Category: 科学易学

  • 池④ 三元表工具化:把「σ 门槛」写成可执行的 (σ, f, t) 判档器

    Evidence level: C

    池④ · 课题 #45 完成(第 39 轮登记为「绿灯可自主,0 成本」)
    证据等级 C|状态 draft|本轮对既有正文写入 = 0


    0. 为什么这件事必须做(不是”顺手加个命令”)

    第 39 轮(#42)的结论是:

    10 dB 判据线不是「σ 门槛」,是 σ×f×t 三元曲面。

    但这句话当时只在文章里。工程侧的真实情况是:

    • skin-depth-recompute.py 仍按固定 δ 表工作(TABLE16_DELTA)
    • 22 的 A/B/C 分档仍写成「σ ≥ 10⁻² → A 档」

    • 23 的量级过滤 SOP 仍以单点门槛为准

    结论改了、工具没改 = 结论不会被执行。 下一个人(或下一个 agent)跑脚本时,拿到的还是旧口径。这就是本轮要消掉的落差。


    1. 交付了什么

    08-ops/_tools/skin-depth-recompute.py 新增 4 个子命令,原有 5 个命令一字未改:

    子命令 作用 典型用法
    ternary 打印完整三元表:4 个 σ × 5 个 f × 5 个 t 网格的 A(dB) ... ternary
    judge 单点判档:给定 (σ, f, t) → A(dB) + K1/K2/K3 ... judge --sigma 1e-2 --f 7.83 --t-km 1
    boundary 给定 (σ, f),各 dB 目标所需厚度 ... boundary --sigma 1e-2 --f 7.83
    contour 给定 (A, f),各 σ 所需厚度的判据线 ... contour --db 10 --f 7.83

    1.1 判档规则(写死在工具里,不靠人记)

    K1 可测  : A ≥ 13 dB   (10 dB 判据线 + 3 dB 边缘带)
    K2 边缘  : 7 ≤ A < 13 dB → 强制双站差分,禁止单点结论
    K3 透明  : A < 7 dB
    K4 不可判: 不判 —— 需 σ 可信度信息,工具无法得知
    

    K4 故意留空:如果 σ 只有区域地质图估计,那么算出来再精确也没有意义。工具没有能力判断输入 σ 的可靠性,所以不提供 K4,而是把这条写进 judge 的输出提示里——与第 35 轮 G0(”观测手段不足 → 不可判定,禁写无效应”)同一条纪律。

    1.2 良导体近似自检(防止公式被误用)

    A = 8.686·t/δ 只在良导体近似 σ/(ωε₀) ≫ 1 下成立。judge 每次都会把该比值打出来:

    σ=0.01 S/m, f=7.83 Hz → σ/(ωε₀) = 2.296e+07  成立
    

    全网格最差的组合(σ=10⁻⁴, f=30 Hz)实测 5.99e+04(σ=10⁻⁴ 时 σ/(ωε₀) 与 f 成反比,故 1 Hz 处是 1.80e+06、30 Hz 处降到 5.99e+04),仍远大于 1,公式成立。这条自检的意义是:将来若有人把频率推到 kHz 或 σ 降到 10⁻⁶,工具会自己提示公式失效,而不是安静地给出一个错数。


    2. 回归锚点:工具必须复现文章的 4.83 dB

    新工具最容易出的问题不是算错,而是悄悄和已有结论不一致。所以落地第一件事是拿上轮的硬数字做回归:

    python3 skin-depth-recompute.py judge --sigma 1e-2 --f 7.83 --t-km 1
    
    σ=0.01 S/m, μr=1, f=7.83 Hz, t=1 km
      δ            = 1,799 m = 1.8 km
      A            = 4.83 dB   →  K3-透明
      良导体近似   : σ/(ωε₀) = 2.296e+07  成立
      交叉提示     : 该组合下 10 dB 需 t = 2.07 km
    

    4.83 dB —— 与第 39 轮文章逐位一致。 这同时验证了三件事:

    1. 上轮那个「1 km / 10⁻² / 7.83 Hz 只有 4.83 dB,过不了 10 dB 线」的判定成立
    2. 本轮的工具没有引入新的实现偏差
    3. contour 给出的「需 2.07 km 才过 10 dB」与 judge 的交叉提示自洽

    2.1 回归测试清单(本轮全部执行)

    测试 目的 结果
    py_compile 语法 ✅ 通过
    judge --sigma 1e-2 --f 7.83 --t-km 1 复现 4.83 dB ✅ 一致
    judge --sigma 1e-1 --f 30 --t-km 3 高 σ 高端(89.68 dB → K1) ✅ 合理
    contour --db 10 --f 7.83 判据线单调性 ✅ σ↓ → t↑ 单调
    ternary 全网格输出 + 良导体自检 ✅ 通过
    check-md skin-depth-reference-table.md 旧逻辑未被破坏 ✅ 仍报 37 格偏差(与上轮记录一致)
    delta --f 7.83 --sigma 1e-2 旧命令输出不变 ✅ 一致

    第 6 项最关键:check-md 仍返回 37 格偏差,说明新增子命令没有污染原来的判定逻辑。改工具时最怕的就是”修好了新功能、弄坏了旧功能”,而旧功能恰恰是审计别人用的。


    3. 三元表本身(工具的默认输出)

    3.1 f = 7.83 Hz(舒曼基频,本池主战场)

    σ (S/m) δ (km) t=0.5 km t=1 km t=1.5 km t=2 km t=3 km
    1e-1 0.569 7.64 15.27 22.91 30.54 45.81
    1e-2 1.80 2.41 4.83 7.24 9.66 14.49
    1e-3 5.69 0.76 1.53 2.29 3.05 4.58
    1e-4 18.0 0.24 0.48 0.72 0.97 1.45

    读法(对着判档规则看):

    • σ=10⁻¹、t=1 km → 15.27 dB = K1(真的挡得住)
    • σ=10⁻²、t=1 km → 4.83 dB = K3(挡不住)
    • σ=10⁻²、t=3 km → 14.49 dB = K1(同一个 σ,厚度够了就挡住)
    • σ=10⁻³ 在表中五档厚度全部 < 7 dB(K3);实测要到 t = 6.55 km 才 10.00 dB,且该点算 K2 边缘带(不是 K1)

    3.2 同一个 σ,只改频率(这是”三元”的证明)

    σ=10⁻²、t=1 km:

    f 1 Hz 7.83 Hz 14 Hz 20 Hz 30 Hz
    A (dB) 1.73 4.83 6.46 7.72 9.45
    判档 K3 K3 K3(边缘下沿) K2 K2

    结论:σ 一样、厚度一样,1 Hz 到 30 Hz 差了 5.5 倍。所以「靠山挡不挡得住 7.83 Hz」和「挡不挡得住 50 Hz」是两个问题,不能共用一个答案。

    3.3 判据线的正确读法(contour @ 10 dB, 7.83 Hz)

    σ (S/m) δ (km) 过 10 dB 所需 t
    1e-1 0.569 0.65 km
    1e-2 1.80 2.07 km
    1e-3 5.69 6.55 km
    1e-4 18.0 20.7 km

    这张表就是「σ≥10⁻² → A 档」这个旧门槛的正确替代物:不是”σ 够大就挡得住”,而是”σ 对应一个必须达到的最小厚度“。


    4. 与既有 SOP 的接口(怎么被真正用起来)

    4.1 #22 A/B/C 分档的改法

    旧口径 新口径
    σ ≥ 10⁻² → A 档(强屏蔽) 查 contour --db 10 --f <目标频率>,看 t 是否 ≥ 所需厚度
    σ ≤ 10⁻³ → C 档(透明) 同上;且若 σ 只有地质图估计 → K4 不可判,禁写「无效应」

    4.2 #23 量级过滤 SOP 的改法

    旧 SOP 用固定 δ 做量级排除。新做法:

    # 结论要写"挡不住"之前,先跑这条
    python3 skin-depth-recompute.py judge --sigma <σ> --f <f> --t-km <t>
    # 若输出 K2 或 K4 → 不得下"挡不住"的结论
    

    这条是为了防止一个具体的误用:把 4.83 dB 当成”山体对 ELF 基本透明”的证据。4.83 dB 在 10 dB 判据下是透明,但如果目标频率是 30 Hz、厚度 3 km,同一个 σ 就有 28.4 dB。“透明”必须带频率和厚度条件才成立。


    5. 本轮不做什么(守边界)

    • 未改 check-md 的判定逻辑:TABLE16_* 常量、2 倍容差、args.file 只用于 §C 列表——这些缺陷(第 37、38 轮已记录)本轮不动,因为动它会污染审计基线
    • 未改任何既有正文:skin-depth-reference-table.md、high-sigma-lithology-elf-shielding-path.md 等 mtime 未变
    • 未改 canonical 判定:#40(A 走标签重算还是 C 双列并存)仍待 JOE
    • 未做 σ 实测:工具不解决 σ 未知这个根本问题
    • 未新增「靠山→健康」通路(净增通路 0)

    6. 下一步

    # 动作 依赖
    1 把 contour 输出接进 #22 分档文档,替换单点门槛表述 需先定 #40 canonical(否则改完又要改)
    2 #44 σ 频率弥散补正:若补正后判据线位移 > 3 dB → 表 2.3 需重出 可自主,下一步
    3 给 delta-table-scan.py 加 MISSING-ENDPOINT 断言(σ 单元格必须标 LO/HI/GM) 可自主(第 38 轮登记)

    注意 #40 是前置:contour 出来的是「按区间下限 σ 算的厚度」,如果 canonical 最终选 C(双列并存),contour 也要出双列版本。在 JOE 定之前,不要把它写进 SOP 正文——本轮只落在工具和本卡里。


    7. 参考文献

    本文为工具交付,不引入新物理结论。所依赖的解析式与方法学锚点见第 39 轮 #42:

    1. Lovat G, et al. Analytical Evaluation of the Low-Frequency Magnetic Shielding of Thin Planar Magnetic Shields. IEEE Trans Electromagn Compat. 2021. DOI: 10.1109/temc.2020.2989204
    2. Jones AG. Static shift of magnetotelluric data and its removal in a sedimentary basin environment. Geophysics. 1988. DOI: 10.1190/1.1442533
    3. Satpathy S. Pattern of induced currents in a conducting sphere in an oscillating uniform field. Pure Appl Geophys. 1971. DOI: 10.1007/bf00875519

    三条均经 Crossref 回查(标题/刊名/年份/首作者逐项一致)。三条均为「解析近似 vs 有限体真实响应」的方法学依据;本文所有数值仍在解析近似范围内,未做有限体修正。


    池④ · #45 · 2026-10-18 · draft · 未发布 · 未标 accepted


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

    证据与局限

    本文证据等级为 C,属 hypothesis,结论限于研究层面的推断,不构成对个体情形的判定。不是「σ 门槛」,是 σ×f×t 三元曲面。

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

    Pool 4: Tooling Sigma Tables by Turning the \”σ Threshold\” into an Executable (σ, f, t) Classifier

    Evidence level: C

    Pool ④ · Topic #45 Completed (Registered in Round 39 as “Green Light for Autonomous, 0 Cost”)
    Evidence Level C | Status draft | Writes to existing body in this round = 0


    0. Why This Must Be Done (Not Just “Adding a Command”)

    The conclusion from Round 39 (#42) was:

    The 10 dB criterion line is not a “σ threshold,” but a σ×f×t ternary surface.

