This is not a precision issue; it is a matter of selecting the wrong object.

    V. Both Thresholds Fall Below the Noise Floor

    The original evidence used two lines: $\hat{\delta}$ drop <1 pp deemed “ineffective,” >3 pp deemed “effective.” Under the common setting of $n=120$ cases/arm and reader disagreement rate 0.25, the single-arm SE($\hat{\delta}$) = 4.51 pp and two-arm difference SE = 6.37 pp:

    • The 1 pp line: Even if there is absolutely no improvement, $\Delta\hat{\delta}$ has 87.5% probability of exceeding 1 pp. That means the correct conclusion “no change” has only a 12.5% chance of being stated.
    • The 3 pp line: When there is truly no improvement, there is a 31.9% probability that $\Delta\hat{\delta} > 3$ pp—over a third of “effective” results are noise masquerading as efficacy.
    • To achieve 80% power for the 3 pp threshold, 4244 cases/arm are needed. The original design with 120 cases has only 7.6% power.

    The two lines themselves are inseparable: at the scale of SE = 6.37 pp, the gap between 1 pp and 3 pp is only 0.31 SE.

    VI. The Fourth Criterion (Correlation Coefficient <0.3) Cannot Decide Under Four Arms

    The original design intended to use “correlation coefficient between $\Delta\delta$ drop and $\Delta\kappa_s$ drop <0.3” to judge whether $\kappa_s$ can serve as an acceptance metric. With only three degrees of freedom across four arms, the 95% confidence interval for the correlation coefficient spans ±1.96 on the atanh scale: true value 0.00 covers (−0.96, +0.96), true value 0.90 also covers (−0.45, +1.00). Any true value covers 0, so this criterion always judges “not significant,” always holds—even in the world where $\kappa_s$ is completely effective. Detecting $\rho=0.5$ with 80% power would require 30 arms.

    Evidence and Limitations

    Verified (numerical simulation in this article, analytically reproducible): $\kappa_s$ and $\delta$ are nearly orthogonal in parameter space; scale error has a ceiling (maximum 8.37 pp within this parameter domain), position error has no ceiling (41.0 pp within the same domain); the qualified $\kappa_s$ zone has zero filtering power for $\delta > 5\text{pp}$; under $n=120$ cases/arm, both the 1 pp and 3 pp thresholds fall below the noise floor.

    Partially verified: Real reader training does improve consistency, but the direction focuses on inexperienced readers—after endoscopic atrophic gastritis grading training, $\kappa$ rose from 0.18/0.32 and remained stable (PMID 29629520); after BI-RADS-specific training, new readers’ $\kappa$ improved from 0.41 to 0.50 ($p=0.01$), while experienced readers’ $\kappa$ changed from 0.48 to 0.46 with no difference (PMID 21899969). This indicates training mainly corrects “insufficient calibration,” not “zero-point divergence”—consistent with this article’s inference that “end-point cases only suppress scale,” but none of these studies reported marginal bias, so it remains unverified whether $\delta$ also dropped.

    To be verified: ① In real training studies, the correlation between $\kappa$ improvement and $\delta$ improvement has never been reported side-by-side—currently only Byrt et al. (PMID 8501467) stand: $\kappa$ is simultaneously polluted by bias and prevalence; reporting only $\kappa$ misleads, so both bias and prevalence metrics must be reported; ② Multi-center breast density studies (PMID 37127220) using Stuart-Maxwell marginal homogeneity tests found significant differences across four categories ($P<0.05$), yet weighted consistency reached 0.86—this is a ready-made example of “high consistency but marginal divergence,” but without training intervention, it cannot answer whether training can fix the margin; ③ This article’s four arms are simulated; whether $\delta$ truly rises under “only end-point cases” in real readers requires empirical testing with ≥120 cases × ≥2 readers × four arms; ④ If boundary cases are used instead, the best achievable precision of $\hat{\delta}$ is limited to ±8.8 pp under $n=120$, so detecting 3 pp would require ≥4244 cases/arm.

    Methodological recommendations (actionable): Any study aiming to improve consistency via “unified scoring standards” must report $\hat{\delta}$ and its CI, and specify whether training materials include boundary cases; claiming “training is effective” must first state the detectable lower limit of $\hat{\delta}$, and improvements below that threshold cannot be called effective.

    Related Reading

    Evidence level: C
    References: PMID 8501467, PMID 29629520, PMID 21899969, PMID 37127220
    L1 Network Status: CORRESPONDENCE_NETWORK.md fully grepped for “判读 / kappa / 一致性 / 信度 / 评分 / 读者 / 校准 / 偏倚” hits only row 192 “生克一致性 | 相生/相克方向能否复现 | 未测” (referring to Five Phases generation/control, not reader reliability); the physical side’s most recent position is row 57 “相位 / 节律 | — | 周期性耦合的关键”, while the physiological/experience side has nine nodes (HRV / skin temperature/skin conductance / breathing rate / EMG / EEG / gastrointestinal motility / qi sensation report / constitution typing) with no node for “reader reliability” or “marginal bias”; the edge table stops at E-050 “六十甲子纳音 | 对应 | 60 个频率” (L1 / open, original text states “no measurable source yet”); the network alignment indicator row reads “L2 and above edges | 0”. This topic has not yet entered the L1 corresponding network, and is registered in science-yixue-l1-gaps.md (cluster “measurement consistency,” item 15).


    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.

  • 真为零的那一阶,也会自己报出 11.4% 的共享

    Evidence level: C

    上一轮把「零点看不见」归因于信号太弱,方向反了:不是测不到零,而是真为零时估计量必然报出一个非零值,而这个值比很多阶的真实值还大。

    一、零点会自己造出一个数

    把每个人在某一阶的相位画成单位圆上的箭头,群体共享成分就是合向量长度除以人数:R̂ = |Σe^{ikθ}|/n,总体真值记为 S_k。

    若所有相位完全随机、真值严格为零,R̂ 应该读出 0。但它不会。合向量是一串随机箭头的和,二维随机游走终点到原点的距离服从 Rayleigh 分布,均值 0.8862/√n。n 取 60、200、480,这个「零点读数」分别是 11.44%、6.27%、4.05%——n=60 时,一个真值为零的阶次会自动报出 11% 的群体共享。这是估计量的几何后果,不是仪器噪声。

    二、那条 1.73 的线,把假阳性率放宽了四倍半

    原判据写「S_k 低于 1.73/√n 视为不可判定」。问题在 1.73。

    Rayleigh 检验的统计量是 nR̂²,真值为零时 2×nR̂² 服从自由度 2 的卡方分布,临界值 nR̂² > 5.9915,即 R̂ > 2.448/√n。拿 1.73 当门槛,等价于 nR̂² > 2.996,零假设下该事件概率为 exp(−1.498) = 22.4%,不是 5%。这条线让「零点附近误判为有信号」的概率从 5% 涨到 22.4%,四倍半。正确门槛:n=60 时 31.60%,n=200 时 17.31%,n=480 时 11.17%。

    三、n=60 时 k≥5 全落在同一个坑里

    按 σ=1.43 小时单峰正态,各阶真值为 93.23、75.55、53.22、32.59、17.34、8.02、3.23、1.13(百分数)。加上零点偏倚(读数抬到 √(S²+(1−S²)/n)),n=60 时读出 93.35、76.03、54.33、34.80、21.50、15.16、13.30、12.96。

    临界线 31.60% 一横:k=1 至 k=3 可量化,k=4 勉强擦边,k≥5 全部低于 31.60%,与真值零的 11.44% 混在同一个灰区里。灰区从 11.44% 到 31.60%,宽 20.16 个百分点——比 k≥5 段全部动态范围还宽。要 k=5 有 80% 检出把握需 n≈288,k=6 需 n≈1386,k=8 需六万八千。

    四、真零与真 8% 在 n=60 下分不开

    判据里最关键的一支:「真值 0」与「真值 = Rayleigh 95 分位」这两种情形,现有读法会不会给出同一个动作?

    会。n=60 时真零读 11.44%、真 8.02% 读 15.16%,差 3.72 个百分点,而单次抽样标准差是 0.603/√n = 5.98 个百分点——差距只有噪声的 0.62 倍。任何「低 ⇒ 扣除」的规则对这两个世界给出同一个动作:扣掉。

    更要紧的是代价不对称。要用 95% 置信区间下限压过零、宣布「确实有共享」,需 n≈776;要用上界压过 8.02%、宣布「确实没有共享」,需 n>12,603,差 9.1 倍。证明「有」比证明「没有」便宜一个数量级,而现有流程默认在做后一件事。

    五、一阶零点不等于这一阶没有结构

    设相位为双峰,两峰相隔 ±1.2 小时。k=5 时 kωd = π/2,cos 恰为零——第 5 阶合向量精确等于零,看起来是教科书式的「该阶无共享」。

    但同一组数据里,该阶的二阶矩 |cos(2kωd)| 此刻等于 1.0000,是满值。合向量为零是因为两峰箭头反相抵消,不是因为没有结构:结构被搬进了方差,而 R̂ 只看一阶矩,看不见它。这种抵消还是位置相关的——峰距 1.0 小时零点是第 6 阶,1.2 小时是第 5 阶,1.5 小时是第 4 阶,2.0 小时是第 3 阶。同一个零点可以落在任何一阶上。

    于是「扣掉零点」在两种世界里后果不同:真随机时扣对了(真 0,无损失);真双峰时扣掉的是被反相藏起来的结构。双峰情形下 k=6 真值 30.90%,若按单峰模型判断当噪声扣掉,损失占总二阶矩 3.1%——这不是小,是「本来有、被算丢了」。

    六、两个陷阱:取最小,和去偏

    一,六阶里找最弱那阶扣掉,等于做了六次选择。全为噪声时单阶落在 5% 分位以下的概率 5%,六阶至少一阶落进去的概率 1−0.95⁶ = 26.5%。哪怕六阶全是纯噪声,这个流程也会稳定扣掉一阶——这是多重比较,不是「恰好发现了零点」。

    二,去偏估计 (nR̂²−1)/(n−1) 不解决问题。真零时 nR̂² 落在 1 以下的概率正好 63.2%,被截断成 0;n=60、真 8.02% 时仍有 50.4% 被压成 0。去偏把「假高」换成了「假零」,而按「低就扣」的规则,假零后果更重。

    七、落地口径

    借用分析化学的 LoB / LoD / LoQ 三层(PMID 18852857):低于 LoB(零点 95 分位,2.448/√n)不给判定、只报「不可判定」;高于 LoD(80% 功效门槛)才允许说「该阶有共享」;高于 LoQ(约 2 倍 LoD)才允许报数值。n=60 时三层是 31.60%、38.01%、76.02%——k≥4 全在不可判定区;要让 k=6 可报数值需 n≈2800。

    代价说明白:这套改法会让现有逐阶表在 k≥5 段变成「不可判定」。但原来那个数本来不承载信息——它主要在报 0.886/√n。

    证据与局限

    已验证:零点读数服从 Rayleigh 分布,E[R]=0.8862/√n、SD=0.603/√n,本文 10 万次蒙特卡洛与解析值在 n=30 至 800 全区间吻合(最大偏差 0.0006);临界值 2.448/√n(nR̂²>5.9915)经数值验证 α=0.05,1.73/√n 对应 α=22.4% 为解析结果。相位相干估计量小样本正偏、已有无偏替代是多篇方法学共识(PMID 20114076、21276857、29247676)。

    部分验证:各阶真值取自上一轮单峰正态设定(σ=1.43 小时),具体数值依赖 σ 估计与特征函数「含 1/2 因子」的约定;若约定不同,k=6 真值可低到 0.64%,所需样本量还要再涨两个数量级。故本文阈值表须随 σ 与约定一起重报,不可单独引用。

    待验证:① 双峰下二阶矩为满值是解析结果,未在真实脉搏相位上验证;② 真实数据是否存在精确到零点的一阶未知,若不存在,第五节风险仅为理论;③ 26.5% 是独立性假设下的值,相邻阶次 R̂ 实际相关,真实误杀率应低于此。

    相关阅读

    证据等级: C
    参考文献: PMID 18852857, PMID 20114076, PMID 21276857, PMID 29247676, PMID 1814155

    L1 网络状态:CORRESPONDENCE_NETWORK.md 全表 grep「零点/噪声底/Rayleigh/不可判定/估计/判据」零命中;物理侧最近位为「相位 / 节律 | — | 周期性耦合的关键」,语义为两振荡锁相,不承载「群体相位合向量的零点读数与检出限」;EDGE_TABLE_science-yixue.md 无「噪声底/检出限/不可判定」条目。→ 该主题尚未进入 L1 对应网络。


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

    The first order of True Zero will also report 11.4% sharing on its own.

    Evidence level: C

    In the previous round, attributing “zero is invisible” to weak signal was in the wrong direction: it is not that zero cannot be measured, but rather when the true value is zero, the estimator inevitably reports a non-zero value, and this value is larger than many true values of higher orders.

    I. Zero itself generates a number

    Plotting the phase of each individual at a given order as an arrow on the unit circle, the shared component of the group is the magnitude of the resultant vector divided by the number of people: R̂ = |Σe^{ikθ}|/n, with the overall true value denoted as S_k.

    If all phases were completely random and the true value strictly zero, R̂ should read 0. It does not. The resultant vector is the sum of random arrows; the distance from the origin of a two-dimensional random walk follows a Rayleigh distribution, with mean 0.8862/√n. For n = 60, 200, 480, this “zero reading” is 11.44%, 6.27%, and 4.05% respectively—when n=60, an order with a true value of zero automatically reports 11% shared. This is a geometric consequence of the estimator, not instrument noise.

    II. The line at 1.73 widens the false-positive rate by four and a half times

    The original criterion stated “S_k below 1.73/√n is considered unquantifiable.” The problem lies with 1.73.

    The statistic for the Rayleigh test is nR̂²; when the true value is zero, 2×nR̂² follows a chi-squared distribution with 2 degrees of freedom, with critical value nR̂² > 5.9915, i.e., R̂ > 2.448/√n. Using 1.73 as the threshold is equivalent to nR̂² > 2.996, under which the probability of this event under the null hypothesis is exp(−1.498) = 22.4%, not 5%. This line increases the probability of misjudging “near zero” as “having signal” from 5% to 22.4%, a four-and-a-half-fold increase. Correct thresholds: 31.60% for n=60, 17.31% for n=200, and 11.17% for n=480.

    III. For n=60, all k≥5 fall into the same pit

    Assuming σ=1.43 hours for a single-modal normal distribution, the true values for each order are 93.23, 75.55, 53.22, 32.59, 17.34, 8.02, 3.23, and 1.13 (percentages). Adding the zero bias (reading raised to √(S²+(1−S²)/n)), for n=60 the readings become 93.35, 76.03, 54.33, 34.80, 21.50, 15.16, 13.30, and 12.96.

    With the critical line at 31.60%: orders k=1 to k=3 are quantifiable, k=4 barely touches the edge, and k≥5 all fall below 31.60%, mixing with the 11.44% of true zero in the same gray zone. The gray zone spans from 11.44% to 31.60%, a width of 20.16 percentage points—wider than the entire dynamic range of the k≥5 segment. To achieve 80% detection power for k=5 requires n≈288, for k=6 requires n≈1386, and for k=8 requires 68,000.

    IV. True zero and true 8% are inseparable under n=60

    The most critical branch of the criterion asks: do the cases “true value = 0” and “true value = Rayleigh 95th percentile” yield the same action under current readings?

    Yes. When n=60, true zero reads 11.44% and true 8.02% reads 15.16%, a difference of 3.72 percentage points, while the single-sample standard deviation is 0.603/√n = 5.98 percentage points—the gap is only 0.62 times the noise. Any “low ⇒ subtract” rule gives the same action to both worlds: subtract.

    More critically, the cost is asymmetric. To use the lower bound of the 95% confidence interval to surpass zero and declare “there is shared,” one needs n≈776; to use the upper bound to surpass 8.02% and declare “there is no shared,” one needs n>12,603, a difference of 9.1 times. Proving “yes” is an order of magnitude cheaper than proving “no,” yet the current procedure defaults to doing the latter.