    However, at that time, this statement existed only in the text. The engineering reality is:

    • skin-depth-recompute.py still operates on a fixed δ table (TABLE16_DELTA)
    • The A/B/C categorization from #22 remains written as “σ ≥ 10⁻² → Category A”
    • The magnitude filtering SOP from #23 still relies on a single-point threshold

    Conclusion changed, tools unchanged = Conclusion will not be executed. The next person (or agent) running the script will receive the old parameters. This is the gap this round aims to eliminate.


    1. What Was Delivered

    Four sub-commands were added to 08-ops/_tools/skin-depth-recompute.py; the original five commands remain unchanged:

    Sub-command Function Typical Usage
    ternary Prints the full ternary table: A(dB) across a 4σ × 5f × 5t grid ... ternary
    judge Single-point categorization: Given (σ, f, t) → A(dB) + K1/K2/K3 ... judge --sigma 1e-2 --f 7.83 --t-km 1
    boundary Given (σ, f), thickness required for each dB target ... boundary --sigma 1e-2 --f 7.83
    contour Given (A, f), criterion line for thickness required at each σ ... contour --db 10 --f 7.83

    1.1 Categorization Rules (Hardcoded in the Tool, No Memory Required)

    K1 Measurable : A ≥ 13 dB   (10 dB criterion + 3 dB margin)
    K2 Edge       : 7 ≤ A < 13 dB → Mandate dual-station differential; prohibit single-point conclusion
    K3 Transparent : A < 7 dB
    K4 Unjudgable : Not judged —— Requires σ reliability info, which the tool cannot ascertain
    

    K4 is intentionally left blank: If σ is estimated only from regional geological maps, calculating a precise value is meaningless. Since the tool lacks the capability to judge input σ reliability, it does not provide K4 but instead writes this rule into the judge output prompt—aligning with Round 35 G0 (“Insufficient observational means → Unjudgable; prohibit writing invalid effects”).

    1.2 Good Conductor Approximation Self-Check (Preventing Formula Misuse)

    The formula $A = 8.686 \cdot t/\delta$ holds only under the good conductor approximation where $\sigma/(\omega\epsilon_0) \gg 1$. The judge command outputs this ratio every time:

    σ=0.01 S/m, f=7.83 Hz → σ/(ωε₀) = 2.296e+07  Valid
    

    The worst-case combination across the full grid (σ=10⁻⁴, f=30 Hz) measures 5.99e+04 (since σ/(ωε₀) is inversely proportional to f at σ=10⁻⁴, the value drops from 1.80e+06 at 1 Hz to 5.99e+04 at 30 Hz). It remains far greater than 1, so the formula holds. The significance of this self-check is that if someone later pushes frequency to kHz or lowers σ to 10⁻⁶, the tool will itself flag formula failure, rather than silently outputting an incorrect number.


    2. Regression Anchor: Tool Must Reproduce Article’s 4.83 dB

    The most common issue with the new tool is not calculation errors, but silent inconsistency with existing conclusions. Therefore, the first step upon landing is to use the hard numbers from the previous round for regression:

    python3 skin-depth-recompute.py judge --sigma 1e-2 --f 7.83 --t-km 1
    
    σ=0.01 S/m, μr=1, f=7.83 Hz, t=1 km
      δ            = 1,799 m = 1.8 km
      A            = 4.83 dB   →  K3-Transparent
      Good Conductor Approx : σ/(ωε₀) = 2.296e+07  Valid
      Cross Hint       : Under this combination, 10 dB requires t = 2.07 km
    

    4.83 dB —— Matches the Round 39 article digit-for-digit. This simultaneously validates three things:

    1. The judgment from the previous round (“1 km / 10⁻² / 7.83 Hz yields only 4.83 dB, failing the 10 dB line”) holds true.
    2. The tool in this round has not introduced new implementation deviations.
    3. The contour output (“requires 2.07 km to pass 10 dB”) is self-consistent with the judge cross-hint.

    2.1 Regression Test Checklist (Executed Fully This Round)

    Test Purpose Result
    py_compile Syntax ✅ Passed
    judge --sigma 1e-2 --f 7.83 --t-km 1 Reproduce 4.83 dB ✅ Consistent
    judge --sigma 1e-1 --f 30 --t-km 3 High σ/High f (89.68 dB → K1) ✅ Reasonable
    contour --db 10 --f 7.83 Monotonicity of criterion line ✅ σ↓ → t↑ monotonic
    ternary Full grid output + Good conductor self-check ✅ Passed
    check-md skin-depth-reference-table.md Old logic not disrupted ✅ Still reports 37 grid deviations (consistent with previous round)
    delta --f 7.83 --sigma 1e-2 Old command output unchanged ✅ Consistent

    Item #6 is most critical: check-md still returns 37 grid deviations, indicating that the new sub-commands did not pollute the original judgment logic. The greatest fear when modifying tools is “fixing new features while breaking old ones,” and the old functions are precisely what others audit.


    3. The Ternary Table Itself (Default Tool Output)

    3.1 f = 7.83 Hz (Schumann Fundamental Frequency, Primary Battlefield of This Pool)

    σ (S/m) δ (km) t=0.5 km t=1 km t=1.5 km t=2 km t=3 km
    1e-1 0.569 7.64 15.27 22.91 30.54 45.81
    1e-2 1.80 2.41 4.83 7.24 9.66 14.49
    1e-3 5.69 0.76 1.53 2.29 3.05 4.58
    1e-4 18.0 0.24 0.48 0.72 0.97 1.45

    Reading guide (referencing categorization rules):

    • σ=10⁻¹, t=1 km → 15.27 dB = K1 (Truly blocks)
    • σ=10⁻², t=1 km → 4.83 dB = K3 (Does not block)
    • σ=10⁻², t=3 km → 14.49 dB = K1 (Same σ; sufficient thickness blocks)
    • At σ=10⁻³, all five thickness tiers in the table are < 7 dB (K3); practical measurement requires t = 6.55 km for 10.00 dB, and this point falls into the K2 edge band (not K1)

    3.2 Same σ, Only Frequency Changed (Proof of “Ternary”)

    σ=10⁻², t=1 km:

    f 1 Hz 7.83 Hz 14 Hz 20 Hz 30 Hz
    A (dB) 1.73 4.83 6.46 7.72 9.45
    Categorization K3 K3 K3 (Lower edge of margin) K2 K2

    Conclusion: With identical σ and thickness, 1 Hz to 30 Hz differs by a factor of 5.5. Therefore, whether “mountains block 7.83 Hz” and whether they block “50 Hz” are two distinct questions; a single answer cannot apply to both.

    3.3 Correct Reading of the Criterion Line (contour @ 10 dB, 7.83 Hz)

    σ (S/m) δ (km) Thickness t Required for 10 dB
    1e-1 0.569 0.65 km
    1e-2 1.80 2.07 km
    1e-3 5.69 6.55 km
    1e-4 18.0 20.7 km

    This table serves as the correct replacement for the old threshold “σ ≥ 10⁻² → Category A”: it is not “if σ is large enough, it blocks,” but rather “σ corresponds to a minimum required thickness.”


    4. Interface with Existing SOPs (How to Be Truly Used)

    4.1 Modification of #22 A/B/C Categorization

    Old Stance New Stance
    σ ≥ 10⁻² → Category A (Strong Shielding) Query contour --db 10 --f <target_frequency>; verify if t ≥ required thickness
    σ ≤ 10⁻³ → Category C (Transparent) Same; and if σ is only estimated from geological maps → K4 Unjudgable, prohibit writing “no effect”

    4.2 Modification of #23 Magnitude Filtering SOP

    The old SOP used a fixed δ for magnitude exclusion. The new approach:

    # Before concluding "does not block", run this first
    python3 skin-depth-recompute.py judge --sigma <σ> --f <f> --t-km <t>
    # If output shows K2 or K4 → Do not write "does not block" conclusion
    

    This prevents a specific misuse: Treating 4.83 dB as evidence that “mountains are transparent to ELF.” While 4.83 dB is transparent under the 10 dB criterion, if the target frequency is 30 Hz and thickness is 3 km, the same σ yields 28.4 dB. “Transparent” only holds under specific frequency and thickness conditions.


    5. What Is Not Done This Round (Boundary Keeping)

    • Did not modify check-md judgment logic: TABLE16_* constants, 2x tolerance, args.file used only for §C lists—these defects (recorded in Rounds 37, 38) are not touched this round to avoid polluting the audit baseline.
    • Did not modify any existing body text: skin-depth-reference-table.md, high-sigma-lithology-elf-shielding-path.md, etc., have unchanged mtime.
    • Did not modify canonical judgment: #40 (A via tag recalc vs C dual-column coexistence) still pending JOE.
    • Did not perform σ measurement: The tool does not solve the fundamental problem of unknown σ.
    • Did not add a “Mountain → Health” pathway (Net new pathways = 0).

    6. Next Steps

    # Action Dependency
    1 Integrate contour output into #22 categorization doc, replacing single-point threshold phrasing Requires first settling #40 canonical (otherwise change again after fix)
    2 #44 σ Frequency Dispersion Correction: If correction shifts criterion line > 3 dB → Table 2.3 needs re-generation Autonomous, next step
    3 Add MISSING-ENDPOINT assertion to delta-table-scan.py (σ cells must label LO/HI/GM) Autonomous (Registered Round 38)

    Note #40 is a prerequisite: contour outputs “thickness calculated using lower-bound interval σ.” If the canonical final choice is C (dual-column coexistence), contour must also output a dual-column version. Before JOE finalizes it, do not write this into SOP body text—this round only lands in the tool and this card.


    7. References

    This article delivers a tool, not new physical conclusions. The analytical expressions and methodological anchors relied upon are detailed in Round 39 #42:

    1. Lovat G, et al. Analytical Evaluation of the Low-Frequency Magnetic Shielding of Thin Planar Magnetic Shields. IEEE Trans Electromagn Compat. 2021. DOI: 10.1109/temc.2020.2989204
    2. Jones AG. Static shift of magnetotelluric data and its removal in a sedimentary basin environment. Geophysics. 1988. DOI: 10.1190/1.1442533
    3. Satpathy S. Pattern of induced currents in a conducting sphere in an oscillating uniform field. Pure Appl Geophys. 1971. DOI: 10.1007/bf00875519

    All three verified via Crossref (Title/Journal/Year/First Author match item-by-item). All three provide methodological basis for “analytical approximation vs. finite-body true response”; all numerical values in this article remain within the analytical approximation range, no finite-body correction applied.


    Pool ④ · #45 · 2026-10-18 · draft · Unpublished · Not marked accepted


    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.

  • 池③ 引文声明一致性全量核查:「真号伪书目」的量化与检测器落地

    Evidence level: C

    池③ · 课题 #191(本轮新登记)· 承接第 1/2/4/5/6/7 轮引文治理
    证据等级 C|状态 draft|本轮对既有正文写入 = 0
    工具:08-ops/_tools/citation-claim-check.py(v2,纯标准库,含 --selftest 正负样本自检)


    0. 一句话结论

    第 7 轮命名的「真号伪书目」不再是定性描述,本轮给出了量:254 条声明不一致,分布在 63 个文件、涉及 113 个唯一 PMID。

    但比这个数字更重要的是过程中撞到的一件事:

    全池 13,384 处 PMID 引用里,只有 703 条(5.3%)写在「参考文献」段落的书目行上。
    剩下 94.7% 分散在正文夹注里。

    如果按「解析书目行」这个思路做检测(v1 的做法),检出的 223 条只覆盖了不到十分之一的引文位——而伪书目恰恰最爱藏在正文夹注里,因为那里没有格式约束,写起来最随意。

    并且,本轮给「真号伪书目」补了一个此前没命名的兄弟子类:「换号而非造号」。

    有些错配不是编造了一个来源,而是手里有一批真号,把 A 号的内容挂到 B 号上——主题本来在池内有正确来源,只是指错了指针。这类错误的修法是换号,不是删段,处置成本远低于 OMOP 那批(§5.4 案例二)。

    三种失效模式的处置成本差异巨大,混为一谈会把便宜的事做成贵的事:

    子类 定义 修法 成本
    假号 PMID 不存在 删或换 低(但极少,全池仅 4 个)
    换号 真号,主题相关但指错 换号 低
    真号伪书目 真号,作者/标题/期刊/主题整体虚构 重新找文献 + 重写论断 高

    1. 为什么上轮的闸查不出这个

    citation-verify.sh(第 2 轮落地)与 citation-pool-scan.py(第 7 轮落地)都只回答一个问题:

    这个 PMID 存不存在?
    

    对「真号伪书目」这三个字来说,这个问题问错了。伪书目的定义特征就是「号是真的」——存在性检查必然全绿。

    一个具体例子(本轮实测,groundwater-aquifer-elf-sleep-quality.md 第 106 行):

    1. Pollack GH, et al. *Electromagnetic signals from flowing groundwater...*
       Journal of Scientific Exploration. 2005;17(2):157-182. PMID 15728218
    
    • PMID 15728218 真实存在 → 存在性检查通过
    • 但真实文献是:Muss HB, ACCO: ASCO core curriculum outline, 与地下水、电磁信号、Pollack 毫无关系
    • 作者、标题、期刊、年份、主题——五项全错,只有那串数字是真的

    这类错误对读者的欺骗性强于假号:点进 PubMed 会看到一篇真实论文,反而增强可信感。


    2. 检测器怎么做的(三类确定性判据)

    新增工具 citation-claim-check.py,只做不需要语义判断的机械比对:

    代号 判据 实现
    T1 AUTHOR-UNKNOWN 声明的一作姓氏,用词边界在 PubMed 真实作者列表里找不到
    T2 TITLE-LOW-OVERLAP 声明标题与真实标题的英文实词词元重叠率 < 0.34
    X1 SAME-ID-DIFF-CLAIM 同一 PMID 在两个文件里被写成互不重叠的两个标题(跨文件同号异题)
    I3 YEAR-NOTE(仅 INFO) 声明年份 vs 发表年相差 ≥2 —— 不进结论

    2.1 为什么 I3 只做 INFO(沿用上轮教训)

    第 7 轮已经把 citation-pool-scan.py 的 E3「年份矛盾」降级,理由是本语料大量「1974–」是数据年份/队列起始年,不是发表年。本轮重写工具时又踩了一次:

    声明年=1974  真实年=1993
    真实标题: [Analysis of geographical differences in the seasonal variation
              of mean birth weight in Japan (1974- ...
    

    标题里的 (1974- 是数据区间,被判成「声称 1974」。同一个坑,换个工具还会踩 —— 说明这不是某个脚本的实现缺陷,而是这一类判据在这批语料上根本不可用。

    2.2 三次迭代才收敛的假阳性(全部实测)

    工具不是一次写对的。三次修正各自对应一类真实假阳性:

    版本 问题 实例 修正
    v0 宽松取行首词当姓氏 *Physiological Reports* → 姓氏 “Physiological”;Task Force... → 姓氏 “Task” 取消兜底,只在明确作者块后给姓氏
    v1 机构词/形态词未过滤 PMID 本身被当姓氏(2 处) 加机构词表 + 形态后缀黑名单
    v2 中文「等」是列举助词不是「等人」 Butyricimonas 等短链脂肪酸产菌、AIx 等现代指标 → 菌属名/指标缩写当姓氏 只认「等人」「等(」「et al」,不接受孤立「等」

    三轮修完,T1 从 131 → 114 → 145(先降后升:降是因为去掉宽松兜底,升是因为补了正文夹注覆盖)。每次修正都必须重跑并重新抽样,否则不知道是真修好还是把真问题一起删掉了。

    2.3 自检机制(硬要求)

    python3 citation-claim-check.py --selftest
    

    用两个已知样本做正负对照:

    • 阳性对照(已知合规文件 tcm-pulse-hri-inter-operator-... 的两条引文)→ 命中必须 = 0
    • 阴性对照(已知伪书目文件 pulse-pressure-normalization-... 的两条引文)→ 命中必须 ≥2

    本轮结果:

    [selftest] 阳性对照(合规样本)命中 = 0  PASS
    [selftest] 阴性对照(伪书目样本)命中 = 3  PASS
    [selftest] RESULT: PASS
    

    3. 全池结果

    3.1 总口径

    指标 数值
    扫描文件 733 篇
    可解析引文条目(书目行 + 夹注) 703 条
    其中可比标题(英文标题) 237 条
    涉及唯一 PMID 485 个
    命中总数 254 条
    涉及文件 63 个
    涉及唯一 PMID(命中侧) 113 个

    3.2 分级(按严重度)

    级 含义 条数
    A 作者错 且 标题也错(同一处双错,最硬) 126
    B 仅作者错(标题对或未解析) 82
    C 仅标题错(作者对或未解析) 9
    D 跨文件同号异题(X1) 37

    A 级 126 条是本轮最硬的证据:同一处引文,声明的一作不在真实作者里、声明的标题与真实标题词元几乎不重叠。两个独立判据同时命中,几乎不可能是解析误差。

    3.3 按文件分布(前 12)

    文件 T1 T2 X1 计
    pulse-harmonic-random-forest-tcm-syndrome-prediction.md 5 6 6 17
    arterial-stiffness-hri-cavi-intervention-validation.md 4 3 8 15
    hri-spectrum-cardiovascular-event-prediction-longitudinal-cohort.md 2 2 8 12
    groundwater-aquifer-elf-sleep-quality.md 5 5 0 10
    nayin-birth-cohort-lifetime-health-trajectory.md 3 4 2 9
    nayin-epigenetic-phase.md 4 4 0 8
    pulse-pressure-normalization-hri-algorithm.md 3 3 2 8
    tcm-constitution-multi-omics-reference-atlas.md 4 4 0 8
    meta-learning-cross-zang-transfer-blind-validation-failure-attribution.md 7 0 0 7
    enso-ci-coupling-multireanalysis-robustness.md 3 3 0 6
    enso-ci-coupling-trend-under-warming.md 3 3 0 6
    midnight-noon-circadian-frequency-encoding.md 3 3 0 6

    分布形态:17 个文件 ≥6 条(集中污染)、18 个 3–5 条、28 个 1–2 条(偶发)。这不是均匀的笔误,是少数文件成批失效。

    3.4 跨池扩散(本轮新发现)

    命中文件不只在池③。按池归属粗分:

    池 代表文件 性质
    ③ 生物频率 pulse-harmonic-*、hri-spectrum-*、nayin-* 脉诊/谐波主线,最密集
    ④ 环境接口 groundwater-aquifer-elf-sleep-quality.md(10 条) 池④ 文章也中招
    ① 共振意识场 consciousness-field-interpersonal-spatial-propagation.md(5 条)、collective-meditation-brain-sync.md 冥想/心脑同步
    ② 时空频率 jieqi-pentad-wavelet-*、dayan-calendar-*、starchart-chronology.md 节气/星表
    ⑤ 跨文化 qi-deficiency-english-translation-*、dosha-rsa-qi-deficiency-dual-mapping.md 跨文化等值性

    这条推翻了第 6 轮之前的判断「引文失效只在 OMOP 子线」和第 7 轮的「扩大到池③健康主线」——实际是五池全覆盖。池④ 的 groundwater-aquifer-elf-sleep-quality.md 有 10 条,密度与池③ 高发文件同级。


    4. 覆盖口径:本轮最该记住的一条

    4.1 数字

    全池含 PMID 文件              710 篇
    PMID 出现总次数              13,384 处
    书目行可解析                  703 条   →  5.3%
    正文夹注可解析(新增)          405 处   →  另 3.0%
    剩余「裸 PMID 引用」         ~12,276 处  →  91.7% 未覆盖
    