    V. Zeroth order does not mean no structure in this order

    Suppose the phase is bimodal, with peaks separated by ±1.2 hours. At k=5, kωd = π/2, so cos is exactly zero—the resultant vector of the 5th order is precisely zero, appearing as a textbook case of “no shared in this order.”

    However, within the same dataset, the second moment of this order |cos(2kωd)| is currently equal to 1.0000, the maximum value. The resultant vector is zero because the two peak arrows cancel each other out, not because there is no structure: structure is shifted into the variance, while R̂ only sees the first moment and cannot see it. This cancellation is position-dependent—the peak separation of 1.0 hour corresponds to the 6th order, 1.2 hours to the 5th, 1.5 hours to the 4th, and 2.0 hours to the 3rd. The same zero can fall on any order.

    Thus, “subtracting zero” has different consequences in the two worlds: when truly random, subtracting is correct (true 0, no loss); when truly bimodal, what is subtracted is the structure hidden by anti-phase cancellation. In the bimodal case, for k=6 the true value is 30.90%; if judged by the single-modal model and subtracted as noise, this loss accounts for 3.1% of the total second moment—this is not small, it is “originally present but calculated away.”

    VI. Two traps: taking the minimum and debiasing

    First, finding the weakest order among six to subtract equals making six choices. When all are noise, the probability of a single order falling below the 5th percentile is 5%, so the probability that at least one of six falls there is 1−0.95⁶ = 26.5%. Even if all six orders are pure noise, this process will stably subtract one order—this is multiple comparisons, not “coincidentally finding zero.”

    Second, the debiased estimator (nR̂²−1)/(n−1) does not solve the problem. When true zero, nR̂² falls below 1 with probability exactly 63.2%, being truncated to 0; when n=60 and true 8.02%, there is still 50.4% chance of being compressed to 0. Debiasing swaps “false high” for “false zero,” and under the “low ⇒ subtract” rule, false zero has heavier consequences.

    VII. Practical thresholds

    Borrowing the three-tier LoB / LoD / LoQ framework from analytical chemistry (PMID 18852857): below LoB (zero’s 95th percentile, 2.448/√n) do not make a decision, only report “unquantifiable”; above LoD (80% power threshold) allow stating “this order has shared”; above LoQ (about twice LoD) allow reporting numerical values. For n=60, the three tiers are 31.60%, 38.01%, and 76.02%—all k≥4 fall into the unquantifiable zone; to make k=6 reportable numerically requires n≈2800.

    Cost clarification: this modification will render the existing per-order table “unquantifiable” for k≥5. But that number originally carried no information—it mainly reported 0.886/√n.

    Evidence and Limitations

    Verified: Zero readings follow the Rayleigh distribution, E[R]=0.8862/√n, SD=0.603/√n; this article’s 100,000 Monte Carlo simulations match analytical values across the full range n=30 to 800 (maximum deviation 0.0006); critical value 2.448/√n (nR̂²>5.9915) verified numerically with α=0.05, and 1.73/√n corresponding to α=22.4% is an analytical result. Phase coherence estimators are positively biased in small samples; unbiased alternatives exist, which is a consensus across multiple methodological papers (PMID 20114076, PMID 21276857, PMID 29247676).

    Partially verified: True values for each order were taken from the previous round’s single-modal normal setting (σ=1.43 hours); specific values depend on σ estimation and the convention of the characteristic function “containing 1/2 factor”; if the convention differs, the true value for k=6 could drop to 0.64%, requiring sample sizes two orders of magnitude higher. Therefore, this article’s threshold table must be re-reported together with σ and the convention, not cited independently.

    Pending verification: ① The second moment being full under bimodal conditions is an analytical result, not yet verified on real pulse phase data; ② Whether real data contains an unknown first order precisely at zero—if not, the risk in section V is purely theoretical; ③ The 26.5% assumes independence, but adjacent orders R̂ are actually correlated, so the real false-negative rate should be lower.

    Related Reading

    Evidence level: C
    References: PMID 18852857, PMID 20114076, PMID 21276857, PMID 29247676, PMID 1814155

    L1 Network status: CORRESPONDENCE_NETWORK.md contains zero hits for “zero/noise floor/Rayleigh/unquantifiable/estimator/criterion”; the closest physical-side entry is “phase/rhythm | — | key to periodic coupling,” meaning two oscillators are phase-locked, not carrying “zero reading of the group resultant vector and detection limit”; EDGE_TABLE_science-yixue.md lacks “noise floor/detection limit/unquantifiable” entries. → This topic has not yet entered the L1 corresponding network.


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

  • 两个把 σ_k 推向相反方向的机制,在这个判据眼里是同一个 12.9%

    Evidence level: C

    上一节留的课题是:逐阶拟合 σ_k,看 σ_k/k 在 k=1 到 6 上是否恒定——变化小于 15% 就判「单参数模型成立」,下降超过 30% 就判「相位去相干机制不成立」。算下来这条判据两头都不灵:一个把 σ_k 往下拉的机制读数 12.93%,一个把 σ_k 往上推的机制读数 12.92%。两个反方向的机制,在判据眼里是同一个数。

    正文

    先看这条判据问的是什么。 它其实在问一件事:σ_k 是不是 k 的函数。如果相位偏移纯粹是「这个人比那个人早或晚多久」这一个数(时间平移),那每一阶看到的都是同一个 σ,σ_k 除以 k 就该是常数;如果它随 k 变,说明还有第二个东西在动。

    好消息是:纯平移下它完美通过,而且过得毫无信息量。 取 σ = 1.43 小时,逐阶 S_k 用 exp(−(kωσ)²) 算出来,再反解回 σ_k:k=1 到 k=8 全部是 1.4300,一位小数都不差,极差 0.00%。但这一条根本不需要测——它是闭式的同义反复。只要 S_k 是由单一 σ 生成的,反解必然还原出那个 σ。判据在「模型对」时通过,和在「模型是自己造出来」时通过,是同一件事。

    第一个坏消息:另一个完全不同的机制也通过。 假设真实情形不是时间平移,而是波形在抖——每一阶相位都在抖,抖的幅度按弧度算各阶相同(s = 0.20 rad,等效约 0.76 小时)。此时 σ_k = √(1.43² + (s/(kω₀))²),k=1 得 1.621、k=2 得 1.480、k=6 得 1.436,极差 12.93%。

    12.93% 小于 15%,判据说:单参数成立。而真实情况是 σ 从 1.62 掉到 1.44,k=1 处被高估了 13.4%。一个明确存在「第二机制」的数据,被判成了「没有第二机制」。原因很直白:15% 这条线是对比值画的,而 s/(kω₀) 这一项在 k=1 时贡献大、k=6 时只剩六分之一,它的极差天然被压在 13% 上下——s 从 0.10 涨到 0.20,极差从 3.3% 涨到 11.4%,一直待在合格区里。这条判据对「恒定角抖动」几乎是恒真的。

    第二个坏消息,也是最要命的一个:反方向的机制给出几乎同一个数。 换成「相位谱非线性」——波速随频率变,k 越高走得越快,第 k 阶相对第 1 阶多出 (k−1)Δ 的相位散差。取 Δ 的组间 SD = 0.150 小时,σ_k = √(1.43² + ((k−1)·0.150)²),得 1.430、1.438、1.461、1.499、1.551、1.615,极差 12.92%。

    12.93% 与 12.92%,差 0.01 个百分点。一个是 σ_k 从 k=1 单调往下掉(1.62→1.44),一个是单调往上爬(1.43→1.62)。判据读的是极差,极差没有方向,所以它读不出这两个世界。 这不是阈值没调好:只要用「最大减最小」,任何一对方向相反的机制只要幅度相近,就必然撞成同一个读数。

    更麻烦的是现实里两个机制都在。 抖动与频散共存时(s = 0.35 rad、Δ 的 SD = 0.10 小时),σ_k 剖面是 1.958、1.579、1.498、1.504、1.508、1.533——先陡降后微升,极差 30.6%。这回判据踩到「超过 30% 即机制不成立」的线上,但结论照样错:它会宣布「相位去相干在脉搏上不成立」,而真实情况是两个机制都在,只是互相抵掉了一段。判据把「两个机制打架」读成了「机制不存在」。

    还有一个更隐蔽的混入项,判据对它彻底盲。 记录脉搏时心搏间期本来就在变,同一个人 10 bpm 的心率标准差很常见。分析若按固定基频做变换,第 k 阶相位会随基频一起漂,等效散差 = SD(f₀) × 传导时间 × 24 小时。SD(f₀) = 0.167 Hz、传导时间 0.20 秒,得 0.800 小时,占 1.43 小时的 56%。这一项在所有阶次上按同一比例贡献,因此精确折叠进 σ,σ_k/k 的恒定性检验对它一点反应都没有。你测出来的 1.43 小时里可能一大半是心率波动,而判据会告诉你「σ_k/k 很恒定,模型很好」。

    顺带说一句:把 1.43 小时全算成测量误差也站不住。DLMO 四种算法的两夜重测信度好至完美、最优算法与四位专家目视均值只差 5 分钟(PMID 36519316,n=31);径向脉搏谐波前十一阶的观察者间 ICC 在 0.83 至 0.96(PMID 27170904,n=22)。仪器和手法没那么粗,1.43 小时更像真实个体差异而非噪声——这恰恰意味着它不该被当成「一个人只有一个数」。

    能救,但要换三个数。 把 y_k = (kω₀σ_k)² 对 x = k² 做直线回归。纯平移时 y_k 恒等于 (ω₀σ)²,截距 0、残差 0;恒定角抖动时 y_k = s² + (ω₀σ)²k²,截距正好是 s²、残差仍是 0;纯频散时 y_k = (ω₀σ)² + ((k−1)Δω₀)²,对 k² 不再线性,截距算出来是负的 −0.208、残差 0.197。三个读数分工明确:截距开方就是那个「每阶同弧度抖动」有多大(s=0.20 时截距 0.0400、开方 0.200),截距为负说明模型设定本身错了(该用 (k−1)² 而不是 k²),残差大小说明两个机制是否同时在。一个正截距加零残差、一个负截距加残差、一个正截距加中等残差——三种情形分得开,而极差那条线一个都分不开。

    代价是样本量。σ̂_k 的均方根误差随阶次恶化:n=60 时 k=1 是 0.130、k=6 是 0.228;n=480 时降到 0.046 和 0.116。要分辨「σ_k/k 变化 15%」(约 0.215 小时),n=60 合成的标准误 0.262 比效应还大;n=480 时 z 约 1.7,勉强;要到 z=2 得 n 约 700。功效实验更直白:s=0.35 rad 的非线性抖动,n=60 检出率 0.2%、n=120 才 31.7%、n=240 到 100%。也就是说,原设计「至少 60 人同日重复记录」只够排除 s 大于 0.5 rad 这类大抖动(n=60 时检出 93.8%),中间那一整段——恰好是最可能真实存在、也最能让「k=6 只剩 0.6%」这个结论失真的那一段——它看不见。

    物理侧还得交代一句。 血管波确实存在频散:研究已在体内观察到一种反对称的弯曲波模态,它高度色散,且不能用经典 Moens-Korteweg 关系描述,需另建梁理论近似,并在软聚合物血管体模上验证(PMID 38717204)。这说明「血管波速随频率变」是真实的物理,但它讲的是另一种模态;对称压力波的各阶相速谱在 PubMed 上没有直接测数。相位差确实可测且有信息量——348 例冠脉造影研究用左右手谐波相位差等指标预测 SYNTAX 评分,AUC 达 0.89 与 0.92(PMID 40396324)——但该文没有报逐阶相位的组间 SD 谱。所以本篇的结论只能停在方法学层面:这条判据没有鉴别力,无论真实频散是多少。

    证据与局限

    已验证: 纯时间平移下 σ_k 逐阶反解恒为 1.4300(k=1 至 8,极差 0.00%);恒定角抖动 s=0.20 rad 的 σ_k 剖面为 1.621/1.480/1.452/1.443/1.438/1.436、极差 12.93%(判据合格);纯频散 SD(Δ)=0.150 h 的剖面为 1.430/1.438/1.461/1.499/1.551/1.615、极差 12.92%(判据合格,方向相反);两机制共存(s=0.35 rad、SD(Δ)=0.10 h)极差 30.6%(误判为「机制不成立」);三参数回归诊断在四种情形下给出可分离的(截距, 残差)组合:纯平移 (0.0000, 0.00000)、纯抖动 (s², 0.00000)、纯频散 (−0.2076, 0.19735)、共存 (0.0302, 0.08771);σ̂_k 的 RMSE 按 n=60/120/240/480 依次为 k=1 的 0.130/0.097/0.065/0.046 与 k=6 的 0.228/0.178/0.142/0.116;检出 s=0.35 rad 所需 n 在 120 与 240 之间。以上为本组纯数值计算(闭式特征函数求值加蒙特卡洛,600 至 800 次重复),可复现。

    部分验证: 血管波存在频散有体内与体模证据(PMID 38717204,反对称弯曲模态,非对称压力波);脉搏谐波前十一阶振幅信度 ICC 0.83 至 0.96(PMID 27170904,n=22,未报相位 SD);谐波相位差可测且与冠脉病变严重度相关(PMID 40396324,n=348,未报逐阶相位组间 SD);DLMO 两夜重测信度好至完美、与目视均值差 5 分钟(PMID 36519316,n=31);测量误差导致回归斜率系统性稀释这一现象在 2858 个变量的重复测量中有系统刻画(PMID 37336529)。

    待验证: ① 对称压力波各阶相速谱从未实测,本篇「频散幅度取 SD(Δ)=0.15 小时」是情景设定而非实测值;② 真实 σ_k 谱一次都没被报过,任何「σ_k/k 恒定」或「σ_k 随 k 降」的判断目前都无观测依据;③ 心率基线贡献占比 56% 是在 SD(f₀)=0.167 Hz、传导 0.20 秒下算的,真实 PTT 与个体心率变异度分布未纳入;④ 三参数回归的截距是两个端点外推出来的,σ̂_k 的估计误差传到截距会被放大多少倍,本篇未算;⑤ 分布形状若偏离单峰正态(前作已证双峰不可行但浅偏斜未排除),闭式 S_k 本身失真,会叠加进 σ̂_k;⑥ 检验工具的已知弱区:多峰情形下 Rayleigh 类检验功效很低,Hermans-Rasson 在单峰几乎同样强、多峰明显更优,但仍对多于两个集中区的偏离低功效(PMID 31391040);核密度带宽选择在小样本(25 至 100)时会显著改变形状读数(PMID 24885339)——两者都提示「形状与谱形状的判断本身需要报其错误率」。

    相关阅读

    证据等级: C
    参考文献: PMID 27170904, PMID 40396324, PMID 38717204, PMID 36519316, PMID 37336529, PMID 31391040, PMID 24885339
    L1 网络状态:CORRESPONDENCE_NETWORK.md 物理侧节点表原文仅列「相位 / 节律 | — | 周期性耦合的关键」一个概括位(语义为两振荡锁相),生理侧九项节点无「节律相位 / 昼夜时相」,时间层仅有「春 夏 长夏 秋 冬」(年尺度)与「子丑寅卯…亥」(历法编码)、粒度止于日;全表检索「频散 / 色散 / 相位谱 / 逐阶 / 阶次 / 相位偏移」零命中;网络对齐指标行原文「L2 以上边数 | 0」;EDGE_TABLE_science-yixue.md 47 个概念中无「节律 / 昼夜 / 生物钟 / 相位谱」条目。该主题尚未进入 L1 对应网络(「节律方差结构」簇第 4 条),本文证据等级据此封顶 C,正文未作任何已验证机制的表述。


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

    Two mechanisms pushing σ_k in opposite directions, which this criterion sees as the same 12.9%

    Evidence level: C

    The question left open in the previous section is: fit σ_k step-by-step and examine whether σ_k/k remains constant for k = 1 to 6. If the variation stays below 15%, we judge “the single-parameter model holds”; if it drops by more than 30%, we judge “the phase decoherence mechanism does not hold.” Calculations show this criterion fails at both ends: one mechanism pulling σ_k downward reads 12.93%, and another pushing σ_k upward reads 12.92%. Two opposing mechanisms appear identical in this criterion.