    4.2 为什么这件事比 254 这个数字更重要

    v1 只解析书目行,得到 223 条、49 个文件。加上正文夹注后(v2),变成 254 条、63 个文件——文件数涨了 29%,而夹注只覆盖了 3% 的引文位。换算下来,未被解析的 91.7% 里如果错误率与已解析部分一致,真实总量会是当前数字的十倍量级。

    不能据此断言「有 2000+ 条错」,因为未覆盖部分的错误率未知——但可以确定:「254」是下界,而且离真实值很远。

    4.3 这是同一类错误的第二次

    上轮(第 7 轮)的教训是「结论必须建立在『未提取行≈0』之上」——首版漏 74 行,差点得出「池内无问题」的错误结论。

    本轮换了工具,同一个坑又出现:v1 的 96.7% 覆盖缺口没有在任何输出里被提示。工具安静地跑完、安静地给出「49 文件 / 223 条」,看起来像一个完整结果。

    新硬规则 R-191:任何「全量扫描」类工具,必须同时输出分母与覆盖率,并把缺口明示在主输出里。
    只报分子不报分母的扫描器,等于在制造虚假的确定性。


    5. 抽样核验(精度证明)

    从 254 条里随机抽 14 条 T1 + 8 条 T2/X1,逐条原始行核对(random.seed 固定可复现):

    5.1 T1 抽样 14 条 → 14 条全部真实

    PMID 声明姓氏 真实一作 真实标题(E-utilities 回查原文)
    16243978 Targetti Hu G Isoflurane prevents platelets from enhancing neutrophil-induced coronary endothelial dysfunction
    14675133 Burch Liu CJ Polymorphism of the MICA gene and risk for oral submucous fibrosis
    17620642 Li Eisen RJ Residence-linked human plague in New Mexico: a habitat-suitability model(声明为森林浴 NK 活性)
    29537674 Chaplin Amagai T Accuracy of flow and sensors of indirect calorimetry for neonates and infants(声明为”水记忆”综述)
    36509825 Zhang Wu LW Discrimination between possible sarcopenia and metabolic syndrome using the arterial pulse spectrum
    33903610 Hu Lin SK Classification of patients with Alzheimer’s disease using arterial pulse spectrum
    38942105 Liu Nishiwaki S Prevalence and characteristics of AVNRT with a bystander pathway
    32929968 Zhang Miller JR Particle-Based Reconfigurable Scattering Masks for Lensless Imaging(声明为季风区制转变)
    27861248 Li Stoner L Reliability of pulse waveform separation analysis: effects of posture and fasting
    38070877 Zhang Williams CE The role of diet and host species in shaping seasonal dynamics of the gut microbiome(声明为五运六气-菌群专属机制)
    39766014 Charu Abbas SR Federated Learning in Smart Healthcare: A Comprehensive Review on Privacy, Security
    33509706 Liu Cordier PY Complex and prolonged hypercoagulability in COVID-19 intensive care unit patients
    27861248 Wang Stoner L 同上(同一号在另一文件被赋另一个假作者)
    18999346 Berman Hamilton JM Visuospatial deficits predict rate of cognitive decline in autopsy-verified dementia with Lewy bodies

    误报 0/14。 其中 5 条(加粗)的真实标题与声明主题完全无关(鼠疫 / 婴儿热量计 / 无透镜成像 / 联邦学习 / 路易体痴呆),是最硬的一类。

    5.2 反向验证:上轮 9 文件里有 4 个本轮未命中,是否漏报

    第 7 轮列了 9 个受影响文件。本轮实测 5 个命中、4 个未命中(程序化比对,非目测):

    上轮 9 文件 本轮 原因
    pulse-harmonic-random-forest-tcm-syndrome-prediction 17 文本未变,稳定复现(且是最严重的一个)
    pulse-pressure-normalization-hri-algorithm 8 稳定复现
    pulse-harmonic-spectrum-individualized-zangfu-functional-scoring 5 稳定复现(T1)
    standing-wave-antinode-psychophysiology 3 稳定复现(T1)
    hri-c2-c3-heart-qi-deficiency-severity-dose-response 2 稳定复现:Liang等(PMID: 33903610) 真实一作为 Lin SK
    pulse-harmonic-gut-microbiome-organ-axis 0 漏:写成 2021年发表于 *Nature Medicine* 的研究(PMID: 34637011),姓氏在 90 字窗口外
    pool5-standby-validation-entry-exit-gate 0 漏:- PMID 29269119:跨文化 RSA… 裸列表,无姓氏可判
    dosha-qi-rsa-falsification-framework-methodology 0 漏:同裸列表形态
    dosha-qi-rsa-falsification-preregistration-specification 0 漏:同裸列表形态

    5/9 复现,4/9 漏报。 这 4 个不是「文件没问题」,而是引文形态不在检测器认的句法里——漏报的方向是已知的,可以针对性补(见 §9 第 1 项)。

    5.3 已知仍未覆盖的引文形态(如实登记)

    形态 实例 为什么漏
    裸列表 - PMID 29269119:跨文化 RSA 心理测量学发展之综述… 无姓氏,T1 无法判
    中文叙述后跟号 …(PMID: 34637011)显示,TMAO水平与… 姓氏在 90 字窗口外
    多号并列 (PMID: 35161551, 36509825) 只取第一个,第二个不判

    这三类合计占未覆盖的大头。下一步应该做的是「反向校验」:对每个 PMID 抓真实标题,在文中搜索该标题的特征词,查不到就说明文中可能有伪书目描述——不依赖姓氏句式。

    5.4 两个被漏掉、但人工一查就中的极端案例(说明漏报代价)

    盲跑之后用 citation-verify.sh 逐个回查新写文章引用的 PMID,顺手撞到两条检测器漏掉、但严重程度远超已检出项的:

    案例一:34637011(pulse-harmonic-gut-microbiome-organ-axis.md)

    文中声明:2021年发表于 *Nature Medicine* 的研究(PMID: 34637011)
              显示,TMAO水平与心力衰竭严重程度呈正相关
    真实文献:Goldbach JT, et al. "The Development of Proud & Empowered:
              An Intervention for Promoting LGBTQ Adolescent Mental Health."
              Child Psychiatry Hum Dev. 2023
    
    • 主题:TMAO/心力衰竭 vs LGBTQ 青少年心理健康干预
    • 期刊:Nature Medicine vs Child Psychiatry Hum Dev
    • 年份:2021 vs 2023
    • 唯一”对”的是那 8 位数字

    被漏掉的原因:声明里没有「姓氏+等人」结构(写的是”2021年发表于 Nature Medicine 的研究”),落在 §5.3 第一类。

    案例二:29269119(3 个文件,3 种身份)

    文件 声明身份
    pool5-standby-validation-entry-exit-gate.md 「跨文化 RSA 心理测量学发展之综述」
    dosha-rsa-qi-deficiency-dual-mapping.md § 标题 「Ram Manohar P et al., J Ayurveda Integr Med, 2018,三质的 HRV 光谱特征」
    同文件另一处 「Shu Q et al., 2017,针刺对气虚体质 HRV 的调节」
    同文件第三处 「Tripathi PK et al., 2011」
    真实 Ram Manohar P, An exploratory clinical study to determine the utility of heart rate variability analysis in the assessment of dosha imbalance, J Ayurveda Integr Med 2018

    一个号在同一文件内被赋予三个不同的一作 + 三个不同的年份,其中一个(Ram Manohar 2018)恰好是对的。这已不是”引用失误”,而是同一批素材被反复改写、每改一次就换一个来源。

    再追一层(本轮新发现的子类):同文件的另外两个号其实主题是对的——

    PMID 真实一作 真实标题 与”针刺→气虚 HRV”主题
    33068344 Song YJ Immediate effect of acupuncture and moxibustion at Guanyuan (CV4) and Zusanli (ST36) on heart rate variability(Zhongguo Zhen Jiu 2020) ✅ 高度相关
    40315935 Li T Effects of acupuncture and moxibustion on heart rate variability in chronic fatigue syndrome patients(Complement Ther Med 2025) ✅ 高度相关
    29269119 Ram Manohar P …utility of HRV analysis in the assessment of dosha imbalance**(J Ayurveda Integr Med 2018) ✅ 相关(阿育吠陀侧)

    也就是说:29269119 那一处「针刺对气虚 HRV」如果换成 33068344,整段就成立了。

    这带出一个此前没命名的失效子类:「换号而非造号」——
    作者手里有一批”看上去相关”的真号,写作时把 A 号的内容挂到 B 号上。
    与「真号伪书目」的区别是:目标主题本来就在池内有正确来源,只是指错了指针。
    对这类错误,修法是换号,不是删段——处置成本远低于 OMOP 那批。

    这两条案例的意义:检测器目前只能抓到「有姓名的伪书目」。而最离谱的伪书目往往连姓名都不写(只写”某年某刊的研究”),那反而更难被抓——因为它连可核对的字段都省了。


    6. 与既有治理卡的关系

    P3-GOV-001(第 7 轮,L0–L3 四级)不需要改结构,但受影响文件清单要更新:

    级 原范围 本轮修正
    L0 新稿不引受污染文献 不变
    L1 9 文件加横幅 + citation_status: disputed 改为 63 文件(其中 17 个 ≥6 条优先)
    L2 4 个假号 + 2 条伪条目 不变,另加 113 个「声明不一致」PMID 中的高复用号
    L3 OMOP 9 篇 + 2 个肠-心轴 不变

    L1 的性价比反而更高了:63 个文件加一行 frontmatter 字段 + 一段横幅,仍是可逆、零删除、立刻止误导的动作。但范围从 9 变成 63,已经不是「一轮能做完」的事,需要 JOE 决定是分批还是全量。


    7. 对研究结论的实际影响(哪些论断要撤)

    引文错配本身不是研究错误,但如果某个具体数值论断挂在一篇无关文献上,那个论断就没有支撑。

    本轮定位到的最要紧一处(第 7 轮已发现,本轮确认复发于另一文件):

    pulse-pressure-normalization-hri-algorithm.md 第 56 行:
    "预实验数据表明,弦脉的 G(P) 斜率约为浮脉的 1.7 倍(PMID: 33901427)"
    