    Main text

    First, what does this criterion actually ask? It asks whether σ_k is a function of k. If the phase shift is purely “this person is earlier or later than that one by a certain amount” (a time shift), then every order yields the same σ, and σ_k divided by k should be constant. If it varies with k, something else is moving.

    The good news: pure translation passes perfectly, but carries no information. Taking σ = 1.43 hours, we compute S_k step-by-step using exp(−(kωσ)²) and solve back for σ_k: for k = 1 to 8, all values are exactly 1.4300, with zero difference at the first decimal place (range 0.00%). But this result requires no measurement—it is a closed-form tautology. If S_k is generated by a single σ, solving back necessarily recovers that σ. The criterion passes when the “model is right” and when the “model is self-generated”—it is the same thing.

    The first bad news: another completely different mechanism also passes. Suppose the true situation is not time translation but waveform jitter—each order’s phase jitters, with the jitter amplitude (in radians) being identical across orders (s = 0.20 rad, equivalent to approximately 0.76 hours). Then σ_k = √(1.43² + (s/(kω₀))²). For k = 1, σ_k = 1.621; for k = 2, σ_k = 1.480; for k = 6, σ_k = 1.436. The range is 12.93%.

    Since 12.93% < 15%, the criterion declares “single-parameter holds.” Yet the truth is σ drops from 1.62 to 1.44, with k = 1 overestimated by 13.4%. A clear “second mechanism” is judged as “no second mechanism.” The reason is straightforward: the 15% line is drawn relative to ratios, while the term s/(kω₀) contributes heavily at k = 1 but only one-sixth at k = 6, so its range is naturally compressed around 13%. As s increases from 0.10 to 0.20, the range grows from 3.3% to 11.4%, staying within the acceptable zone. This criterion is almost always true for “constant angular jitter.”

    The second bad news, and the most fatal: the opposite-direction mechanism yields almost the same number. Consider “phase spectrum nonlinearity”—wave speed varies with frequency, so higher k moves faster, adding (k−1)Δ phase dispersion between the k-th and 1st orders. Taking the inter-group SD of Δ = 0.150 hours, σ_k = √(1.43² + ((k−1)·0.150)²). The results are 1.430, 1.438, 1.461, 1.499, 1.551, 1.615, with a range of 12.92%.

    The difference between 12.93% and 12.92% is just 0.01 percentage points. In one case, σ_k monotonically decreases from k = 1 (1.62 → 1.44); in the other, it monotonically increases (1.43 → 1.62). The criterion reads the range, which has no direction, so it cannot distinguish these two worlds. This is not due to a poorly set threshold: as long as “maximum minus minimum” is used, any pair of opposite-direction mechanisms with similar magnitudes will inevitably collide into the same reading.

    Even more troubling is that both mechanisms coexist in reality. When jitter and dispersion occur together (s = 0.35 rad, SD of Δ = 0.10 hours), the σ_k profile is 1.958, 1.579, 1.498, 1.504, 1.508, 1.533—steeply dropping then slightly rising, with a range of 30.6%. This time the criterion hits the “exceeds 30% means mechanism fails” line, yet the conclusion remains wrong: it declares “phase decoherence does not hold in the pulse,” while the truth is both mechanisms are present, merely canceling each other out over a segment. The criterion reads “two mechanisms fighting” as “no mechanism exists.”

    There is also a more hidden confounding term that the criterion completely misses. When recording the pulse, inter-beat intervals naturally vary; a standard deviation of 10 bpm is common for the same individual. If analysis uses a fixed fundamental frequency, the phase of the k-th order drifts with it, creating an equivalent dispersion equal to SD(f₀) × conduction time × 24 hours. With SD(f₀) = 0.167 Hz and conduction time = 0.20 seconds, this equals 0.800 hours, which is 56% of 1.43 hours. This term contributes proportionally across all orders, so it folds precisely into σ, making the constancy test of σ_k/k unresponsive. Much of the measured 1.43 hours may actually be heart rate fluctuation, yet the criterion would tell us “σ_k/k is very constant; the model is good.”

    By the way, treating all 1.43 hours as measurement error is untenable. DLMO reliability across two nights using four algorithms is perfect; the optimal algorithm differs from the mean of four expert visual assessments by only 5 minutes (PMID 36519316, n=31); inter-observer ICC for the first eleven harmonics of radial pulse is between 0.83 and 0.96 (PMID 27170904, n=22). Instruments and methods are not that coarse; 1.43 hours is more likely real individual variation than noise—exactly why it should not be treated as “one person has only one number.”

    A rescue is possible, but requires three numbers. Perform linear regression of y_k = (kω₀σ_k)² against x = k². Under pure translation, y_k is constantly (ω₀σ)², with intercept 0 and residual 0; under constant angular jitter, y_k = s² + (ω₀σ)²k², with intercept exactly s² and residual still 0; under pure dispersion, y_k = (ω₀σ)² + ((k−1)Δω₀)², which is no longer linear in k², yielding a negative intercept of −0.208 and residual 0.197. The three readings have clear roles: the square root of the intercept gives the magnitude of “same-radian jitter per order” (s = 0.20 yields intercept 0.0400, square root 0.200); a negative intercept indicates the model specification itself is wrong (should use (k−1)² instead of k²); the residual magnitude indicates whether both mechanisms are active. A positive intercept with zero residual, a negative intercept with residual, and a positive intercept with moderate residual—these three cases are clearly separated, whereas the range line cannot distinguish any of them.

    The cost is sample size. The root mean square error of σ̂_k worsens with order: for n = 60, k = 1 is 0.130 and k = 6 is 0.228; for n = 480, they drop to 0.046 and 0.116. To detect “σ_k/k changes by 15%” (about 0.215 hours), the standard error for n = 60 is 0.262, larger than the effect; for n = 480, z is about 1.7, barely sufficient; to reach z = 2, n must be about 700. Power analysis is even clearer: for s = 0.35 rad nonlinear jitter, detection rates are 0.2% for n = 60, 31.7% for n = 120, and 100% for n = 240. Thus, the original design of “at least 60 people with repeated records on the same day” can only exclude large jitters like s > 0.5 rad (detection rate 93.8% for n = 60); the middle segment—exactly where real effects most likely exist and where “k = 6 drops to only 0.6%” would be distorted—is invisible.

    A physical note must also be added. Vascular waves do exhibit dispersion: research has observed an antisymmetric bending wave mode in vivo, which is highly dispersive and cannot be described by the classic Moens-Korteweg relation; instead, a beam theory approximation must be constructed and validated on soft polymer vascular phantoms (PMID 38717204). This confirms “vascular wave speed varies with frequency” is real physics, but it refers to another mode; the inter-order phase velocity spectrum of symmetric pressure waves is not directly measured in PubMed. Phase differences are indeed measurable and informative—348 coronary angiography studies used left-right harmonic phase difference and other metrics to predict SYNTAX scores, achieving AUC of 0.89 and 0.92 (PMID 40396324)—but that paper did not report the inter-group SD spectrum of phase step-by-step. Therefore, this paper’s conclusion can only stop at the methodological level: this criterion lacks discriminatory power, regardless of the true dispersion magnitude.

    Evidence and Limitations

    Verified: Under pure time translation, σ_k solved step-by-step is constantly 1.4300 (k = 1 to 8, range 0.00%); for constant angular jitter s = 0.20 rad, the σ_k profile is 1.621/1.480/1.452/1.443/1.438/1.436, with range 12.93% (criterion passes); for pure dispersion SD(Δ) = 0.150 h, the profile is 1.430/1.438/1.461/1.499/1.551/1.615, with range 12.92% (criterion passes, opposite direction); when both mechanisms coexist (s = 0.35 rad, SD(Δ) = 0.10 h), the range is 30.6% (misjudged as “mechanism fails”); the three-parameter regression diagnosis yields separable (intercept, residual) combinations across four cases: pure translation (0.0000, 0.00000), pure jitter (s², 0.00000), pure dispersion (−0.2076, 0.19735), coexistence (0.0302, 0.08771); the RMSE of σ̂_k for n = 60/120/240/480 is k = 1: 0.130/0.097/0.065/0.046 and k = 6: 0.228/0.178/0.142/0.116; detection of s = 0.35 rad requires n between 120 and 240. These are pure numerical calculations by this group (closed-form characteristic function evaluation plus Monte Carlo, 600 to 800 repetitions), reproducible.

    Partially verified: Vascular wave dispersion has in vivo and phantom evidence (PMID 38717204, antisymmetric bending mode, asymmetric pressure wave); inter-observer ICC for the first eleven harmonics of pulse wave amplitude is 0.83 to 0.96 (PMID 27170904, n=22, phase SD not reported); harmonic phase differences are measurable and correlate with coronary lesion severity (PMID 40396324, n=348, step-by-step phase inter-group SD not reported); DLMO two-night retest reliability is perfect, differing from visual mean by 5 minutes (PMID 36519316, n=31); regression slope systematic dilution due to measurement error has been systematically characterized in repeated measures of 2,858 variables (PMID 37336529).

    To be verified: ① The inter-order phase velocity spectrum of symmetric pressure waves has never been directly measured; this paper’s “dispersion magnitude set to SD(Δ) = 0.15 hours” is a scenario assumption, not an observed value; ② The true σ_k spectrum has never been reported once, so any judgment of “σ_k/k constant” or “σ_k decreases with k” currently lacks observational basis; ③ The heart rate baseline contribution of 56% was calculated under SD(f₀) = 0.167 Hz and conduction time 0.20 seconds, without incorporating real PTT distribution and individual heart rate variability; ④ The intercept of the three-parameter regression is extrapolated from two endpoints, and how much the estimation error of σ̂_k amplifies at the intercept was not calculated in this paper; ⑤ If the distribution shape deviates from single-peaked normal (previous work showed double-peaked is infeasible but shallow skewness was not excluded), the closed-form S_k itself becomes distorted, adding into σ̂_k; ⑥ Known weak areas of the testing tool: under multi-peaked conditions, Rayleigh-type tests have low power; Hermans-Rasson is almost equally strong for single-peaked but clearly superior for multi-peaked, yet still has low power against deviations involving more than two concentration regions (PMID 31391040); kernel density bandwidth selection significantly alters shape readings in small samples (25 to 100) (PMID 24885339)—both suggest “judgment of shape and spectral shape itself must report its error rate.”

    Related Reading

    Evidence level: C
    References: PMID 27170904, PMID 40396324, PMID 38717204, PMID 36519316, PMID 37336529, PMID 31391040, PMID 24885339
    L1 network status: The original physical-side node table in CORRESPONDENCE_NETWORK.md lists only “phase / rhythm | — | key periodic coupling” as a summary position (meaning two oscillators phase-locked); the nine physiological-side nodes lack “rhythm phase / circadian phase”; the time layer includes only “Spring Summer Late Summer Autumn Winter” (annual scale) and “Zi Chou Yin Mao… Hai” (calendar encoding), with granularity stopping at day; searching the full table for “dispersion / chromatic dispersion / phase spectrum / step-by-step / order / phase shift” yields zero hits; the network alignment indicator row originally states “L2 and above edges | 0”; EDGE_TABLE_science-yixue.md contains none of the 47 concepts related to “rhythm / circadian / biological clock / phase spectrum.” This topic has not yet entered the corresponding L1 network (cluster #4 of “rhythm variance structure”), so this paper’s evidence level is capped at C, with no claims of verified mechanisms in the main text.


    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.

  • 被叫作双峰的那个分布,压了一整篇论文的中心峰从来没低过头

    Evidence level: C

    上一节留下的课题是「把相位分布从单峰换成晨晚型双峰,看逐阶共享占比 S_k 会不会在中间阶次塌下去」。算完才发现第一步就错了:那套参数化里中心峰的高度恒为中心两侧的 3.00 倍,根本没有凹陷——叫它双峰,只是因为它有三个峰。而真正能满足「双峰」且「总标准差仍是 1.43 小时」的参数组合,只剩一条很窄的带,窄到刚好把一个精确零点装在里面。

    正文

    一、三个峰不等于双峰

    原设想是:60% 的人相位在中心、各 20% 落在 ±d,调节两侧峰的宽度 w 让总标准差始终保持 1.43 小时——这样只换形状、不换 σ,与上一篇的比值判据正交,可以单独做。

    这套参数下总标准差确实恒为 1.43(实测 1.430),但峰高比从头到尾是常数:中心峰 / 肩峰 = 3.00(w = 0.05 时 3.0,w = 0.6 时仍是 3.0,w = 1.0 才降到 1.9)。也就是说中心那块密度永远比两侧高,密度曲线只有三个极大值,中间的谷不存在。一个有三个峰、且中心最高的分布,统计上不是双峰,是主峰带两个肩。

    这不是算术小事。整条推导的动机是「有一撮人偏得特别远,会在高阶留下不衰减的旁瓣」,而肩峰恰恰不产生旁瓣——它离中心太近了。用它算出来的逐阶 S_k 与正态几乎重合,于是「正态近似够用」会被这个实验证实,但证实的其实是一个单峰分布。用一个不是双峰的双峰去检验双峰假设,结论必然是「双峰不存在」。

    二、σ = 1.43 装不下一个真双峰

    要真双峰(中心低于两侧),只能是 50/50:两峰在 ±m,各自宽 w,约束 m² + w² = 1.43²。扫下来,中心/峰高比 ≤ 1 的可行域是 w ≤ 0.99 小时、m ≥ 1.02 小时。一旦 w 超过 1.0,中心又比两侧高了,退回单峰。

    换句话说,在总离散只有 1.43 小时的约束下,你要摆出一个真双峰,两侧峰心必须至少分开 2.04 小时,而两个峰各自的宽度不能超过 0.99 小时。总离散就这么大,双峰已经把预算全用完了。

    这与实测有直接冲突:极端晨型与极端晚型的外周钟基因相位差实测 2.61 h(Per3)与 3.08 h(Nr1d2),均值 2.86 h(PMID 27103930,14 人)。如果人群真的是两峰各半,峰间距就在 2.86 小时量级,对应的总标准差至少 1.43 小时、更可能更大。所以「σ = 1.43」与「真双峰」这两个前提不能同时为真——至少有一个要被换掉,而上一篇整张 S_k 表的绝对数值全部建立在 σ = 1.43 上。

    三、可行域里刚好装着一个精确零点

    真双峰下特征函数的组间相关项是 cos(kωm),ω = 2π/24。它在 kωm = π/2 时归零,解出 m = 6/k 小时:k=3 → 2.0 h,k=4 → 1.5 h,k=5 → 1.2 h,k=6 → 1.0 h。

    可行域 m ∈ [1.02, 1.43] 里,唯一落进去的是 k=5 的 1.2 小时(此时 w = 0.778,中心/峰高比 0.193,是货真价实的双峰)。实测:S5 = 0.00000,S4 = 0.222(正态 0.326),S6 = 0.147(正态 0.080,高出 0.067)。k=6 的零点在 1.0 小时,刚好差 0.02 小时落在域外——边界擦肩。

    S5 精确归零的后果比「高阶回升」严重得多。现有流程对 S_k 的读法是:S_k 低 ⇒ 该阶主要是各人各抖 ⇒ 必须个体化校正掉。若观测到 S5 ≈ 0,标准流程会判定「第 5 阶纯属个体差异,整体扣除」,而真相是群体在这一阶完美反相抵消——把一段真实的群体信号当成噪声扣掉了。这不是校正不足,是校正方向反了。相位分布的双峰性本身是被反复讨论的表型(EPISONO 队列 857 人中 16% 判为双峰型,PMID 28426387;但那是问卷作答模式,不是相位形态学),而大样本人群睡眠时型近高斯也有横断面证据(PMID 17936039,MCTW 五万余人)——两者并不互斥,「多数人近高斯 + 少数极端型」这条混合恰好产生浅凹陷,而浅凹陷在总 σ 固定的条件下正好能碰出零点。

    更要紧的是零点根本看不见。S5 的估计量是 |Σe^{i5θ}|/n,在真值为 0 时它的分布不是 0 附近的一个窄峰,而是 Rayleigh 分布,期望 0.886/√n。实测:n = 40 估出 0.140、n = 80 估出 0.099、n = 200 估出 0.063。也就是说,真值为 0 的那一阶,在上一篇设计的 80 人样本里会显示「仍有 10% 共享」——而正态在第 5 阶是 0.173,两者只差 43%,零点与正常值在同一个量级上。若要把这个噪声底压到上一篇自己定的可测门槛 S_k > 0.02 以下,需要 n ≈ 1960。真值是不是 0,在这个样本量级上是一个不可判定的问题。

    而且零点很窄:峰间距从 1.20 h 偏到 1.18 h,S5 立刻从 0 升到 0.015;偏到 1.15 h 就是 0.035,偏到 1.10 h 是 0.064。要在数据里认出这个零点,峰间距必须被确定到 ±0.02 小时以内——而 DLMO 单次测量的个体内 SD 是 0.6 小时,是那个窗口的 30 倍。

    四、噪声一加,差异只剩在最测不到的那一阶

    DLMO 单次测量的个体内 SD 就有 0.6 小时。把这个噪声卷进双峰(m = 1.298、w = 0.6)后,与正态的逐阶差变成:k=1 −0.001、k=2 −0.015、k=3 −0.057、k=4 −0.126、k=5 −0.058、k=6 +0.134。

    前五阶全是负数(双峰反而更低,方向与「双峰留旁瓣」的直觉相反),唯一保留正号、也唯一保留可检出幅度的是第 6 阶。而第 6 阶恰恰是上一篇已经证明「在 r > 1.76 时全灭」的那一阶。能看见形状差异的那一阶,就是测不到的那一阶。

    五、两个自检手段,一个误报一个漏报

    第一个是「模型指纹」:log(−ln S_k) 对 log k 的斜率,正态真值恰为 2.000。1500 次蒙特卡洛实测:正态、n = 40 时斜率均值 1.924、95%CI 半宽仅 0.146,覆盖真值 2 的比例只有 42.7%;n = 200 时均值 1.987、半宽 0.063、覆盖 33.2%——样本越大 CI 越窄,误报率反而升到 67%。而在双峰(w = 0.6)下斜率只有 1.81–1.84,CI 半宽却宽到 0.23–0.28,覆盖 2 的比例 72%–88%。同一把尺子:对真单峰有一半以上概率误报,对真双峰有七成以上概率放过。这条指纹不能用来给闭式背书。

    第二个是直接比 S6。w = 0.6 的双峰与正态,S6 真值差 0.049–0.068。蒙特卡洛给出所需样本量:n = 400 时标准化差仅 1.28(不可分),n = 800 才到 1.88,n = 1600 才 2.69。而上一篇设计的实测方案是 ≥ 80 人同日两次 DLMO——差一个数量级。在可行样本量内,形状检验这件事做不了。分布形状的判别力与实验可行性,也是一把剪刀差。

    证据与局限

    已验证: 60/20/20 参数化下中心峰/肩峰高度比恒为 3.00(w = 0.05 至 0.6 实测不变),该分布不是双峰;真双峰需 w ≤ 0.99 h、m ≥ 1.02 h;零点族 m = 6/k 在可行域内唯一可达 k=5(1.2 h),此时 S5 = 0.00000、S6 = 0.147(正态 0.080);加 0.6 h 噪声后仅 k=6 差值为正;斜率指纹对正态 n=40/200 的覆盖率 42.7%/33.2%,对双峰 72%/88%;S5 真值为 0 时估计量服从 Rayleigh,n=40/80/200 实测均值 0.140/0.099/0.063(理论 0.886/√n)。以上均为本组纯数值计算(IEEE 双精度,闭式特征函数直接求值),可复现。

    部分验证: 晨晚型相位差 2.61/3.08 h 有实测(PMID 27103930,n=14,毛囊钟基因三时点,样本极小、仅极型人群);人群睡眠时型近高斯有横断面支持(PMID 17936039,MCTW 五万余人,为问卷而非 DLMO 相位);双峰表型占 16% 有队列数据(PMID 28426387,EPISONO n=857,问卷算法判别,非相位形态学)。

    待验证: ① 真实 DLMO 相位分布的形状从未做过大样本直方图拟合——所有「双峰」讨论都在问卷层面,相位层面是空白;② 若真实分布确有浅凹陷,其峰间距 m 落在 1.02–1.43 h 这段窄带内的概率有多大,无任何数据;③ m 稍偏离 1.2 h 时零点会移动而非消失,零点两侧 S5 上升的速度(本组只算了 1.2 h 附近 ±0.05 h)与偏倚方向未系统扫过;④ 双峰判别工具本身有已知偏倚:偏度会系统性抬高双峰检出率,BC 与 Hartigan dip 在部分配置下不一致(PMID 22806703 的系统评估),本组未把这些工具的误判率纳入模拟。

    相关阅读

    证据等级: C
    参考文献: PMID 27103930, PMID 28426387, PMID 17936039, PMID 22806703, PMID 34396981
    L1 网络状态:CORRESPONDENCE_NETWORK.md 物理侧节点表原文仅列「相位 / 节律 | — | 周期性耦合的关键」一个概括位(语义为两振荡锁相),生理侧九项节点无「节律相位/昼夜时相」,时间层仅有「春 夏 长夏 秋 冬」(年尺度)与「子丑寅卯…亥」(历法编码)、粒度止于日;网络对齐指标行原文「L2 以上边数 | 0」;EDGE_TABLE_science-yixue.md 47 个概念中无「节律/昼夜/生物钟」条目。该主题尚未进入 L1 对应网络(「节律方差结构」簇第 3 条),本文证据等级据此封顶 C,正文未作任何已验证机制的表述。


    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 distribution called bimodal, its central peak in the entire paper never dipped below the pressure

    Evidence level: C

    The question left open in the previous section was “whether shifting the phase distribution from single-peaked to morning-evening double-peaked would cause the stepwise shared fraction $S_k$ to collapse at intermediate orders.” Upon completing the calculation, it was discovered that the first step was wrong: in that parameterization, the height of the central peak is constantly 3.00 times that of the flanking peaks, meaning there is no depression—calling it double-peaked merely because it has three peaks. The only parameter combination that truly satisfies “double-peaked” while maintaining a “total standard deviation of 1.43 hours” is a very narrow band, narrow enough to fit exactly one precise zero point inside it.

    Main Text

    1. Three Peaks Does Not Equal Double-Peak

    The original hypothesis was: 60% of people have phases centered, with 20% each falling at ±d; adjusting the width $w$ of the flanking peaks to keep the total standard deviation constantly at 1.43 hours—changing only the shape, not $\sigma$, making it orthogonal to the ratio criterion of the previous paper and allowing separate analysis.

    Under these parameters, the total standard deviation is indeed constantly 1.43 (measured as 1.430), but the peak-height ratio remains constant throughout: central peak / shoulder peak = 3.00 (3.0 when $w = 0.05$, still 3.0 when $w = 0.6$, dropping only to 1.9 when $w = 1.0$). Thus, the density in the center is always higher than on the sides; the density curve has only three maxima, and the intermediate valley does not exist. A distribution with three peaks, where the center is highest, is statistically not double-peaked but a main peak with two shoulders.

    This is not a trivial arithmetic matter. The motivation for the entire derivation was “a small group deviates particularly far, leaving non-decaying side lobes at higher orders,” whereas the shoulder peaks do not produce side lobes—they are too close to the center. The stepwise $S_k$ calculated using them almost coincides with the normal distribution, so “the normal approximation suffices” would be confirmed by this experiment, but what is confirmed is actually a single-peaked distribution. Using a non-double-peaked distribution to test the double-peaked hypothesis inevitably leads to the conclusion that “the double peak does not exist.”

    2. $\sigma = 1.43$ Cannot Accommodate a True Double-Peak

    For a true double peak (center lower than sides), it must be 50/50: two peaks at ±m, each with width $w$, constrained by $m^2 + w^2 = 1.43^2$. Scanning this, the feasible region where central/peak height ratio ≤ 1 is $w \leq 0.99$ hours, $m \geq 1.02$ hours. Once $w$ exceeds 1.0, the center becomes higher than the sides again, reverting to single-peaked.

    In other words, under the constraint of total dispersion of only 1.43 hours, to arrange a true double peak, the centers of the two flanking peaks must be at least 2.04 hours apart, while the width of each peak cannot exceed 0.99 hours. The total dispersion is so limited that the double peak exhausts the entire budget.

    This directly conflicts with measurement: the peripheral clock gene phase difference between extreme morning types and extreme evening types is measured at 2.61 h (Per3) and 3.08 h (Nr1d2), with a mean of 2.86 h (PMID 27103930, n=14). If the population were truly half-and-half double-peaked, the peak separation would be on the order of 2.86 hours, implying a total standard deviation of at least 1.43 hours, and more likely larger. Therefore, the premises “$\sigma = 1.43$” and “true double peak” cannot both be true—at least one must be replaced, and the entire $S_k$ table of the previous paper is built absolutely on $\sigma = 1.43$.

    3. Exactly One Precise Zero Point Fits in the Feasible Region

    Under true double peak, the inter-group correlation term of the characteristic function is $\cos(k\omega m)$, where $\omega = 2\pi/24$. It vanishes when $k\omega m = \pi/2$, solving to $m = 6/k$ hours: $k=3 \to 2.0$ h, $k=4 \to 1.5$ h, $k=5 \to 1.2$ h, $k=6 \to 1.0$ h.

    In the feasible region $m \in [1.02, 1.43]$, the only value that falls inside is $k=5$ at 1.2 hours (here $w = 0.778$, central/peak height ratio 0.193, a genuine double peak). Measurement: $S_5 = 0.00000$, $S_4 = 0.222$ (normal 0.326), $S_6 = 0.147$ (normal 0.080, higher by 0.067). The zero point for $k=6$ is at 1.0 hour, just 0.02 hours outside the domain—a boundary miss.

    The consequence of $S_5$ precisely vanishing is far more severe than “higher-order resurgence.” The existing procedure interprets $S_k$ as: low $S_k \Rightarrow$ that order is mainly individual jitter $\Rightarrow$ must be individually corrected out. If $S_5 \approx 0$ is observed, the standard procedure would judge “the 5th order is purely individual difference, subtracted from the whole,” while the truth is the group perfectly cancels out in this order—a real group signal is treated as noise and subtracted. This is not insufficient correction; it is correction in the wrong direction. The double-peaked nature of phase distribution itself is a repeatedly discussed phenotype (16% judged as double-peaked in the EPISONO cohort of 857 people, PMID 28426387; but that is questionnaire response mode, not phase morphology), and large-sample population chronotypes near Gaussian also have cross-sectional evidence (PMID 17936039, MCTW over 50,000 people)—these do not mutually exclude. The mixed pattern of “most near Gaussian + few extremes” exactly produces a shallow depression, and under fixed total $\sigma$, this shallow depression precisely hits the zero point.

    More critically, the zero point is invisible. The estimator for $S_5$ is $|\sum e^{i5\theta}|/n$; when the true value is 0, its distribution is not a narrow peak near 0 but a Rayleigh distribution, with expectation $0.886/\sqrt{n}$. Measurement: $n=40$ estimates 0.140, $n=80$ estimates 0.099, $n=200$ estimates 0.063. Thus, the order with true value 0 would show “still 10% shared” in the previous paper’s designed sample of 80 people—while normal is 0.173 at the 5th order, differing by only 43%, the zero and normal values are on the same order of magnitude. To suppress this noise floor below the measurable threshold $S_k > 0.02$ set by the previous paper, one needs $n \approx 1960$. Whether the true value is 0 is an undecidable question at this sample size.

    Moreover, the zero point is very narrow: shifting peak separation from 1.20 h to 1.18 h immediately raises $S_5$ from 0 to 0.015; shifting to 1.15 h gives 0.035, shifting to 1.10 h gives 0.064. To recognize this zero point in the data, peak separation must be determined within ±0.02 hours—yet the individual SD of DLMO single measurement is 0.6 hours, thirty times larger than that window.

    4. Adding Noise Makes Differences Visible Only at the Least Detectable Order

    The individual SD of DLMO single measurement is already 0.6 hours. Rolling this noise into the double peak ($m = 1.298, w = 0.6$), the stepwise difference from normal becomes: $k=1 \,-0.001$, $k=2 \,-0.015$, $k=3 \,-0.057$, $k=4 \,-0.126$, $k=5 \,-0.058$, $k=6 \,+0.134$.

    The first five orders are all negative (the double peak is actually lower, contrary to the intuition of “double peak leaves side lobes”), and the only order retaining a positive sign—and the only one retaining detectable amplitude—is the 6th order. Yet the 6th order is precisely the one already proven to “completely vanish when $r > 1.76$” in the previous paper. The order where shape differences can be seen is the one that cannot be measured.

    5. Two Self-Checks: One False Positive, One Missed Detection

    The first is “model fingerprint”: the slope of $\log(-\ln S_k)$ vs. $\log k$; for normal, the true value is exactly 2.000. Monte Carlo simulation of 1500 runs: for normal, $n=40$, the mean slope is 1.924 with 95% CI half-width only 0.146, covering the true value 2 with probability 42.7%; for $n=200$, mean is 1.987, half-width 0.063, covering 33.2%—larger samples narrow the CI but raise the false positive rate to 67%. Under double peak ($w=0.6$), the slope is only 1.81–1.84, yet the CI half-width widens to 0.23–0.28, covering 2 with probability 72%–88%. With the same ruler: more than half chance of false positive for true single peak, more than 70% chance of missing true double peak. This fingerprint cannot be used to endorse closed-form solutions.

    The second is direct comparison of $S_6$. For double peak with $w=0.6$ vs. normal, the true difference in $S_6$ is 0.049–0.068. Monte Carlo gives required sample size: $n=400$ yields standardized difference only 1.28 (indistinguishable), $n=800$ reaches 1.88, $n=1600$ reaches 2.69. Yet the experimental design of the previous paper is ≥80 people with two DLMO measurements on the same day—off by an order of magnitude. Within feasible sample sizes, shape testing cannot be done. The discriminative power of distribution shape and experimental feasibility are mismatched like scissors.

    Evidence and Limitations

    Verified: Under the 60/20/20 parameterization, the central peak/shoulder peak height ratio is constantly 3.00 (measured unchanged from $w=0.05$ to $w=0.6$); this distribution is not double-peaked; true double peak requires $w \leq 0.99$ h, $m \geq 1.02$ h; the zero-family $m=6/k$ is uniquely reachable at $k=5$ (1.2 h) within the feasible region, where $S_5 = 0.00000$, $S_6 = 0.147$ (normal 0.080); adding 0.6 h noise leaves only $k=6$ with positive difference; the slope fingerprint covers normal $n=40/200$ at 42.7%/33.2%, and double peak at 72%/88%; when $S_5$ true value is 0, the estimator follows Rayleigh, with measured means for $n=40/80/200$ at 0.140/0.099/0.063 (theoretical $0.886/\sqrt{n}$). All are pure numerical calculations by this group (IEEE double precision, closed-form characteristic function directly evaluated), reproducible.

    Partially Verified: Morning-evening phase differences of 2.61/3.08 h have measurement (PMID 27103930, n=14, three time points of hair clock genes, extremely small sample, only extreme phenotypes); population chronotypes near Gaussian have cross-sectional support (PMID 17936039, MCTW over 50,000 people, but questionnaire rather than DLMO phase); double-peak phenotype accounts for 16% has cohort data (PMID 28426387, EPISONO n=857, questionnaire algorithm judgment, not phase morphology).