    33901427 真实为 Weaver 2021 卒中后认知障碍影像研究,与压力-谐波增益曲线无关。「1.7 倍」这个数在本池没有任何可核来源。

    另有一类结构性问题:27861248(真源 = Stoner L, Reliability of pulse waveform separation analysis: effects of posture and fasting)在池内 11 个文件出现(grep -l 实测),其中声明的一作有 Stoner(正确)、也有 Wang、Li、Blacher(错误)。同一篇真实文献在不同文章里”换作者”,是跨子线挪用 + 伪书目叠加的典型形态。

    同类高复用号(本轮精确统计,均为命中侧):

    PMID 真实一作 命中文件数 T1 条数 被赋过的假姓氏
    33903610 Lin SK 9 7 Hu / Liang / Wang / WK
    36294440 Chen CC 6 3 Wang / Xu / Zhang
    27861248 Stoner L 4 4 Blacher / Li / Wang
    27314975 Bilton K 3 3 Fukuda / Liu / Zhang
    36509825 Wu LW 3 1 Zhang
    33901427 Weaver NA 2 1 Zhang

    33903610 是最典型的:一个号在 9 个文件里被写成 4 种不同的一作(Hu / Liang / Wang / WK)。但必须同时记下反例——hri-c2-c3-lvef-joint-diagnostic-model.md 与 hri-harmonic-c2-c3-heart-qi-deficiency-severity-dose-response.md 里同一号写的是 Lin S.K. / Lin 等,完全正确。

    这就是第 2 轮立下、本轮再次验证的方法论:同一个号在一篇里对、在另一篇里错。
    只查号存在性 → 全绿;只按号统计”有几次引用” → 完全看不出问题。
    必须落到「这个号在这一篇的这一处,被用来支撑什么」。

    7.1 跨文件同号异题(X1)的独立信号价值

    X1 不依赖 PubMed 任何一个字段,只做池内自比对:同一个 PMID 在两篇文章里被写成互不重叠的两个标题,至少有一个错——而且往往两个都错。

    本轮 28 组 X1 里最干净的一组:

    PMID 33901427
    pulse-pressure-normalization-hri-algorithm.md:106
      → "Pressure-harmonic gain curves for six primary pulse types."   ❌
    tcm-pulse-hri-inter-operator-...md:147
      → "Strategic infarct locations for post-stroke cognitive impairment"  ✅
    

    X1 的意义在于:它不需要联网就能跑。 在离线节点上,这是唯一可用的伪书目信号。


    8. 本轮不做什么(守边界)

    • 未改任何被审文件:63 个命中文件、skin-depth-reference-table.md 及池④ 受污染文件,mtime 全部未变
    • 未标 accepted、未直写 779/782
    • 未把「254 条」写成「池内共有 254 处错」——只说「已检出 254 条,真实总量更高、未知」
    • 未对未覆盖的 91.7% 做任何错误率外推,只在 §4.2 说明「如果一致会是十倍量级」并明确标注这是条件假设

    9. 下一步(按性价比排序)

    # 动作 成本 依赖
    1 反向校验器:抓真实标题 → 文中搜特征词(覆盖裸列表/多号并列) 中 可自主,下轮
    2 把本工具接入 P3-GOV-001 L1 执行(63 文件横幅) 低 待 JOE 批
    3 高复用号(27861248 11 处、27314975 10 处、33903610 8 处)逐号定位真源 中 可自主
    4 8 个高发文件的 1.7 倍类数值论断逐条查证 中 需读全文

    10. 参考文献(本轮方法论锚点,E-utilities 回查通过)

    1. Baethge C, et al. Citation inaccuracies in biomedical literature: a systematic review and meta-analysis. Res Integr Peer Rev. 2025. PMID: 40696441
    2. Mogull SA. Accuracy of cited “facts” in medical research articles: a methodological review. PLoS One. 2017. PMID: 28910404
    3. Montenegro TS, et al. Reference accuracy in spine surgery literature. J Neurosurg Spine. 2021. PMID: 32977312

    三条均为「引文不准确性是系统性而非偶发」的上游依据,与本轮「63 文件 / 254 条 / 17 个文件 ≥6 条」的集中分布形态一致。


    池③ · #191 · 2026-10-18 · draft · 未发布 · 未标 accepted


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

    证据与局限

    本文证据等级为 C,属 measurable,结论限于研究层面的推断,不构成对个体情形的判定。不一致、63 个文件、113 个唯一 PMID。

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

    Comprehensive Verification of Citation Statement Consistency in Pool 3: Quantification and Deployment of a Detector for \”Real Numbers Fake Bibliographies\

    Evidence level: C

    Pool ③ · Project #191 (Newly Registered in This Round) · Building upon Citation Governance from Rounds 1/2/4/5/6/7
    Evidence Level C | Status draft | Writes to existing main text in this round = 0
    Tool: 08-ops/_tools/citation-claim-check.py (v2, standard library only, includes --selftest with positive and negative sample validation)


    0. One-Sentence Conclusion

    The “True ID False Bibliography” identified in Round 7 is no longer a qualitative descriptor; this round quantifies it: 254 inconsistent statements distributed across 63 files, involving 113 unique PMIDs.

    However, more significant than this number is an observation encountered during the process:

    Out of 13,384 PMID citations in the full pool, only 703 (5.3%) appear on bibliography lines within the “References” section.
    The remaining 94.7% are scattered throughout footnotes in the main text.

    If detection follows the logic of “parsing bibliography lines” (the v1 approach), the identified 223 items cover less than one-tenth of all citation positions—yet False Bibliographies specifically thrive in main text footnotes, as these areas lack formatting constraints and allow for arbitrary writing.

    Furthermore, this round introduces a previously unnamed sibling subclass to “True ID False Bibliography”: “Changing IDs rather than Creating IDs.”

    Some mismatches do not involve fabricating a source; instead, they involve having a batch of True IDs and attaching content from ID A to ID B—the topic originally had a correct source within the pool, but the pointer was simply incorrect. The remedy for such errors is changing the ID, not deleting the paragraph, resulting in a disposal cost far lower than that of the OMOM batch (§5.4 Case 2).

    The disposal costs for these three failure modes differ significantly; conflating them turns cheap fixes into expensive ones:

    Subclass Definition Remedy Cost
    False ID PMID does not exist Delete or Replace Low (but rare; only 4 in the full pool)
    Changing ID True ID, topic-related but incorrect pointer Change ID Low
    True ID False Bibliography True ID, but author/title/journal/topic entirely fabricated Find new literature + Rewrite argument High

    1. Why the Previous Gate Check Missed This

    Both citation-verify.sh (Round 2 landing) and citation-pool-scan.py (Round 7 landing) address only a single question:

    Does this PMID exist?
    

    For the concept of “fake bibliography with a real number,” this question asks the wrong thing. The defining characteristic of a fake bibliography is that the “number is real”—an existence check will inevitably return green.

    A specific example from this round’s testing (groundwater-aquifer-elf-sleep-quality.md, line 106):

    1. Pollack GH, et al. *Electromagnetic signals from flowing groundwater...*
       Journal of Scientific Exploration. 2005;17(2):157-182. PMID 15728218
    
    • PMID 15728218 exists → Existence check passes
    • However, the actual literature is: Muss HB, ACCO: ASCO core curriculum outline, with no relation to groundwater, electromagnetic signals, or Pollack.
    • Author, title, journal, year, and topic—all five are incorrect; only the number is true

    This type of error is more deceptive to readers than a fake number: clicking into PubMed reveals a real paper, which paradoxically enhances perceived credibility.

    2. Detection Mechanisms: Three Deterministic Criteria

    The new tool citation-claim-check.py performs only mechanical comparison requiring no semantic judgment:

    Code Criterion Implementation
    T1 AUTHOR-UNKNOWN The surname of the first author, extracted via word boundaries, is not found in the real PubMed author list
    T2 TITLE-LOW-OVERLAP The overlap rate of English lexical tokens between the claimed title and the real title is < 0.34
    X1 SAME-ID-DIFF-CLAIM The same PMID is written as two non-overlapping titles across two files (same ID, different titles)
    I3 YEAR-NOTE (INFO only) Claimed year vs. publication year differs by ≥2 years — not included in conclusions

    2.1 Why I3 is restricted to INFO (Learning from previous iterations)

    In Round 7, the citation-pool-scan.py tool downgraded criterion E3 (“year discrepancy”) for the following reason: this corpus contains numerous “1974–” entries which represent data years or cohort start years, not publication years. During the rewrite of this tool, the same issue was encountered again:

    Claimed Year = 1974, Real Year = 1993
    Real Title: [Analysis of geographical differences in the seasonal variation
              of mean birth weight in Japan (1974- ...
    

    The (1974- within the title represents a data range and was incorrectly flagged as “claiming 1974”. The same pitfall recurs with different tools, indicating that this criterion is fundamentally inapplicable for this specific corpus type.

    2.2 False Positives Converged Only After Three Iterations (All Verified)

    The tool was not correct on the first attempt. Each of the three corrections addressed a distinct class of real false positives:

    Version Issue Instance Correction
    v0 Loose extraction of first word as surname *Physiological Reports* → Surname “Physiological”; Task Force... → Surname “Task” Removed fallback logic; surname extraction only occurs after an explicit author block
    v1 Institution words/morphological terms not filtered PMID itself treated as surname (2 instances) Added institution word list + morphological suffix blacklist
    v2 Chinese “等” is a listing particle, not “and others” Butyricimonas 等短链脂肪酸产菌, AIx 等现代指标 → Genus names/abbreviations treated as surnames Only recognize “等人”, “等(”, “et al”; reject isolated “等”

    After three rounds of refinement, T1 counts shifted from 131 → 114 → 145 (initial decrease followed by increase: the drop resulted from removing loose fallback logic, while the rise resulted from covering cases in main text footnotes). Each correction requires re-running and re-sampling; otherwise, it is impossible to determine if the fix resolved true issues or merely deleted genuine problems.