    Pending Verification: ① The shape of true DLMO phase distribution has never been fitted with large-sample histograms—all “double peak” discussions are at the questionnaire level, leaving phase morphology blank; ② If true distribution does have shallow depression, what is the probability that its peak separation $m$ falls within the narrow band 1.02–1.43 h, with no data; ③ When $m$ slightly deviates from 1.2 h, the zero point shifts rather than disappears; the rate of $S_5$ rise on both sides of the zero (this group only calculated near ±0.05 h around 1.2 h) and the direction of bias have not been systematically scanned; ④ The double-peak discrimination tools themselves have known biases: skewness systematically elevates double-peak detection rate, while BC and Hartigan dip disagree under some configurations (systematic evaluation in PMID 22806703); this group has not incorporated the misjudgment rates of these tools into simulations.

    Further Reading

    Evidence Level: C
    References: PMID 27103930, PMID 28426387, PMID 17936039, PMID 22806703, PMID 34396981
    L1 Network Status: The original physical-side node table in CORRESPONDENCE_NETWORK.md lists only “phase / rhythm | — | key periodic coupling” as a summary position (meaning two oscillators phase-locked); the physiological-side nine nodes lack “rhythm phase / circadian phase”; the time layer includes only “spring summer long-summer autumn winter” (annual scale) and “zi chou yin mao… hai” (calendar encoding), with granularity stopping at day; the network alignment indicator row reads “L2 and above edges | 0”; among 47 concepts in EDGE_TABLE_science-yixue.md, there is no entry for “rhythm / circadian / biological clock.” This topic has not entered the corresponding L1 network (cluster #3 of “rhythm variance structure”), so this paper’s evidence level is capped at C, with no claims of verified mechanisms in the main text.


    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.

  • 标准差之比这把尺子,量不出尺子挪了地方:一个精确等于 1.000 的合格读数,背后可以是八个百分点的偏倚

    Evidence level: C

    上一篇建议用「两位判读者连续打分的标准差之比」来替代那个从没测过的偏倚 δ。这个建议只有一个漏洞,但它致命:δ 由两个自由度决定,而这个比值只读其中一个。漏掉的那个自由度,能把 δ 直接推到八个点,同时让这个比值精确等于 1.000。

    正文

    上一篇把偏倚 δ 的量级归给「两位读者心里的尺子刻度是否同源」,并提出一个改法:让每位读者对同一批人的变化幅度做 0–100 连续打分,看两人打分的标准差之比 κ_s;比值在 0.9–1.1 就是尺子同源、δ 结构上在 1pp 量级。理由是连续量的方差估计效率远高于比例差,精度能高一个数量级。

    这个改法错在两处:漏了一个自由度;「效率高一个数量级」这句话,方向是反的。

    一、κ_s 对位置是精确盲的

    每位读者对同一个人打分 s_j = a_j·x + b_j,x 是真实变化程度。于是读者的全部行为只有两个自由度:尺度 a(尺子多长)和位置 b(零点在哪)。κ_s = a₁/a₂ 只读第一个。

    现在做纯平移:a₁ = a₂ = 1,b₁ = 0,b₂ = τ。此时第二位打分恒等于第一位加常数,两人打分的标准差完全相同,κ_s 精确等于 1.000——不是「约等于」,是恒等式。但主档占比照样分家:

    τ(单位=共同潜变量 SD) 0.2 0.4 0.6 0.9
    κ_s 1.000 1.000 1.000 1.000
    δ(m=0.75) 0.9pp 3.7pp 8.1pp 17.1pp
    δ(m=0.90) 0.7pp 2.7pp 6.0pp 13.4pp

    (δ 定义为「读者一主档占比 − 读者二」,正值表示读者二更倾向把人有变化、主档更少。)

    要把 δ 推到 5pp 这条硬线,纯平移只需 τ = 0.47 SD(m=0.75)或 0.55 SD(m=0.90)。0.47 个标准差有多大:m=0.75 时两条切口原本在 ±1.15 SD、下侧「判进加重」与上侧「判进改善」各占 12.5%,平移后落到 −0.68 与 1.62,下侧涨到 24.7%、上侧压到 5.3%。读者二那边「加重」的人数直接翻了一倍——这种差异把两张边际表并排一看就露,但 κ_s 报出来是 1.000,一个数都察觉不到。

    二、门在两个方向上同时出错,错法随 m 变

    反解「δ 不超过 5pp」允许的 κ_s 区间(纯尺度差):

    m 0.60 0.75 0.85 0.90 0.93
    允许下限 0.898 0.901 0.890 0.875 0.858
    允许上限 1.110 1.114 1.143 1.192 1.284

    低档位(m≤0.80)与名义 ±0.10 大体吻合,误差 5%–7%,改法看着没问题。但主档一集中,区间就朝两边撕开:m=0.90 到 [0.875, 1.192],m=0.93 到 [0.858, 1.284]。而真实三档裁决的主档占比普遍在 0.8–0.95。

    以 m=0.90 为例,±0.10 同时干两件坏事:真 κ_s=1.15 的读者组,δ 只有 4.2pp、完全合格,却落在门外的「超界」侧;真 κ_s=0.88 的读者组,δ 已到 4.8pp、贴着硬线,反而被判合格。把「合法区」和「名义门」叠加,落在门外却仍合格的宽度占整个合法区:m=0.75 是 6.6%,m=0.85 是 20.7%,m=0.90 是 36.8%,m=0.93 是 53.0%。同一条线,在主档集中时三分之一到一半的合法配置会被误判,而没有任何固定数字能同时修好两边。

    三、「效率高一个数量级」这句话,方向是反的

    上一篇的推理混淆了两件事——κ_s 估得准,和δ 换算得准。中间隔着一个汇率。

    δ 对 κ_s 的斜率在 κ_s=1 处是 22.9–48.3 pp/单位(m=0.70 最高 48.3,m=0.90 是 33.9,m=0.95 是 22.9)。要把 δ 的 95% 置信区间半宽压到 2pp,κ_s 的标准误得压到 0.0215;按 σ 之比的抽样分布反解,每读者需要约 2155 人。而 δ 直接测(假定不一致率 10%)只要约 960 人。贵 2.2 倍,不是便宜十倍。

    只有 m≥0.92 才翻转(m=0.95 时 505 对 960,省一半)。小样本更糟:n=60 时走 κ_s 换回来的 δ 区间半宽是 12.3pp,直测是 8.0pp。

    四、这道门在小样本上自己就过不去

    真 κ_s 严格等于 1.0 时,估计值落进 [0.90, 1.10] 的概率:n=30 只有 40.7%,n=60 是 55.6%,n=120 是 72.4%,n=300 是 91.6%,要到 n≈388 才有 95%。一项 n=60 的研究,哪怕两人尺子一模一样,也有一半以上概率被判「尺度不合格」。

    这和上一篇批判过的「n=60 时 δ 判据永远给出不合格」是同一个病,只换了个参数。实况更不乐观:用 kappa 的一致性研究样本量中位数只有 71(四分位 50–233),而 82 项一致性研究里只有 27 项(33%)给了样本量依据(PMID 36123642)。多数研究正好落在这道门最不可靠的位置,且没人算过自己该有多少人。样本量本就该按目标精度反解(PMID 22560852;改用 ICCa 同理,须按区间宽度反解,PMID 39285135)。

    五、改法:把双向判据改成单向

    κ_s 唯一站得住的是单向:κ_s 超界 ⇒ 必存在尺度差 ⇒ δ 必不为零(量级还得看 m)。反向不成立——κ_s 合格什么都不说明,因为位置差与它无关。

    非要用 κ_s 当前置筛选,两件事必须同做:① 同报位置参数 τ̂(两人切口中点之差,单位=共同潜变量 SD),或干脆直接报 δ̂ 加置信区间;② 样本量按 δ 的目标精度反解后乘 2.2。τ̂ 的估法有个坑:把两人连续打分做回归,斜率给尺度比、截距给位置差,但两人的测量误差都会压低斜率,普通最小二乘会系统性低估尺度差,须用主轴回归或 Deming 回归。

    这套处境在方法学上早有对应:固定偏倚(整体系统差)与比例偏倚(随量级变化的差)必须分开检验,只报一个平均差和一致性界限会把系统差藏起来(PMID 2868172)。连续打分的标准差之比读的正是「比例偏倚」,它天生看不见「固定偏倚」。Kendall 早就指出 kappa 混着偏倚与流行度两个可分离指数,任一单指标都可能漏检另一个维度(PMID 8501467)。高一致率与显著边际异质性可以同时成立,这个反例在真实放射判读数据里已被直接量到(PMID 37127220)。

    证据与局限

    已验证:κ_s 对纯平移精确不敏感(a₁=a₂ 时 σ₁/σ₂ 恒为 1,算术恒等);δ 是尺度与位置的二元函数(潜变量阈值模型解析结果);纯尺度门下 δ=5pp 的临界 κ_s 随 m 外扩,名义 ±0.10 在 m≥0.85 上误判率 20.7%–53.0%(数值反解);真值 1 时 κ_s 的估计量服从 √F 分布,n≈388 才有 95% 落入 ±0.10(F 分布解析);固定偏倚与比例偏倚须分开检验(PMID 2868172 直接论述)。

    部分验证:潜变量阈值模型假设单一潜在维度、两位读者共用同一条 x 轴。真实判读可能混入记忆效应、病例顺序效应,这些既不是纯尺度也不是纯位置,会同时污染两个参数,本模型不给这一项。δ 直测的 n≈960 按不一致率 10% 算,真实三档裁决可能在 5%–30%,对应 n 从 480 到 2881。所有数字都是模型推演,气功外气三档裁决里没有任何实测的 κ_s 或 τ。

    待验证:① 真实三档判读数据上 κ_s 与 δ 的相关系数到底多少——上一篇给的准入线是 r>0.7,本文只证了「纯平移机制下 r 可以接近 0」,没测真实数据。② τ̂ 用主轴回归估计的偏差与稳健性,尚无数值。③ 「锚定样例训练」压 κ_s 时能否同时压住 τ——这是两个自由度,只训一个可能压不住另一个。

    证据等级: C
    参考文献: PMID 2868172, PMID 22560852, PMID 37127220, PMID 39285135, PMID 8501467, PMID 36123642

    相关阅读

    L1 网络状态:CORRESPONDENCE_NETWORK.md 三域节点表中,物理侧仅「相位 / 节律 | — | 周期性耦合的关键」,生理侧九项节点(HRV / 皮温皮电 / 呼吸率 / EMG / EEG / 胃肠 / 气感报告 / 体质分型)无「判读者」「评分者」「连续打分」任何节点;全表检索「判读 / 信度 / 一致率 / 尺度 / 偏倚 / 代理量」零命中;EDGE_TABLE_science-yixue.md 47 个概念中无相关条目;网络对齐指标行原文「L2 以上边数 | 0」。该主题尚未进入 L1 对应网络。


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

    The ratio of standard deviations as a ruler cannot measure when the ruler has shifted: a qualifying reading precisely equal to 1.000 can hide an eight-percentage-point bias behi…

    Evidence level: C

    In the previous post, it was suggested that “the ratio of standard deviations between two raters scoring consecutively” should replace the bias δ, which was never measured. This suggestion has only one flaw, but it is fatal: δ is determined by two degrees of freedom, while this ratio reads only one. The missing degree of freedom can push δ directly to eight points, while keeping this ratio exactly equal to 1.000.

    Main Text

    The previous post attributed the magnitude of bias δ to “whether the scale markings in the minds of two readers share the same origin,” and proposed a modification: have each reader give continuous scores from 0–100 on the magnitude of change for the same batch of people, and examine the ratio of standard deviations κ_s of their scores; a ratio between 0.9 and 1.1 indicates the scales share the same origin, and δ is structurally on the order of 1 part per hundred (pp). The rationale is that the variance estimation efficiency of continuous quantities is far superior to that of proportional differences, allowing precision one order of magnitude higher.

    This modification errs in two ways: it omits one degree of freedom; and the phrase “one order of magnitude more efficient” is reversed in direction.

    I. κ_s is precisely blind to location

    Each reader’s score for person $j$ is $s_j = a_j \cdot x + b_j$, where $x$ represents the true magnitude of change. Thus, a reader’s entire behavior involves only two degrees of freedom: scale $a$ (the length of the ruler) and location $b$ (where the zero point lies). The ratio $\kappa_s = a_1/a_2$ reads only the first.

    Now consider pure translation: $a_1 = a_2 = 1$, $b_1 = 0$, $b_2 = \tau$. In this case, the second reader’s scores are always equal to the first plus a constant; the standard deviations of their scores are identical, so $\kappa_s$ is exactly 1.000—not “approximately,” but an identity. Yet the proportion of main records still diverges:

    τ (units = common latent variable SD) 0.2 0.4 0.6 0.9
    κ_s 1.000 1.000 1.000 1.000
    δ (m=0.75) 0.9pp 3.7pp 8.1pp 17.1pp
    δ (m=0.90) 0.7pp 2.7pp 6.0pp 13.4pp

    (δ is defined as “reader one’s proportion of main records minus reader two”; positive values indicate reader two tends to mark people as having change, while main records are fewer.)

    To push δ to the hard limit of 5pp, pure translation requires $\tau = 0.47$ SD (for $m=0.75$) or $0.55$ SD (for $m=0.90$). What is 0.47 standard deviations? When $m=0.75$, the two cut-offs originally lie at ±1.15 SD, with “判进加重” (判进加重) and “判进改善” (判进改善) each occupying 12.5% on the lower and upper sides respectively; after translation, they fall to −0.68 and 1.62, with the lower side rising to 24.7% and the upper side compressed to 5.3%. The number of people marked “加重” (加重) by reader two doubles directly—such differences become obvious when the two marginal tables are placed side-by-side, yet $\kappa_s$ reports 1.000, revealing nothing.

    II. The gate errs in both directions, with error depending on m

    Solving backwards for the allowable $\kappa_s$ interval under “δ not exceeding 5pp” (pure scale difference):

    m 0.60 0.75 0.85 0.90 0.93
    Allowable lower bound 0.898 0.901 0.890 0.875 0.858
    Allowable upper bound 1.110 1.114 1.143 1.192 1.284

    The low档位 (m ≤ 0.80) roughly matches the nominal ±0.10, with errors of 5%–7%, making the modification appear acceptable. But when main records concentrate, the interval tears apart on both sides: for $m=0.90$ it becomes [0.875, 1.192], and for $m=0.93$ it becomes [0.858, 1.284]. In real three-tier adjudication, the proportion of main records is generally between 0.8 and 0.95.

    Taking $m=0.90$ as an example, the nominal ±0.10 simultaneously does two harmful things: a group with true $\kappa_s=1.15$ has δ only 4.2pp, fully qualified, yet falls on the “out-of-gate” side; a group with true $\kappa_s=0.88$ has δ already at 4.8pp, touching the hard limit, yet is judged qualified. Overlaying the “legal zone” and the “nominal gate,” the width that falls outside yet remains qualified accounts for: $m=0.75$ is 6.6%, $m=0.85$ is 20.7%, $m=0.90$ is 36.8%, and $m=0.93$ is 53.0%. Along the same line, when main records concentrate, one-third to half of all legally valid configurations are misjudged, and no fixed number can simultaneously fix both sides.

    III. “One order of magnitude more efficient” is reversed in direction

    The previous post’s reasoning conflates two matters—that $\kappa_s$ is estimated accurately, and that δ is converted accurately—separated by an exchange rate.

    The slope of δ with respect to $\kappa_s$ at $\kappa_s=1$ ranges from 22.9 to 48.3 pp per unit (highest at $m=0.70$ with 48.3, $m=0.90$ yields 33.9, and $m=0.95$ gives 22.9). To compress the half-width of δ’s 95% confidence interval to 2pp, the standard error of $\kappa_s$ must be reduced to 0.0215; solving backwards via the sampling distribution of the ratio of σs, each reader requires approximately 2155 people. Measuring δ directly (assuming an inconsistency rate of 10%) requires only about 960 people. This is 2.2 times more expensive, not ten times cheaper.

    Only when $m \geq 0.92$ does the direction flip (for $m=0.95$, 505 versus 960, saving half). Small samples fare worse: at $n=60$, the half-width of δ recovered via $\kappa_s$ is 12.3pp, whereas direct measurement yields 8.0pp.

    IV. This gate cannot even pass itself with small samples

    When the true $\kappa_s$ is strictly 1.0, the probability that the estimate falls into [0.90, 1.10] is: $n=30$ gives only 40.7%, $n=60$ yields 55.6%, $n=120$ reaches 72.4%, and $n=300$ achieves 91.6%; it takes $n \approx 388$ to reach 95%. A study with $n=60$, even if both rulers are identical, has over a 50% chance of being judged “scale-不合格”.

    This matches the criticism leveled in the previous post that “at $n=60$, the δ criterion always yields failure,” differing only in parameter. Reality is even less optimistic: the median sample size for kappa consistency studies is 71 (interquartile range 50–233), and among 82 consistency studies, only 27 (33%) provided sample size justification (PMID 36123642). Most studies fall precisely where this gate is most unreliable, and no one has calculated how many subjects they should have. Sample size ought to be derived backwards from target precision (PMID 22560852; similarly for ICCa, it must be solved backwards via interval width, PMID 39285135).

    V. Modification: Convert the bidirectional criterion into unidirectional

    The only robust aspect of $\kappa_s$ is unidirectional: $\kappa_s$ out-of-bounds ⇒ there must be a scale difference ⇒ δ cannot be zero (the magnitude still depends on $m$). The reverse does not hold—$\kappa_s$ being qualified tells us nothing, because location difference is unrelated to it.

    If we insist on using $\kappa_s$ as a pre-screen, two things must be done simultaneously: ① report the location parameter $\hat{\tau}$ (the difference between midpoints of the two cut-offs, units = common latent variable SD), or simply report $\hat{\delta}$ with its confidence interval; ② derive sample size by multiplying the target precision for δ by 2.2. Estimating $\hat{\tau}$ has a pitfall: regressing the two continuous scores gives slope as scale ratio and intercept as location difference, but measurement errors from both readers depress the slope; ordinary least squares systematically underestimates scale difference, requiring principal axis regression or Deming regression.

    This situation has long counterparts in methodology: fixed bias (overall systematic difference) and proportional bias (difference varying with magnitude) must be tested separately; reporting only an average difference and consistency limits hides systematic bias (PMID 2868172). The ratio of standard deviations of continuous scores reads precisely “proportional bias,” which inherently cannot see “fixed bias.” Kendall long ago pointed out that kappa mixes two separable indices—bias and prevalence—and any single indicator may miss the other dimension (PMID 8501467). High consistency rates and significant marginal heterogeneity can coexist, a counterexample directly measured in real radiological reading data (PMID 37127220).

    Evidence and Limitations

    Verified: $\kappa_s$ is precisely insensitive to pure translation (when $a_1=a_2$, $\sigma_1/\sigma_2$ is always 1, an arithmetic identity); δ is a binary function of scale and location (analytical result from latent variable threshold model); under pure scale gates, the critical $\kappa_s$ for $\delta=5pp$ expands outward with $m$, and the nominal ±0.10 misjudgment rate reaches 20.7%–53.0% for $m \geq 0.85$ (numerical inversion); when the true value is 1, the estimator of $\kappa_s$ follows a $\sqrt{F}$ distribution, requiring $n \approx 388$ for 95% to fall within ±0.10 (analytical F distribution); fixed bias and proportional bias must be tested separately (PMID 2868172 explicitly discusses this).

    Partially Verified: The latent variable threshold model assumes a single latent dimension and that two readers share the same x-axis. Real adjudication may mix memory effects, case-order effects, which are neither pure scale nor pure location, contaminating both parameters simultaneously; this model does not include such a term. Direct measurement of δ requires $n \approx 960$ assuming an inconsistency rate of 10%, while real three-tier adjudication may range from 5%–30%, corresponding to $n$ from 480 to 2881. All numbers are model-derived; there is no empirical $\kappa_s$ or $\tau$ in the three-tier adjudication of qigong external qi.

    To Be Verified: ① What is the actual correlation coefficient between $\kappa_s$ and δ on real three-tier adjudication data—the准入线 (admission threshold) given in the previous post is $r > 0.7$, but this paper only proves that “under pure translation mechanism, $r$ can approach 0,” without measuring real data. ② The bias and robustness of $\hat{\tau}$ estimated via principal axis regression, lacking numerical values. ③ Whether “anchored sample training” suppressing $\kappa_s$ can simultaneously suppress $\hat{\tau}$—these are two degrees of freedom; training only one may not constrain the other.

    Evidence Level: C
    References: PMID 2868172, PMID 22560852, PMID 37127220, PMID 39285135, PMID 8501467, PMID 36123642

    Further Reading

    L1 Network Status: In the three-domain node table of CORRESPONDENCE_NETWORK.md, the physical side contains only “phase/rhythm | — | key of periodic coupling”; the physiological side has nine nodes (HRV / skin temperature/skin conductance / breathing rate / EMG / EEG / gastrointestinal / qi sensation report / constitution typing) but no nodes for “reader/rater/continuous scoring”; searching the entire table for “reading/reliability/consistency/scale/bias/proxy quantity” yields zero hits; EDGE_TABLE_science-yixue.md has 47 concepts but no relevant entries; the network alignment indicator row reads “L2 and above edge count | 0”. This topic has not yet entered the L1 corresponding network.


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

  • 这个比值判据的两头都是盲区:要测准 σ 得靠高阶,可高阶早就测不到了

    Evidence level: C

    上一节把整张 S_k 表的命门压在一个从未测过的比值 r = σ_脉搏 / σ_DLMO 上,并给了它一套三档判据。算下来,这套判据在它最该生效的两端(r 远大于 1、r 远小于 1),恰好都是数据给不出答案的地方——而要测准 σ 需要的那个阶次,脉搏侧早就衰减到噪声以下了。

    正文

    一、判据要的是什么

    上一篇把节律校正的取舍压在一个数上:个体相位偏移 σ。而脉搏这条线自己没有 σ,是从褪黑素起始时间(DLMO)借来的。借用合法与否,取决于两者相位离散是否同量级(1.43 小时),于是有了这个比值

    r = σ_脉搏 / σ_DLMO
    

    原判据三档:r 落在 0.7–1.4 → 沿用 DLMO 的 σ,整张表照搬;r > 2 → 中枢与外周解耦,脉搏侧必须自测,绝对数值全部作废;r < 0.5 → 外周被中枢强锁相,个体化校正的危害比原估计小一个量级,紧迫性降级。

    这三档听起来很稳。问题在于,r 从来不是直接测出来的——它是两个 σ 各自反解之后再相除的比值。所以「r 测得准不准」这个问题,要先拆成「σ 测得准不准」。

    二、σ 的精度随阶次单调变好,快得离谱

    由 S_k = exp(−(kωσ)²) 反解:

    σ = √(−ln S_k) / (kω)
    

    误差传递给出 σ 的相对误差 ≈ 0.5 · |Δln S| / |ln S_k|。关键在于分母 |ln S_k| 随阶次暴涨。取 σ = 1.43 小时、S_k 的组间抽样误差 0.05:

    阶次 k S_k |ln S_k| σ 相对误差
    1 0.8692 0.140 17.8%
    2 0.5709 0.561 4.5%
    3 0.2833 1.261 2.0%
    4 0.1062 2.242 1.1%

    一阶几乎废掉(17.8% 的误差,判不了 0.7 与 1.4 之间的任何区别),三阶就有 2.0%。想要 σ,就得往高阶走。

    三、但脉搏侧往高阶走是走不通的

    脉搏谐波的 S_k 是同一个式子,只是把 σ 换成 σ_脉搏 = r·σ_DLMO。r 越大,衰减越快。要让一个阶次可用,S_k 至少得在 0.02 以上(再低就埋进测量噪声)。反解这个门槛:

    S_k > 0.02  ⇒  r < √(−ln 0.02) / (kωσ_DLMO)
    

    代入 σ_DLMO = 1.43:

    阶次 k 可测的 r 上界 σ 反解误差
    1 5.28 17.8%
    2 2.64 4.5%
    3 1.76 2.0%
    4 1.32 1.1%

    对照一下:精度够用的阶次(k≥3,误差 ≤2%)只能测到 r < 1.76;能测到 r > 2 的阶次,只剩 k=2 这一根独苗。

    四、剪刀差:判「解耦」时,你没有任何自检余地

    这就是本篇要说的第一件事。判据里最重的那一档是「r > 2 → 必须自测」。而 r 一旦超过 1.76,k=3 及以上的 S_k 全部低于 0.02,等于全灭,只剩 k=2 一阶可用。

    单阶可用意味着什么?意味着你失去了唯一的一种内部检查手段:多阶一致性。正常情况下,k=2 与 k=3 各自反解出的 σ 应该吻合;若不吻合,说明模型(正态相位分布)不成立、或者某一阶的测量有问题。可当 r > 1.76 时,k=3 已经测不到了——你恰恰在最需要证据的那个区间,只剩一个数、一套口径、零余量做交叉验证。

    两条曲线在这里迎面撞上:σ 的精度随 k 上升,可测性随 k 下降。它们的交点落在 r ≈ 1.76,而判据的分档线(1.4 与 2.0)正好跨在交点两侧。判据最想区分的那一段,正好是测量能力最薄的那一段。

    五、另一头更糟:r < 0.5 这一支,信息量直接枯竭

    原判据的降级支是 r < 0.5。直觉上「σ 小」应该好测——离散小、S_k 接近 1,多干净。算下来正好相反。

    σ_pulse = 0.715 小时(r = 0.5)时:S₁ = 0.9830,S₂ = 0.9338,S₃ = 0.8571。全部挤在 1 附近,|ln S| 极小,于是相对误差爆开:

    σ_pulse S₁ S₂ S₃ k=1 误差 k=2 误差 k=3 误差
    1.43 h 0.8692 0.5709 0.2833 17.8% 4.5% 2.0%
    1.00 h 0.9338 0.7602 0.5396 36.5% 9.1% 4.1%
    0.72 h 0.9651 0.8675 0.7263 70.4% 17.6% 7.8%
    0.50 h 0.9830 0.9338 0.8571 145.9% 36.5% 16.2%

    r = 0.5 时,一阶的 σ 反解误差 145.9%——这个数已经完全无意义了;二阶 36.5%,同样判不了 0.5 与 0.7 的差别。

    再直接算:要区分 r = 0.5 与 r = 0.7,σ = 0.715 与 1.001 两点的 S_k 差多少?

    k S(0.715) S(1.001) 差 可分?
    1 0.9656 0.9336 0.0319 不可分(<0.05)
    2 0.8692 0.7598 0.1094 可分
    3 0.7295 0.5390 0.1906 可分

    一阶不可分。也就是说,判「降级」这件事,比判「自测」更难——它需要的相对精度更高,而它手上的信号更弱。

    六、判据本身还有 0.80 宽的洞

    三档划分留了两段没有归属:r ∈ [0.5, 0.7](宽 0.20)与 r ∈ [1.4, 2.0](宽 0.60),合计 0.80。这段空白占到 r ∈ [0, 4] 的 20%;若只看生理上可能的 [0.3, 3.0],占 29.6%。

    判断这两个洞是不是「真洞」,要跟测量分辨率比。单阶(k=2,误差 4.5%)反解出的 r 的 95% 置信区间半宽约 0.22。

    • [0.5, 0.7] 宽 0.20 < 2×0.22 = 0.44 → 这个洞比尺子的最小刻度还窄,等于不存在
    • [1.4, 2.0] 宽 0.60 > 0.44 → 这是真洞,落进去的样本无法归类

    蒙特卡洛确认了这一点(se = 0.22):

    真 r 判「降级」 判「放行」 判「未定义」 判「自测」
    0.5 50% 18% 32% —
    0.6 33% 32% 35% —
    0.7 18% 50% 32% —
    1.0 1% 89% 10% —
    1.4 — 50% 50% 0%
    1.7 — 8% 83% 9%
    2.0 — 0% 50% 50%

    真 r = 0.7(判据的下界),只有一半被判放行,18% 被判降级——方向反了。真 r = 2.0(自测门槛),一半判未定义、一半判自测。真 r = 1.7,83% 判未定义,这一档实际上不可判。

    七、所以该报的不是 r

    结论是:这条判据不该以「r 落在哪个区间」的形式使用。改法有三条:

    第一,报两个 σ 各自的区间,不要报比值点值。 r 是相除得来的,分母的误差会传染给整个判据;单独报 σ_DLMO = 1.43 (95%CI 1.31–1.61) 与 σ_脉搏 = 1.20 (95%CI 1.05–1.38),读者可以自己看两个区间重不重叠。

    第二,判据改成「比值区间是否完全落在某一档内」。 若 [σ_脉搏/σ_DLMO] 的区间跨越了分档线,就该写「不可判」,而不是取中点硬分一档。按上表,r = 0.6 与 r = 1.7 这两段基本都该写「不可判」。

    第三,r < 1.76 时必须用 k=2 与 k=3 双阶交叉验证。 两阶反解出的 σ 若差超过 15%,说明正态相位模型不成立,此时 S_k 表整体失效——这是唯一能在数据内部抓出模型错误的检查,成本很低,不能省。

    一句话概括:这个判据问的是「中枢和外周的抖动一样大吗」,而它的答案精度取决于高阶谐波,高阶谐波又只有在中枢和外周抖动差不多的时候才测得到。 它在你最想知道答案的地方最沉默。

    证据与局限

    已验证:S_k = exp(−(kωσ)²) 的相位去相干推导与十万次蒙特卡洛一致(前作已验,偏差 <0.003);σ 反解的误差传递式 0.5·|Δln S|/|ln S_k| 由一阶泰勒展开直接得出,与 3000 次蒙特卡洛的 CI 吻合;脉搏谐波前十一阶的重测信度 ICC 在 0.83–0.96(PMID 27170904,22 人,TD01C 系统),说明「逐阶 S_k 可测」这件事在仪器层面有基础;DLMO 相位个体差异大、组内 ICC 0.67–0.85(PMID 34396981、PMID 28452285);晨型与晚型的外周钟基因表达相位差 2.6–3.1 小时(PMID 27103930),是 r 可能明显偏离 1 的生理依据。

    部分验证:双峰表型在人群中的占比有实测(16%,PMID 28426387,EPISONO 队列 857 人),但那是问卷意义上的「答案在晨型与晚型之间摇摆」,并非相位分布的形态学证据,不能直接当作「相位偏移呈双峰」的替代。

    待验证:① r 的真实值从未测过,本文所有关于「判据在两端失效」的推论都建立在「r 可能落在 0.7–2.6 各处」这个前提上,若实测 r 恰好接近 1,本篇的批评虽仍成立但实际影响有限;② 本文的「可测门槛 S_k > 0.02」是拍的,真实门槛取决于脉搏谐波振幅信噪比与记录时长,未做过灵敏度分析;③ 全部推导假定相位偏移服从单峰正态,若为双峰,则 S_k 在高阶会出现非单调回升,本文的「高阶不可测」结论需要重写(另题处理);④ 未讨论 σ 随阶次本身变化的情形(相位谱非线性),那会直接推翻单参数模型。

    L1 网络状态:CORRESPONDENCE_NETWORK.md 三域节点表中,物理侧仅列「相位 / 节律 | — | 周期性耦合的关键」,语义为两个振荡之间的锁相,不是日内生物钟相位的个体间离散度;生理侧九项节点(HRV / 皮温皮电 / 呼吸率 / EMG / EEG / 胃肠 / 气感 / 体质)无「节律相位」「昼夜时相」任何节点;时间层只有「春 夏 长夏 秋 冬|五季」(年尺度)与「子丑寅卯…亥|十二地支」(历法编码),粒度止于日、无日内钟点层。边表止于 E-050(纳音↔频率,L1、open,现代证据栏原文「尚无任何可测来源」);网络对齐指标行原文「L2 以上边数 | 0」。EDGE_TABLE_science-yixue.md 47 个概念中无「节律」「昼夜」「生物钟」条目,最接近的 时辰(42 篇 / A 级 9)指时辰与临床风险的关联,不承载相位离散结构。→ 该主题尚未进入 L1 对应网络,本文封顶 C。

    相关阅读

    证据等级: C
    参考文献: PMID 27170904, PMID 34396981, PMID 28452285, PMID 27103930, PMID 28426387


    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。

    Blindness at both ends of this ratio criterion: measuring σ accurately requires higher orders, yet those higher orders are already unmeasurable.