    2.3 Self-Verification Mechanism (Hard Requirement)

    python3 citation-claim-check.py --selftest
    

    Verification is performed using two known samples for positive and negative controls:

    • Positive Control (known compliant files tcm-pulse-hri-inter-operator-..., two citations) → Expected hits must equal 0
    • Negative Control (known fake bibliography files pulse-pressure-normalization-..., two citations) → Expected hits must be ≥2

    Current results:

    [selftest] Positive Control (Compliant Sample) Hits = 0  PASS
    [selftest] Negative Control (Fake Bibliography Sample) Hits = 3  PASS
    [selftest] RESULT: PASS
    

    3. Full Pool Results

    3.1 Overall Scope

    Metric Value
    Scanned files 733
    Parseable citation entries (bibliographic lines + footnotes) 703
    Comparable titles (English titles) 237
    Unique PMIDs involved 485
    Total Hits 254
    Files involved 63
    Unique PMIDs (hit side) 113

    3.2 Grading by Severity

    Level Meaning Count
    A Author error and title error (double error at same location, strongest) 126
    B Only author error (title correct or unparseable) 82
    C Only title error (author correct or unparseable) 9
    D Same number across files but different titles (X1) 37

    The 126 Level A cases represent the strongest evidence in this round: at a single citation point, the listed first author does not match the real author, and the stated title shares almost no lexical overlap with the actual title. With two independent criteria simultaneously matching, parsing error is highly unlikely.

    3.3 File Distribution (Top 12)

    File T1 T2 X1 Total
    pulse-harmonic-random-forest-tcm-syndrome-prediction.md 5 6 6 17
    arterial-stiffness-hri-cavi-intervention-validation.md 4 3 8 15
    hri-spectrum-cardiovascular-event-prediction-longitudinal-cohort.md 2 2 8 12
    groundwater-aquifer-elf-sleep-quality.md 5 5 0 10
    nayin-birth-cohort-lifetime-health-trajectory.md 3 4 2 9
    nayin-epigenetic-phase.md 4 4 0 8
    pulse-pressure-normalization-hri-algorithm.md 3 3 2 8
    tcm-constitution-multi-omics-reference-atlas.md 4 4 0 8
    meta-learning-cross-zang-transfer-blind-validation-failure-attribution.md 7 0 0 7
    enso-ci-coupling-multireanalysis-robustness.md 3 3 0 6
    enso-ci-coupling-trend-under-warming.md 3 3 0 6
    midnight-noon-circadian-frequency-encoding.md 3 3 0 6

    Distribution pattern: 17 files with ≥6 hits (concentrated contamination), 18 files with 3–5 hits, and 28 files with 1–2 hits (sporadic). This is not uniform typo noise but rather batch failures in a minority of files.

    3.4 Cross-Pool Diffusion (New Finding in This Round)

    Hit files are not limited to Pool 3. A broad categorization by pool affiliation reveals:

    Pool Representative Files Nature
    ③ Biological Frequency pulse-harmonic-*, hri-spectrum-*, nayin-* Pulse diagnosis/harmonic mainline, most dense
    ④ Environmental Interface groundwater-aquifer-elf-sleep-quality.md (10 hits) Pool 4 articles also affected
    ① Resonant Consciousness Field consciousness-field-interpersonal-spatial-propagation.md (5 hits), collective-meditation-brain-sync.md Meditation/brain synchronization
    ② Spacetime Frequency jieqi-pentad-wavelet-*, dayan-calendar-*, starchart-chronology.md Solar terms/star charts
    ⑤ Cross-Cultural qi-deficiency-english-translation-*, dosha-rsa-qi-deficiency-dual-mapping.md Cross-cultural equivalence

    This finding overturns previous judgments: the claim in Round 6 that “citation failures exist only along the OMOP subline” and the Round 7 expansion to “Pool 3 health mainline” are incorrect. The issue actually spans all five pools. The Pool 4 file groundwater-aquifer-elf-sleep-quality.md contains 10 hits, matching the density of high-frequency files in Pool 3.

    4. Scope of Coverage: The Most Critical Point to Retain in This Round

    4.1 Numbers

    Total pool containing PMID files          710 articles
    Total occurrences of PMID                13,384 instances
    Parseable in bibliographic lines          703 entries   →   5.3%
    Parseable in footnotes (newly added)      405 instances   →   Additional 3.0%
    Remaining "bare PMID citations"          ~12,276 instances   →   91.7% uncovered
    

    4.2 Why This Matters More Than the Figure “254”

    Version 1 parsed only bibliographic lines, yielding 223 entries and 49 files. With the inclusion of footnotes (Version 2), this increased to 254 entries and 63 files—a 29% increase in file count, while footnotes covered only 3% of citations. Consequently, if the error rate in the uncovered portion aligns with that of the parsed section, the true total would be an order of magnitude larger than the current figure.

    One cannot assert that “there are over 2,000 errors” because the error rate in the uncovered portion remains unknown. However, it is certain that “254” represents a lower bound, and one is far from the true value.

    4.3 This Is the Second Occurrence of the Same Error Type

    The lesson from the previous round (Round 7) was that “conclusions must be based on ‘unextracted lines ≈ 0′”—the initial version missed 74 lines, nearly leading to the erroneous conclusion that “no issues exist within the pool.”

    In this round, using a different tool, the same pitfall reappeared: the 96.7% coverage gap in Version 1 was not flagged in any output. The tool ran silently, quietly reported “49 files / 223 entries,” and presented itself as a complete result.

    New Hard Rule R-191: Any tool performing “full-scan” operations must simultaneously output the denominator and coverage rate, explicitly stating the gap in the main output.
    Scanners that report only the numerator without the denominator create a false sense of certainty.


    5. Sampling Verification (Precision Validation)

    From a random sample of 14 T1 and 8 T2/X1 items out of 254 total, each was verified against the original record (random.seed fixed for reproducibility):

    5.1 T1 Sampling: 14 Items → All 14 Verified as Authentic

    PMID Declared Surname Actual First Author Actual Title (Verified via E-utilities)
    16243978 Targetti Hu G Isoflurane prevents platelets from enhancing neutrophil-induced coronary endothelial dysfunction
    14675133 Burch Liu CJ Polymorphism of the MICA gene and risk for oral submucous fibrosis
    17620642 Li Eisen RJ Residence-linked human plague in New Mexico: a habitat-suitability model (Declared as forest bathing NK activity)
    29537674 Chaplin Amagai T Accuracy of flow and sensors of indirect calorimetry for neonates and infants (Declared as “water memory” review)
    36509825 Zhang Wu LW Discrimination between possible sarcopenia and metabolic syndrome using the arterial pulse spectrum
    33903610 Hu Lin SK Classification of patients with Alzheimer’s disease using arterial pulse spectrum
    38942105 Liu Nishiwaki S Prevalence and characteristics of AVNRT with a bystander pathway
    32929968 Zhang Miller JR Particle-Based Reconfigurable Scattering Masks for Lensless Imaging (Declared as monsoon regime shift)
    27861248 Li Stoner L Reliability of pulse waveform separation analysis: effects of posture and fasting
    38070877 Zhang Williams CE The role of diet and host species in shaping seasonal dynamics of the gut microbiome (Declared as Five Phases-Six Qi-bacteria specific mechanism)
    39766014 Charu Abbas SR Federated Learning in Smart Healthcare: A Comprehensive Review on Privacy, Security
    33509706 Liu Cordier PY Complex and prolonged hypercoagulability in COVID-19 intensive care unit patients
    27861248 Wang Stoner L Same as above (Same PMID assigned to a different fake author in another file)
    18999346 Berman Hamilton JM Visuospatial deficits predict rate of cognitive decline in autopsy-verified dementia with Lewy bodies

    False Positive Rate: 0/14. Among these, 5 items (bolded) had actual titles completely unrelated to the declared theme (plague/neonatal calorimetry/lensless imaging/federated learning/Lewy body dementia), representing the most severe category of error.

    5.2 Reverse Verification: Of the 9 files from Round 7, 4 were not hit this round; were they missed?

    Round 7 listed 9 affected files. This round’s empirical test yielded 5 hits and 4 misses (programmatic comparison, not visual inspection):

    Round 7 File This Round Reason
    pulse-harmonic-random-forest-tcm-syndrome-prediction 17 Text unchanged, stable reproduction (and most severe)
    pulse-pressure-normalization-hri-algorithm 8 Stable reproduction
    pulse-harmonic-spectrum-individualized-zangfu-functional-scoring 5 Stable reproduction (T1)
    standing-wave-antinode-psychophysiology 3 Stable reproduction (T1)
    hri-c2-c3-heart-qi-deficiency-severity-dose-response 2 Stable reproduction: Liang et al. (PMID: 33903610) has actual first author Lin SK
    pulse-harmonic-gut-microbiome-organ-axis 0 Missed: Written as “2021 study published in Nature Medicine” (PMID: 34637011); surname falls outside 90-character window
    pool5-standby-validation-entry-exit-gate 0 Missed: - PMID 29269119: Cross-cultural RSA... bare list, no surname for judgment
    dosha-qi-rsa-falsification-framework-methodology 0 Missed: Same bare list format
    dosha-qi-rsa-falsification-preregistration-specification 0 Missed: Same bare list format

    5/9 Reproduced, 4/9 Missed. These 4 are not “files without issues,” but rather citation formats outside the detector’s recognized syntax—the direction of the miss is known and can be targeted (see §9, Item 1).

    5.3 Known Citation Formats Still Not Covered (Registered as Such)

    Format Instance Reason for Miss
    Bare list - PMID 29269119: Cross-cultural RSA psychometrics development review... No surname; T1 cannot be judged
    Chinese narrative followed by number ... (PMID: 34637011) shows TMAO levels correlate with... Surname falls outside 90-character window
    Multiple IDs in parallel (PMID: 35161551, 36509825) Only first taken; second not judged

    These three categories account for the majority of uncovered cases. The next step should be “reverse verification”: For each PMID, retrieve the real title via API, search for characteristic terms within the text; if not found, it suggests a fake bibliographic description may exist in the text—independent of surname syntax.

    5.4 Two Extreme Cases Missed but Immediately Identified Manually (Illustrating Miss Cost)

    After blind execution, citation-verify.sh was used to re-check PMIDs cited in newly written articles, accidentally encountering two cases missed by the detector but with severity far exceeding already detected items:

    Case 1: 34637011 (pulse-harmonic-gut-microbiome-organ-axis.md)

    Declared in text: 2021 study published in *Nature Medicine* (PMID: 34637011)
              showing positive correlation between TMAO levels and heart failure severity
    Actual Literature: Goldbach JT, et al. "The Development of Proud & Empowered:
              An Intervention for Promoting LGBTQ Adolescent Mental Health."
              Child Psychiatry Hum Dev. 2023
    
    • Theme: TMAO/Heart Failure vs LGBTQ Adolescent Mental Health Intervention
    • Journal: Nature Medicine vs Child Psychiatry Hum Dev
    • Year: 2021 vs 2023
    • Only the 8 digits match correctly

    Reason for Miss: The declaration lacked the “Surname + et al.” structure (wrote “2021 study published in Nature Medicine”), falling under Category 1 of §5.3.