    Evidence level: C

    The previous section pinned the critical point of the entire $S_k$ table on a ratio $r = \sigma_{\text{pulse}} / \sigma_{\text{DLMO}}$ that has never been measured, and provided a three-tier criterion for it. Calculated out, this criterion fails precisely at both ends where it should be most effective ($r \gg 1$ and $r \ll 1$)—the very places where the data yields no answer. Moreover, the harmonic order required to measure $\sigma$ accurately has already decayed below the noise level on the pulse side.

    Main Text

    I. What the Criterion Requires

    The previous article pinned the choice of rhythm correction on a single number: individual phase shift $\sigma$. However, the pulse line itself lacks $\sigma$ and borrows it from the melatonin onset time (DLMO). Whether this borrowing is legitimate depends on whether the phase dispersion of both variables is of the same order (1.43 hours), leading to this ratio:

    r = σ_pulse / σ_DLMO
    

    The original criterion has three tiers: $r$ falls in 0.7–1.4 → use the $\sigma$ of DLMO and copy the whole table; $r > 2$ → central and peripheral systems decouple, the pulse side must self-test, and all absolute values are void; $r < 0.5$ → the periphery is strongly phase-locked by the central system, the danger of individualized correction is one order of magnitude smaller than originally estimated, and urgency is downgraded.

    These three tiers sound stable. The problem is that $r$ is never directly measured—it is the ratio of two $\sigma$s each inverted separately. Therefore, “how accurate is $r$?” must first be broken down into “how accurate is $\sigma$?”.

    II. Accuracy of $\sigma$ Improves Monotonically with Order, Faster Than Expected

    From $S_k = \exp(-(k\omega\sigma)^2)$, we invert:

    σ = √(−ln S_k) / (kω)
    

    Error propagation gives the relative error of $\sigma$ as approximately $0.5 \cdot |\Delta\ln S| / |\ln S_k|$. The key is that the denominator $|\ln S_k|$ explodes with harmonic order. Taking $\sigma = 1.43$ hours and the between-group sampling error of $S_k$ as 0.05:

    Order k S_k |ln S_k| σ relative error
    1 0.8692 0.140 17.8%
    2 0.5709 0.561 4.5%
    3 0.2833 1.261 2.0%
    4 0.1062 2.242 1.1%

    The first order is almost useless (17.8% error, cannot distinguish between 0.7 and 1.4), while the third order has only 2.0%. To obtain $\sigma$, one must go to higher orders.

    III. But Going to Higher Orders on the Pulse Side Is Impossible

    The $S_k$ of pulse harmonics follows the same formula, except $\sigma$ is replaced by $\sigma_{\text{pulse}} = r \cdot \sigma_{\text{DLMO}}$. As $r$ increases, decay accelerates. For a harmonic order to be usable, $S_k$ must be at least 0.02 (otherwise it is buried in measurement noise). Inverting this threshold:

    S_k > 0.02  ⇒  r < √(−ln 0.02) / (kωσ_DLMO)
    

    Substituting $\sigma_{\text{DLMO}} = 1.43$:

    Order k Upper bound of measurable r σ inversion error
    1 5.28 17.8%
    2 2.64 4.5%
    3 1.76 2.0%
    4 1.32 1.1%

    Comparing these: orders with sufficient accuracy ($k \geq 3$, error $\leq 2\%$) can only measure $r < 1.76$; orders capable of measuring $r > 2$ are limited to just the single $k=2$ case.

    IV. The Scissors Gap: No Self-Check When Judging “Decoupling”

    This is the first point of this article. The heaviest tier in the criterion is “$r > 2$ → must self-test.” Once $r$ exceeds 1.76, all $S_k$ for $k \geq 3$ drop below 0.02, effectively vanishing, leaving only the $k=2$ order available.

    What does having a single order mean? It means you lose the only internal check: multi-order consistency. Normally, $\sigma$ inverted from $k=2$ and $k=3$ should agree; if they do not, either the model (normal phase distribution) fails or one order has measurement issues. But when $r > 1.76$, $k=3$ is no longer measurable—you are precisely in the interval needing evidence, yet you have only one number, one scale, and zero margin for cross-validation.

    Two curves collide here: accuracy of $\sigma$ improves with $k$, while measurability declines with $k$. Their intersection lies at $r \approx 1.76$, while the criterion’s tier boundaries (1.4 and 2.0) straddle this intersection. The segment the criterion most wants to distinguish is exactly where measurement capability is weakest.

    V. The Other End Is Worse: The $r < 0.5$ Branch Has Directly Dried Information

    The downgrade branch of the original criterion is $r < 0.5$. Intuitively, “small $\sigma$” should be easy to measure—low dispersion, $S_k$ close to 1, very clean. Calculated out, it is exactly the opposite.

    When $\sigma_{\text{pulse}} = 0.715$ hours ($r = 0.5$): $S_1 = 0.9830$, $S_2 = 0.9338$, $S_3 = 0.8571$. All crowd near 1, making $|\ln S|$ extremely small, thus blowing up the relative error:

    σ_pulse S₁ S₂ S₃ k=1 error k=2 error k=3 error
    1.43 h 0.8692 0.5709 0.2833 17.8% 4.5% 2.0%
    1.00 h 0.9338 0.7602 0.5396 36.5% 9.1% 4.1%
    0.72 h 0.9651 0.8675 0.7263 70.4% 17.6% 7.8%
    0.50 h 0.9830 0.9338 0.8571 145.9% 36.5% 16.2%

    At $r = 0.5$, the first-order $\sigma$ inversion error is 145.9%—this number is completely meaningless; the second-order error is 36.5%, still unable to distinguish between 0.5 and 0.7.

    Directly calculating: to distinguish $r = 0.5$ from $r = 0.7$, i.e., $\sigma = 0.715$ vs $1.001$, how much do their $S_k$ values differ?

    k S(0.715) S(1.001) Difference Distinguishable?
    1 0.9656 0.9336 0.0319 No (<0.05)
    2 0.8692 0.7598 0.1094 Yes
    3 0.7295 0.5390 0.1906 Yes

    The first order is indistinguishable. That means judging “downgrade” is harder than judging “self-test”—it requires higher relative precision, yet the signal it has is weaker.

    VI. The Criterion Itself Has a 0.80-Wide Hole

    The three-tier division leaves two segments unassigned: $r \in [0.5, 0.7]$ (width 0.20) and $r \in [1.4, 2.0]$ (width 0.60), totaling 0.80. This blank occupies 20% of $r \in [0, 4]$; if we consider only physiologically possible $[0.3, 3.0]$, it accounts for 29.6%.

    To judge whether these two holes are “real holes,” we must compare them with measurement resolution. For the single order ($k=2$, error 4.5%), the 95% confidence interval half-width of inverted $r$ is approximately 0.22.

    • $[0.5, 0.7]$ width 0.20 < $2 \times 0.22 = 0.44$ → this hole is narrower than the smallest scale of the ruler, effectively non-existent
    • $[1.4, 2.0]$ width 0.60 > 0.44 → this is a real hole; samples falling into it cannot be classified

    Monte Carlo simulations confirm this ($se = 0.22$):

    True r Judge “downgrade” Judge “release” Judge “undefined” Judge “self-test”
    0.5 50% 18% 32% —
    0.6 33% 32% 35% —
    0.7 18% 50% 32% —
    1.0 1% 89% 10% —
    1.4 — 50% 50% 0%
    1.7 — 8% 83% 9%
    2.0 — 0% 50% 50%

    True $r = 0.7$ (lower bound of the criterion) is judged “release” only half the time, and “downgrade” 18%—the direction is reversed. True $r = 2.0$ (self-test threshold) is half judged “undefined” and half “self-test.” True $r = 1.7$, 83% are judged “undefined”—this tier is actually unjudgeable.

    VII. Therefore, We Should Not Report $r$

    The conclusion is: this criterion should not be used in the form of “$r$ falls into which interval.” Three modifications are proposed:

    First, report the intervals of both $\sigma$s separately, not the ratio point value. Since $r$ is obtained by division, errors in the denominator propagate to the entire criterion; separately reporting $\sigma_{\text{DLMO}} = 1.43$ (95% CI 1.31–1.61) and $\sigma_{\text{pulse}} = 1.20$ (95% CI 1.05–1.38) allows readers to check whether the two intervals overlap.

    Second, change the criterion to “whether the ratio interval completely falls within one tier.” If the interval of $[\sigma_{\text{pulse}}/\sigma_{\text{DLMO}}]$ crosses a tier boundary, it should be written as “unjudgeable,” rather than taking the midpoint to force a tier. According to the table above, $r = 0.6$ and $r = 1.7$ should both be written as “unjudgeable.”

    Third, when $r < 1.76$, must use dual-order cross-validation with $k=2$ and $k=3$. If the $\sigma$ inverted from two orders differs by more than 15%, it indicates the normal phase model fails, and the entire $S_k$ table becomes invalid—this is the only check that can catch model errors within the data itself, at low cost, and cannot be skipped.

    In one sentence: This criterion asks “Do the central and peripheral systems jitter equally?”, but its answer precision depends on higher-order harmonics, which are only measurable when the central and peripheral systems jitter similarly. It is most silent exactly where you want the answer most.

    Evidence and Limitations

    Verified: The phase decoherence derivation of $S_k = \exp(-(k\omega\sigma)^2)$ and the Monte Carlo simulation of 100,000 runs are consistent (previous work verified, deviation <0.003); the error propagation formula for $\sigma$ inversion $0.5 \cdot |\Delta\ln S| / |\ln S_k|$ is directly obtained from first-order Taylor expansion and matches the CI of 3,000 Monte Carlo runs; the retest reliability ICC of the first eleven pulse harmonics is 0.83–0.96 (PMID 27170904, 22 subjects, TD01C system), indicating that “measurability of each-order $S_k$” has an instrumental basis; DLMO phase individual differences are large, with within-group ICC 0.67–0.85 (PMID 34396981, PMID 28452285); the phase difference of peripheral clock gene expression between morning-type and evening-type is 2.6–3.1 hours (PMID 27103930), providing physiological basis for $r$ possibly deviating significantly from 1.

    Partially verified: The proportion of bimodal phenotype in the population has been measured (16%, PMID 28426387, EPISONO cohort of 857 subjects), but that refers to the questionnaire sense of “answer oscillating between morning-type and evening-type,” not morphological evidence of phase distribution, and cannot directly serve as a substitute for “phase shift exhibiting bimodal.”

    To be verified: ① The true value of $r$ has never been measured; all inferences in this article about “the criterion failing at both ends” are based on the premise that $r$ may fall anywhere in 0.7–2.6; if the measured $r$ happens to be close to 1, the criticism in this article still holds but its practical impact is limited; ② The “measurability threshold $S_k > 0.02$” in this article is estimated; the true threshold depends on the signal-to-noise ratio of pulse harmonic amplitude and recording duration, and no sensitivity analysis has been performed; ③ All derivations assume phase shift follows a unimodal normal distribution; if it is bimodal, $S_k$ would show non-monotonic rebound at higher orders, and the conclusion “higher orders are unmeasurable” needs rewriting (to be handled separately); ④ The case where $\sigma$ itself changes with order (nonlinear phase spectrum) has not been discussed, which would directly overturn the single-parameter model.

    L1 Network Status: In the CORRESPONDENCE_NETWORK.md three-domain node table, the physical side lists only “phase / rhythm | — | key to periodic coupling,” semantically meaning phase-locking between two oscillators, not inter-individual phase dispersion of the circadian clock; the nine physiological side nodes (HRV / skin temperature and conductance / breathing rate / EMG / EEG / gastrointestinal / qi sensation / constitution) have no nodes for “rhythm phase” or “circadian phase”; the time layer only includes “Spring Summer Long-Summer Autumn Winter | Five Seasons” (annual scale) and “Zi Chou Yin Mao… Hai | Twelve Earthly Branches” (calendar encoding), with granularity stopping at day, lacking intra-day clock-hour layer. The edge table stops at E-050 (Na Yin ↔ Frequency, L1, open, original text of modern evidence column states “no measurable source yet”); the network alignment indicator row originally states “L2 and above edge count | 0”. EDGE_TABLE_science-yixue.md among 47 concepts has no “rhythm,” “circadian,” or “biological clock” entries; the closest 时辰 (42 articles / A-level 9) refers to the association between time periods and clinical risk, not carrying phase dispersion structure. → This topic has not yet entered the L1 corresponding network; this article is capped at C.

    Further Reading

    Evidence Level: C
    References: PMID 27170904, PMID 34396981, PMID 28452285, PMID 27103930, PMID 28426387


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

  • 「两人把剩余质量各分一半」这个假设其实不承重:三档总变差距离有闭式解,它只在一条边界外才偏离 δ

    Evidence level: C

    上一篇担心等分假设一破、天花板就被压得更低。把它算成闭式之后,结论要反过来写:那个假设从来不是承重件,真正控制一切的是一个比值——主档差有多大,相对于主档外面还剩多少空间。

    正文

    上一篇把整条天花板推导压在一个参数化上:两位判读者把主档(比如「无变化」)之外的质量各自等分给另外两档。等分带来一个漂亮的巧合——两条边际的总变差距离恰好等于主档占比差 δ,于是上限 1−δ 干净得能背下来。上一篇担心:真实判读里可能一位偏乐观、一位偏保守,次档分配不同向,等分一破,TVD 就大于 δ,上限被压得更低。

    这个担心的方向对,但幅度算错了——而且错在「等分假设很脆弱」这个前提本身。

    先把闭式解摆出来。三档裁决下,两条边际分布 p、q 的总变差距离有一个恒等式:TVD = max{|p₁−q₁|, |p₂−q₂|, |p₃−q₃|}。因为三档概率差相加为零,必有一档的差恰等于另两档之差的和,所以半 L1 范数就等于最大的那条边际差。这与参数化无关,是纯代数,随机取三百万组边际验证误差在 10⁻¹⁶ 量级。

    第一个推论就解掉了上一篇一半的担心:TVD ≥ |p主−q主| = δ,恒成立。所以 p₀ 的上限 1−δ 从来不会被违反,等分假设根本不是这个上界的承重件。

    第二个推论更硬:决定「TVD 会不会超过 δ」的,是主档那一档的差是不是三条差里最大的那条。等分假设只影响另两档,而主档差 δ 已经把地板焊死了。

    再看方向性。同向错位(两人都把剩余的更多分给「改善」)完全无害:两人的偏移互相抵消,TVD 恒等于 δ,一分不动——枚举扫过 ε 从 0 到 0.4,TVD 一直是 0.0500。只有反向错位(一位偏乐观、一位偏保守)才有代价,而且有死区。设两人把剩余质量分给第二档的比例分别是 0.5+ε 与 0.5−ε,两侧剩余质量 s_A、s_B,可解出临界点:

    ε* = δ / (2(s_A+s_B))

    当 ε ≤ ε*,TVD 精确等于 δ,错位被主档差完全盖住;越过后 TVD 线性上升,斜率就是 s_A+s_B。

    对称情形 m_A = m̄+δ/2、m_B = m̄−δ/2 时,ε* = δ / (4(1−m̄))。这个式子很反直觉:主档越集中,死区越宽,等分假设越安全。m̄=0.90、δ=0.05 时 ε*=0.125——两位读者对剩余质量的分配要相差 25 个百分点(一人 62.5%/37.5%,另一人 37.5%/62.5%),TVD 才开始动。而 m̄=0.70、δ=0.05 时 ε*=0.042,死区只剩 8.3%。

    唯一的控制器是比值 r = δ/(1−m̄):主档差相对于主档外面剩余空间有多大。死区边界就是 ε* = r/4,死区占「最大可能错位 0.5」的比例则是 r/2。r=0.1 时只覆盖 5%(ε*=0.025),r=0.2 覆盖 10%,r=0.5 覆盖 25%,r=1 覆盖 50%,r=2 覆盖 100%(此时 ε*=0.5,任何偏离都安全)。整条推导只剩一个无量纲数。