    Case 2: 29269119 (3 Files, 3 Identities)

    File Declared Identity
    pool5-standby-validation-entry-exit-gate.md “Cross-cultural RSA psychometrics development review”
    dosha-rsa-qi-deficiency-dual-mapping.md § Title “Ram Manohar P et al., J Ayurveda Integr Med, 2018, HRV spectral features of the three doshas”
    Another location in same file “Shu Q et al., 2017, Acupuncture regulation of HRV in Qi-deficiency constitution”
    Third location in same file “Tripathi PK et al., 2011”
    Actual Ram Manohar P, An exploratory clinical study to determine the utility of heart rate variability analysis in the assessment of dosha imbalance, J Ayurveda Integr Med 2018

    One PMID was assigned three different first authors and three different years within the same file; one (Ram Manohar 2018) happens to be correct. This is no longer a “citation error,” but rather the same batch of materials being rewritten repeatedly, switching sources each time.

    One layer deeper (new subclass discovered this round): The other two PMIDs in the same file actually have correct themes:

    PMID Actual First Author Actual Title Relevance to “Acupuncture → Qi-deficiency HRV” Theme
    33068344 Song YJ Immediate effect of acupuncture and moxibustion at Guanyuan (CV4) and Zusanli (ST36) on heart rate variability (Zhongguo Zhen Jiu 2020) ✅ Highly relevant
    40315935 Li T Effects of acupuncture and moxibustion on heart rate variability in chronic fatigue syndrome patients (Complement Ther Med 2025) ✅ Highly relevant
    29269119 Ram Manohar P …utility of HRV analysis in the assessment of dosha imbalance* (J Ayurveda Integr Med* 2018) ✅ Relevant (Ayurvedic side)

    In other words: If 29269119 in the “Acupuncture → Qi-deficiency HRV” section were replaced by 33068344, the paragraph would hold.

    This reveals a previously unnamed failure subclass: “Swapping IDs rather than fabricating them”——
    The author possesses a batch of “seemingly relevant” real IDs; during writing, they attach the content of ID A to ID B.
    The distinction from “Real ID Fake Bibliography” is that the target theme already has a correct source in the pool; it is merely pointing at the wrong needle.
    For such errors, the remedy is to swap the ID, not delete the paragraph—disposal cost is far lower than the OMOM batch.

    Significance of these two cases: The detector currently only catches “fake bibliographies with surnames.” However, the most egregious fake bibliographies often omit names entirely (writing only “a study from a certain year and journal”), making them even harder to catch—as they save the very field needed for verification.

    6. Relationship with Existing Governance Cards

    P3-GOV-001 (Round 7, Levels L0–L3) does not require structural changes; however, the list of affected files must be updated:

    Level Original Scope Current Revision
    L0 New drafts do not cite contaminated literature Unchanged
    L1 9 files with banners + citation_status: disputed Expanded to 63 files (17 of which have ≥6 high-priority citations)
    L2 4 fake IDs + 2 false entries Unchanged, plus 113 high-reuse IDs from “inconsistent declarations” in PMID
    L3 OMOP 9 papers + 2 gut-heart axes Unchanged

    L1 offers improved cost-effectiveness: Adding 63 files, a single frontmatter field, and a banner remains a reversible, zero-deletion action that immediately halts misleading citations. However, expanding the scope from 9 to 63 files is no longer a task completable in a single iteration, requiring JOE to decide whether to proceed in batches or implement fully.

    7. Practical Implications for Research Conclusions (Retractions Required)

    A citation mismatch is not inherently a research error; however, if a specific numerical claim is attached to an irrelevant document, that claim lacks support.

    The most critical location identified in this round (first detected in Round 7, confirmed to recur in another file):

    pulse-pressure-normalization-hri-algorithm.md Line 56:
    "Pre-experimental data indicates that the G(P) slope of the thready pulse is approximately 1.7 times that of the floating pulse (PMID: 33901427)"
    

    The PMID 33901427 actually refers to Weaver 2021, a study on imaging in post-stroke cognitive impairment, and has no relation to pressure-harmonic gain curves. The figure “1.7 times” has no verifiable source in this corpus.

    Another category of structural issues: PMID 27861248 (True Source = Stoner L, Reliability of pulse waveform separation analysis: effects of posture and fasting) appears in 11 files within the corpus (verified via grep -l). While some entries correctly list Stoner as the first author, others incorrectly list Wang, Li, or Blacher. Having the same real document appear with different first authors across different articles is a typical manifestation of cross-subline misappropriation and fabricated bibliographic stacking.

    Other high-reuse PMIDs (precisely counted in this round, all hits):

    PMID True First Author Files Hit T1 Count False Last Names Assigned
    33903610 Lin SK 9 7 Hu / Liang / Wang / WK
    36294440 Chen CC 6 3 Wang / Xu / Zhang
    27861248 Stoner L 4 4 Blacher / Li / Wang
    27314975 Bilton K 3 3 Fukuda / Liu / Zhang
    36509825 Wu LW 3 1 Zhang
    33901427 Weaver NA 2 1 Zhang

    33903610 is the most typical case: a single PMID is written as four different first authors (Hu / Liang / Wang / WK) across 9 files. However, a counter-example must also be noted: in hri-c2-c3-lvef-joint-diagnostic-model.md and hri-harmonic-c2-c3-heart-qi-deficiency-severity-dose-response.md, the same PMID is correctly listed as Lin S.K. / Lin et al..

    This is the methodology established in Round 2 and verified again this round: A single PMID is correct in one article but incorrect in another.
    Checking only for the existence of a PMID yields all green; counting only how many times a PMID is cited reveals no obvious issues.
    One must drill down to “what specific claim in this article is supported by this PMID at this location.”

    7.1 Independent Signal Value of Cross-File Same-ID Different-Title (X1)

    X1 does not rely on any PubMed field; it performs only internal comparison within the corpus: if the same PMID is written with two non-overlapping titles in two different articles, at least one is incorrect—and often both are.

    Among the 28 groups of X1 identified in this round, the cleanest group is:

    PMID 33901427
    pulse-pressure-normalization-hri-algorithm.md:106
      → "Pressure-harmonic gain curves for six primary pulse types."   ❌
    tcm-pulse-hri-inter-operator-...md:147
      → "Strategic infarct locations for post-stroke cognitive impairment"  ✅
    

    The significance of X1 lies in its ability to run offline. On an isolated node, this is the only available pseudo-bibliographic signal.

    8. Scope Limitations (Boundary Preservation)

    • No modifications to audited documents: The 63 flagged files, skin-depth-reference-table.md, and the contaminated files in Pool ④ retain their original modification times (mtime).
    • No labels of “accepted” status were applied, nor were the values 779/782 written directly.
    • The count of “254 items” was not rephrased as “there are 254 errors within the pool”; instead, it was stated that “254 instances were detected, with the actual total being higher and currently unknown.”
    • No error rate extrapolation was performed for the 91.7% of data not covered; Section 4.2 merely noted that “if consistent, the magnitude would be tenfold” and explicitly marked this as a conditional hypothesis.

    9. Next Steps (Ranked by Cost-Effectiveness)

    # Action Cost Dependency
    1 Reverse Validator: Search for characteristic terms within real titles (covering bare lists and multiple numbers listed together) Medium Autonomous; dependent on next round
    2 Integrate this tool into the P3-GOV-001 L1 execution (63-file banner) Low Pending JOE approval
    3 Locate true sources for high-repetition numbers (27861248 appearing 11 times, 27314975 appearing 10 times, 33903610 appearing 8 times) Medium Autonomous
    4 Verify each claim of the “1.7 times” numerical assertion across the 8 high-frequency files Medium Requires reading full texts

    10. References (Methodological Anchors for This Round; Verified via E-utilities)

    1. Baethge C, et al. Citation inaccuracies in biomedical literature: a systematic review and meta-analysis. Res Integr Peer Rev. 2025. PMID: 40696441
    2. Mogull SA. Accuracy of cited “facts” in medical research articles: a methodological review. PLoS One. 2017. PMID: 28910404
    3. Montenegro TS, et al. Reference accuracy in spine surgery literature. J Neurosurg Spine. 2021. PMID: 32977312

    All three sources provide upstream evidence that citation inaccuracies are systematic rather than incidental, aligning with the concentrated distribution pattern observed in this round (63 files / 254 citations / 17 files with ≥6 citations).


    Pool 3 · #191 · 2026-10-18 · draft · unpublished · not accepted


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

  • 「用水十分钟,氡升一倍」:那条回落曲线测的是房子,不是水

    Evidence level: B

    你想量的是”水放出了多少氡”,拟合出来的却主要是”这间屋子每小时换几次气”。一条曲线上叠着两个完全不同的物理量,而它们长得一模一样。

    正文

    近水住宅室内氡偏高,已经有实测支撑:广东温泉酒店用水时段,浴室氡比无用水时段高 510%–1200%,卧室高 110%–410%(PMID 21382658)。于是”用水脉冲”成了很自然的量化目标——测一条曲线,看它升多高、多快回落,就能算出水的贡献率和时间常数。这个想法在实验设计上有五道坎,每一道都能把结论改向。

    第一道坎:回落的快慢,几乎全由换气决定。 氡-222 半衰期 3.82 天,对应衰变常数约 0.0076 h⁻¹;而实测住宅换气率在 0.12–3.46 h⁻¹,中位约 1.15 h⁻¹(PMID 22894100)。换气比放射性衰变快一两个数量级,所以回落段拟合出的时间常数 τ≈1/(λ+AER) 基本就是 1/AER。这一点在越南大叻的逐时监测里看得最清楚:同一房间在”加强自然通风”与”关闭”两种状态下,一区质量平衡拟合出的有效去除参数分别是 0.499 与 0.315 h⁻¹(下层)、0.338 与 0.213 h⁻¹(上层),浓度降了 64.3% 与 43.1%(PMID 42790328)。换句话说,你从回落速度里读不出水的任何特征,只能读出这栋房子有多漏风。