    越过死区之后,TVD 的形状也是线性的:令 t = ε/ε*,则 TVD/δ = (1+t)/2(t>1)。也就是说 TVD 相对 δ 的超出量只由 t 决定,与 δ 本身、与 m̄ 本身都无关——m̄ 与 δ 的全部作用都被吸收进 ε* 这一个尺度里。t=2 时 TVD 是 δ 的 1.5 倍,t=4 时是 2.5 倍。相应地 κ_max 的跌落:m̄=0.90、δ=0.05 下 t=1 时 0.731,t=2 时 0.599(跌 18.1%),t=3 时 0.470(跌 35.8%),t 到 4 顶到 ε=0.5 的理论极限,只剩 0.344(跌 52.9%)。注意这个「跌 18%」正是上一篇举出的那个数字——但它对应的是 t=2、即 ε 恰好是死区边界的两倍,不是「ε 稍大一点」。

    回到上一篇的具体担心:ε 的典型量级若真是 0.02–0.05,在 m̄=0.90、δ=0.05 的死区(ε*=0.125)里被整个包住,t 只有 0.16–0.4,TVD 仍精确等于 δ,映射照用不误。要造出上一篇举的那个例子(A=(0.925,0.05,0.025)、B=(0.875,0.025,0.10)),得让 A 把剩余 0.075 里的 0.05 全押给第二档、B 把剩余 0.125 里的 0.10 押给第三档——两人的偏离量分别是 0.167 与 0.30,且方向相反。折算到对称参数族,这个例子等价于 t=2(TVD/δ=1.5),即 ε 恰好是死区边界的两倍——那是两位读者把剩余质量几乎反向推向两个极端,不是「有点风格差异」。

    所以上一篇留下的四条判据需要整体改写。原判据把「ε<0.01」当作安全线、把「ε>δ 就说明主档差作代理失效」当作失效线,两条线的量纲都不对:ε 与 δ 不可直接比,必须除上剩余空间。正确形式是:r = δ/(1−m̄) ≤ 0.5 时(死区已覆盖 25% 的错位空间)等分假设可作工程近似;r 升高、ε 越过 r/4 才需要改用 TVD 报「两边际总变差距离」;只有当 ε 越过死区且「两人次档偏向确实系统相反」时,才谈得上「判的方向不同」而非「判得不准」。

    最后是报告口径。既然 κ_max = 1 − TVD/(1−Σp_k q_k),而 TVD 不是 δ 的单调函数,「偏倚」这个词就该换。实测固定 (m̄, TVD) 后扫遍不同边际形状,κ_max 的跨度只有 0.0015–0.0078,远小于 0.05 的带宽——说明 (m̄, TVD) 这个二元组几乎是充分的,可替代单一 δ 作报告项。规范上不妨直接要求同报 m̄ 与 TVD,δ 退为参考。

    这里有个容易读反的地方:κ_max 对 TVD 是单调下降的(p_o^max = 1−TVD 是恒等式,与边际形状无关),所以真正决定天花板的从头到尾都是 TVD,δ 只是 TVD 的一个下界和一个常用代理。等分假设的全部作用,是让这个下界变成等号。换句话说,上一篇问「δ 能不能当 TVD 用」,答案是:δ 永远是 TVD 的下界,等号条件只需 ε ≤ ε*,而这个条件在高集中区(m̄≥0.9)宽到几乎必然满足。

    一句话收口:等分假设不承重,δ→上限的方向永远成立;数值上只需要一个比值 r = δ/(1−m̄),r ≤ 0.5 时放心用 δ,超出后按 TVD/δ = (1+t)/2 换算,t = ε/(r/4)。

    证据与局限

    已验证:三档下 TVD = 三条边际差绝对值的最大值,这是纯代数恒等式(随机三百万组边际,误差 10⁻¹⁶);TVD ≥ δ 恒成立,故 1−δ 上界无条件有效;同向错位下 TVD 精确等于 δ(ε 扫 0–0.4 不变);反向错位的临界点 ε* = δ/(2(s_A+s_B)) 与二分解实测逐点吻合;闭式 TVD = max(δ, δ/2+ε(s_A+s_B)) 全参数域误差 10⁻¹⁶;κ_max = 1 − TVD/(1−p_e) 由恒等式 p_o^max = Σmin(p,q) = 1−TVD 直接得出。

    部分验证:把 δ 换成 (m̄, TVD) 二元组的充分性只是数值上「近似」成立(固定 (m̄,TVD) 后 κ_max 跨度 0.0015–0.0078),没有解析证明残差上界;死区宽度依赖对称参数化 m_A = m̄±δ/2,真实研究里两位读者的主档占比未必对称,非对称情形 ε* 会移动(主档差越大、动作越晚,方向不变)。文献侧只印证了「边际是 kappa 的独立自由度」这一大类结论(PMID 8501467 把 κ 分解成偏倚、流行度、观察一致三项;PMID 8621993 指出 κ 的上下限完全由边际决定;PMID 26095449 指出 κ 类扩展系数「prone to prevalence and marginal distribution issues」,并改用有序 GLMM 模型化路径;PMID 41230191 同样指出加权 κ 类系数对边际分布与类别流行度敏感),没有任何文献直接给过本文这条三档 TVD 闭式与死区。

    待验证:真实三档裁决数据里 ε 的典型量级到底是多少——本文只能给出「若 ε≤0.05 则安全」的判别,没有实测分布;δ/(1−m̄) 在已发表研究中的实际落点未知;反向错位是否与读者资历、判读顺序、是否有回看机制相关,完全没查。

    相关阅读

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

    证据等级: C
    参考文献: PMID 8501467, PMID 8621993, PMID 2348207, PMID 25168681, PMID 26095449, PMID 41230191, PMID 37127220


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

    The assumption that \”two people each take half of the remaining mass\” does not bear weight: there is a closed-form solution for the total variation distance across three tiers,…

    Evidence level: C

    In the previous discussion, there was concern that breaking the equal-split assumption would depress the ceiling further. By deriving a closed-form expression, the conclusion must be inverted: that assumption never bears structural weight; what truly controls everything is a ratio—the magnitude of the primary class difference relative to the remaining space outside the primary class.

    Body

    The previous derivation compressed the entire ceiling argument into a single parameterization: two raters split the mass equally between the other two classes, excluding the primary class (e.g., “No Change”). This equal-split assumption yields a convenient coincidence—the sum of the total variation distances at the two margins exactly equals the primary class proportion difference δ, making the upper bound 1−δ clean enough to memorize. The previous concern was: in real raters, one might be optimistic and the other conservative, splitting the secondary classes differently; if equal-split fails, TVD exceeds δ, and the ceiling is depressed further.

    This direction of concern is correct, but the magnitude calculation was wrong—and the error lies in the premise itself that “the equal-split assumption is fragile.”

    Let us first present the closed-form solution. Under three-class adjudication, the total variation distance between two marginal distributions p and q has an identity: TVD = max{|p₁−q₁|, |p₂−q₂|, |p₃−q₃|}. Because the sum of probability differences across three classes is zero, exactly one class difference equals the sum of the other two; thus, half the L1 norm equals the largest marginal difference. This result is independent of parameterization and purely algebraic; random sampling of three million marginal pairs verified the error at 10⁻¹⁶ magnitude.

    The first corollary resolves half of the previous concern: TVD ≥ |p_main−q_main| = δ, always holds. Therefore, the upper bound 1−δ for p₀ is never violated; the equal-split assumption is not the load-bearing component of this bound.

    The second corollary is stronger: whether “TVD exceeds δ” depends on whether the primary class difference is the largest among the three differences. The equal-split assumption affects only the other two classes, while the primary class difference δ already welds the floor shut.

    Now consider directionality. Co-directional misalignment (both raters assign more of the remainder to “Improvement”) is entirely harmless: their offsets cancel each other, TVD remains exactly δ, unchanged—enumerating ε from 0 to 0.4 yields TVD consistently at 0.0500. Only counter-directional misalignment (one optimistic, one conservative) incurs cost, and it has a dead zone. Suppose the two raters allocate the remainder mass to the second class in proportions 0.5+ε and 0.5−ε, with remaining masses s_A and s_B; solving yields the critical point:

    ε* = δ / (2(s_A+s_B))

    When ε ≤ ε*, TVD exactly equals δ, with misalignment completely masked by the primary class difference; beyond this, TVD rises linearly with slope s_A+s_B.

    In the symmetric case m_A = m̄+δ/2, m_B = m̄−δ/2, we obtain ε* = δ / (4(1−m̄)). This formula is counterintuitive: the more concentrated the primary class, the wider the dead zone, and the safer the equal-split assumption. With m̄=0.90, δ=0.05, ε*=0.125—raters must differ by 25 percentage points in remainder allocation (one at 62.5%/37.5%, the other at 37.5%/62.5%) before TVD moves. Conversely, with m̄=0.70, δ=0.05, ε*=0.042, the dead zone shrinks to 8.3%.

    The sole controller is the ratio r = δ/(1−m̄): how large the primary class difference is relative to the remaining space. The dead zone boundary is ε* = r/4, and the dead zone occupies r/2 of the “maximum possible misalignment 0.5”. When r=0.1, it covers only 5% (ε*=0.025); r=0.2 covers 10%; r=0.5 covers 25%; r=1 covers 50%; r=2 covers 100% (ε*=0.5, so any deviation is safe). The entire derivation reduces to this single dimensionless number.

    Beyond the dead zone, TVD’s shape remains linear: let t = ε/ε*, then TVD/δ = (1+t)/2 (t>1). That is, the excess of TVD over δ depends only on t, not on δ itself or m̄ itself—all effects of m̄ and δ are absorbed into the single scale ε*. When t=2, TVD is 1.5 times δ; when t=4, it is 2.5 times δ. Correspondingly, the drop in κ_max: with m̄=0.90, δ=0.05, t=1 gives 0.731, t=2 gives 0.599 (drop 18.1%), t=3 gives 0.470 (drop 35.8%); at t=4 (ε=0.5, the theoretical limit), it drops to 0.344 (drop 52.9%). Note that this “18% drop” matches the number cited previously—but it corresponds to t=2, i.e., ε exactly twice the dead zone boundary, not “ε slightly larger.”

    Returning to the specific concern: if ε’s typical scale is indeed 0.02–0.05, it lies entirely within the dead zone (ε*=0.125) for m̄=0.90, δ=0.05, so t ranges from 0.16 to 0.4, and TVD remains exactly δ; the mapping holds without error. To construct the example cited previously (A=(0.925,0.05,0.025), B=(0.875,0.025,0.10)), rater A must assign all 0.05 of the remainder 0.075 to the second class, while rater B assigns all 0.10 of the remainder 0.125 to the third class—their deviations are 0.167 and 0.30, respectively, in opposite directions. In terms of the symmetric parameter family, this example is equivalent to t=2 (TVD/δ=1.5), i.e., ε exactly twice the dead zone boundary—meaning the raters pushed the remainder mass almost oppositely toward two extremes, not “a slight stylistic difference.”

    Thus, the four criteria left by the previous discussion require comprehensive revision. The original criteria treated “ε<0.01” as the safety line and “ε>δ” as the failure line, but both have mismatched dimensions: ε and δ cannot be directly compared; they must be divided by the remaining space. The correct form is: when r = δ/(1−m̄) ≤ 0.5 (the dead zone covers 25% of misalignment space), the equal-split assumption serves as an engineering approximation; when r increases and ε exceeds r/4, one must switch to TVD reporting “the sum of marginal total variation distances”; only when ε crosses the dead zone and “the two raters’ secondary class biases are systematically opposite” can we speak of “different adjudication directions” rather than “inaccurate adjudication.”

    Finally, reporting format. Since κ_max = 1 − TVD/(1−Σp_k q_k) and TVD is not a monotonic function of δ, the term “bias” should be replaced. After fixing (m̄, TVD) and sweeping different marginal shapes, κ_max spans only 0.0015–0.0078, far narrower than the 0.05 bandwidth—indicating (m̄, TVD) is nearly sufficient as a pair and can replace single δ as the report item. Normatively, we may directly require reporting both m̄ and TVD, with δ relegated to reference.

    A point easy to misread: κ_max monotonically decreases with TVD (p_o^max = 1−TVD is an identity, independent of marginal shape), so what truly determines the ceiling throughout is TVD; δ serves only as a lower bound and common proxy for TVD. The full role of the equal-split assumption is to make this lower bound an equality. In other words, the previous question “Can δ serve as TVD?” has the answer: δ is always the lower bound of TVD; the equality condition requires ε ≤ ε*, and this condition holds almost certainly in high-concentration regions (m̄≥0.9), where the dead zone is wide.

    In one sentence: the equal-split assumption does not bear weight; the direction δ→upper bound always holds; numerically, only one ratio r = δ/(1−m̄) matters: use δ confidently when r ≤ 0.5; beyond that, convert via TVD/δ = (1+t)/2, where t = ε/(r/4).

    Evidence and Limitations

    Verified: Under three classes, TVD equals the maximum of the absolute differences of the three marginals, a pure algebraic identity (randomly sampled three million marginal pairs yielded error 10⁻¹⁶); TVD ≥ δ always holds, so the upper bound 1−δ is unconditionally valid; under co-directional misalignment, TVD exactly equals δ (ε swept 0–0.4 remained constant); the critical point for counter-directional misalignment ε* = δ/(2(s_A+s_B)) matches empirical pointwise verification; the closed-form TVD = max(δ, δ/2+ε(s_A+s_B)) has error 10⁻¹⁶ across the full parameter domain; κ_max = 1 − TVD/(1−p_e) follows directly from the identity p_o^max = Σmin(p,q) = 1−TVD.

    Partially Verified: The sufficiency of replacing δ with the (m̄, TVD) pair holds numerically as an “approximation” (fixing (m̄, TVD) yields κ_max spanning 0.0015–0.0078), but lacks analytical proof of residual upper bounds; the dead zone width depends on symmetric parameterization m_A = m̄±δ/2, whereas in real studies the primary class proportions of two raters may not be symmetric, so in asymmetric cases ε* shifts (larger primary class difference delays the onset of TVD increase, without changing direction). Literature only corroborates the general conclusion that “marginal distribution is an independent degree of freedom for kappa” (PMID 8501467 decomposes κ into bias, prevalence, and marginal agreement; PMID 8621993 states κ’s bounds are entirely determined by marginals; PMID 26095449 notes κ-type extension coefficients are “prone to prevalence and marginal distribution issues” and adopts ordered GLMM modeling; PMID 41230191 similarly highlights sensitivity of weighted κ-type coefficients to marginal distribution and category prevalence), but no literature directly provides this paper’s three-class TVD closed-form and dead zone.

    Pending Verification: What is the typical scale of ε in real three-class adjudication data—this paper can only state “if ε≤0.05, it is safe,” without empirical distribution; the actual placement of δ/(1−m̄) in published studies remains unknown; whether counter-directional misalignment correlates with rater experience, adjudication order, or presence of review mechanisms has not been examined.

    Related Reading

    L1 Network Status: The CORRESPONDENCE_NETWORK.md edge table ends at E-050 (Na Yin ↔ Frequency, L1 | open); the modern evidence column reads “No measurable source yet“; the network alignment indicator row reads “Edges ≥ L2 | 0“. The three-domain node table lacks any registration for “adjudication / rater / marginal distribution / agreement rate” (full table search for “adjudication / decision / agreement rate / marginal / bias / bandwidth / TVD” yields zero hits); the physical side “phase / rhythm” semantics refers to “two oscillators phase-locked,” not marginal distribution comparison. Thus, this paper’s evidence level caps at C, and the topic has not yet entered the L1 corresponding network (registered to l1-gaps).

    Evidence Level: C
    References: PMID 8501467, PMID 8621993, PMID 2348207, PMID 25162681, PMID 26095449, PMID 41230191, PMID 37127220


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