    第二道坎:仪器自己也有时间常数。 连续氡监测仪对瞬态浓度的响应从来不是即时的,七十年代就有专门讨论如何解释这类响应的论文(PMID 528213)。水中氡在线提取的三种装置差别更大:喷雾室与流通膜响应相近,而浸没式膜管对浓度变化的滞后显著更大(PMID 22835087)。若仪器 τ 与房间 τ 同量级,拟合出的衰减常数是两者的卷积——你测的是探测器,不是房子,更不是水。

    第三道坎:贡献率在数学上不可辨识。 一区质量平衡只有一个观测序列,却要同时估”水的源项”和”土壤气的源项”两条曲线,方程欠定。大叻那篇作者自己写得很直白:拟合出的有效源项 Eeff 只是”模型推断的源-去除平衡”,不是独立测量的源强(PMID 42790328)。所以可解释的”贡献率”必须靠外部约束换来——水龙头取样测水氡、关水期做基线、或土壤气独立监测。广东那项研究之所以能报出百分比,靠的是”用水时段 vs 无用水时段”的对照,不是模型分解。

    第四道坎:事件会重叠,基线还在自己晃。 先做个算术。一次八分钟淋浴约用 60 L 水,水氡 100 Bq/L 即带 6000 Bq,其中约 92% 释入浴室(PMID 24803512),落进 12 m³ 空间若瞬时混匀可到 460 Bq/m³ 量级;而按 1.15 h⁻¹ 的换气率,八分钟只换掉约 15%。所以峰值看得见,回落却要 50 分钟上下——五十分钟内洗第二个人,脉冲就叠成平台了。淋浴、洗衣四十分钟、做饭二十分钟在早晚集中出现,间隔短于 τ 时单事件 τ 根本拟合不出来。更要紧的是背景本身有节律:室内外压差梯度是房间”呼吸”的关键驱动,日周期由室内温度的昼夜循环触发,多日周期由风速造成的负压触发(PMID 29908412)。而用水高峰恰好落在温差的日周期高峰上——两个节律同相,时间上就分不开,只能靠随机化用水时刻的交叉设计,或按中断时间序列的规范事前指定影响模型、处理自相关与过度分散(PMID 27283160)。

    第五道坎:峰值浓度不等于峰值剂量。 剂量走的是子体与未附着份额那条路:中国实测未附着份额 9.3%–16.9%,却贡献 30.2%–46.2% 的剂量(PMID 23146814),且该份额本身随环境变化(PMID 27171653)。而室内平衡因子中位仅 0.3,且在 1–370 Bq/m³ 范围内子体/氡比基本不变、只轻微依赖通风(PMID 6094394)——这意味着浓度冲高时子体不会等倍跟上,短脉冲的剂量增量小于按峰值线性外推。室内空间差异同样不能忽略:同一宅内房间之间的剂量变异达 58%–61%,房与房之间 66%–74%(PMID 30747753)。反例更极端:罗马两户在春夏数小时内升到 20000 Bq/m³,峰值位置、时长与发生季节十年后复现一致,而季节校正因子会把这种低估放大(PMID 37019261)。

    要做事,就做五件。 ① 报告分钟级上升沿与峰值高度,回落 τ 只用来交叉验证换气率,不当”水的指纹”。② 仪器响应时间先用阶跃或脉冲注入标定,要求 τ_仪器 ≤ 1/5 τ_房间。③ 用水信号用流量计精确到分钟,不靠自报日志;条件允许就随机化用水时刻。④ 水氡冷热水分别取样作源项外部约束——注意热水端在淋浴过程中已损失约 92% 的氡(PMID 24803512)。⑤ 报剂量时走子体与未附着份额,并设阴性对照:无人用水期,氡应回落到土壤气基线。

    风水这边,《葬书》讲”气乘风则散,界水则止”,是把水当界面用——这是选址的空间判据,尺度是年。而水的物理作用在时间上恰恰相反:它不止气,它一次次把气放出来。两句并不冲突,冲突的是把”年尺度的选址判断”和”分钟尺度的事件脉冲”压成同一个平均浓度再互相比对。争论大多不是关于事实,是关于尺度。

    证据与局限

    已验证: 用水时段室内氡显著抬升,浴室 510%–1200%、卧室 110%–410%(PMID 21382658);浴室局部转移因子达 1/2300,淋浴过程水氡损失约 92%(PMID 24803512);一区质量平衡可拟合有效去除参数,但源项不可独立辨识(PMID 42790328);连续监测仪与在线提取装置存在响应滞后(PMID 528213, 22835087)。
    部分验证: 换气率量级与衰减常数的对应关系明确(PMID 22894100),但住宅换气率的个体差异极大,单点测量代表性有限;压差驱动”呼吸”的机制清楚(PMID 29908412),但用水与温差的同相混杂程度尚无实测分解。
    待验证: 单次淋浴事件的”贡献率”与”时间常数”在真实住宅中从未被去卷积分离;脉冲期剂量是否显著低于按峰值外推的估计,缺直接测量;季节校正因子在极端脉冲型住宅中的适用性未知。

    相关阅读

    证据等级: B
    参考文献: PMID 21382658, PMID 24803512, PMID 42790328, PMID 22894100, PMID 528213, PMID 22835087, PMID 29908412, PMID 27283160, PMID 23146814, PMID 27171653, PMID 6094394, PMID 30747753, PMID 37019261


    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。

    \”Ten Minutes of Water, Radon Doubles\”: The Declining Curve Measures the House, Not the Water

    Evidence level: B

    You are trying to measure “how much radon the water released,” but the fit primarily captures “how many times this room exchanges air per hour.” Two completely different physical quantities are superimposed on a single curve, yet they look identical.

    Main Text

    Indoor radon levels in homes near water are already supported by empirical data: during water usage periods at a hot spring hotel in Guangdong, radon levels in bathrooms were 510%–1200% higher than during non-usage periods, and in bedrooms 110%–410% higher (PMID 21382658). Consequently, the “water usage pulse” became a natural quantification target—measure a curve to see how high it rises and how fast it falls back, then calculate the water’s contribution rate and time constant. This idea faces five hurdles in experimental design, each capable of redirecting the conclusion.

    First Hurdle: The speed of decline is determined almost entirely by air exchange. Radon-222 has a half-life of 3.82 days, corresponding to a decay constant of approximately 0.0076 h⁻¹; meanwhile, measured residential air exchange rates range from 0.12–3.46 h⁻¹, with a median of 1.15 h⁻¹ (PMID 22894100). Air exchange is faster than radioactive decay by one or two orders of magnitude, so the time constant τ≈1/(λ+AER) fitted from the decline phase is essentially 1/AER. This is most clearly visible in hourly monitoring in Da Lat, Vietnam: in the same room under “enhanced natural ventilation” versus “closed” conditions, zone mass balance fitting yielded effective removal parameters of 0.499 and 0.315 h⁻¹ (lower level), and 0.338 and 0.213 h⁻¹ (upper level), with concentrations dropping by 64.3% and 43.1% respectively (PMID 42790328). In other words, you cannot read any characteristic of water from the decline speed; you can only read how leaky this building is.

    Second Hurdle: The instrument itself has a time constant. Continuous radon monitors never respond instantaneously to transient concentrations; papers discussing how to interpret such responses date back to the 1970s (PMID 528213). Differences are even greater among three online extraction devices for radon in water: spray chambers and flow-through membranes show similar responses, whereas immersed membrane tubes exhibit significantly larger lag against concentration changes (PMID 22835087). If the instrument τ is of the same order as the room τ, the fitted decay constant is a convolution of both—you are measuring the detector, not the house, and certainly not the water.

    Third Hurdle: Contribution rates are mathematically unidentifiable. Zone mass balance has only one observed sequence but must simultaneously estimate two curves: the “water source term” and the “soil gas source term,” making the system underdetermined. The authors of the Da Lat study wrote this very plainly themselves: the fitted effective source term Eeff is merely a “model-inferred source-removal balance,” not an independently measured source strength (PMID 42790328). Therefore, interpretable “contribution rates” must be obtained through external constraints—such as sampling water radon at faucets, establishing baselines during water-off periods, or independent soil gas monitoring. The study in Guangdong was able to report percentages not because of model decomposition, but due to the contrast between “water usage periods” and “non-usage periods.”

    Fourth Hurdle: Events overlap, and the baseline still fluctuates. Let’s do some arithmetic. An eight-minute shower uses approximately 60 L of water; at a water radon level of 100 Bq/L, this carries 6000 Bq, of which approximately 92% is released into the bathroom (PMID 24803512). If mixed instantaneously in a 12 m³ space, this reaches levels around 460 Bq/m³; however, with an air exchange rate of 1.15 h⁻¹, only about 15% is exchanged in eight minutes. Thus, the peak is visible, but the return takes roughly 50 minutes—if you shower a second person within those fifty minutes, the pulse stacks into a plateau. Showers, laundry (40 minutes), and cooking (20 minutes) cluster in the early and late evenings with intervals shorter than τ, making single-event τ fitting impossible. More importantly, the background itself has rhythms: indoor-outdoor pressure gradient is the key driver of the room’s “breathing,” triggered by diurnal temperature cycles and multi-day cycles caused by wind-induced negative pressure (PMID 29908412). Water usage peaks coincide with the peak of the diurnal temperature cycle—two rhythms in phase, they cannot be separated temporally—requiring instead a randomized cross-design of water usage times or pre-specifying impact models according to interrupted time series norms to handle autocorrelation and overdispersion (PMID 27283160).

    Fifth Hurdle: Peak concentration does not equal peak dose. Dose follows the path of daughters and unattached fraction: Chinese measurements show an unattached fraction of 9.3%–16.9%, yet it contributes 30.2%–46.2% of the dose (PMID 23146814), and this fraction itself varies with the environment (PMID 27171653). Meanwhile, the median indoor equilibrium factor is only 0.3, and within the range of 1–370 Bq/m³, the daughter-to-radon ratio remains essentially unchanged, depending only slightly on ventilation (PMID 6094394)—meaning daughters do not follow proportionally when concentration spikes, so dose increments for short pulses are less than linear extrapolation from peaks. Spatial differences within the same dwelling also cannot be ignored: dose variation between rooms in the same house reaches 58%–61%, and between houses 66%–74% (PMID 30747753). Counter-examples are even more extreme: in Rome, two households rose to 20,000 Bq/m³ over several hours in spring and summer; the peak position, duration, and seasonal occurrence repeated consistently ten years later, while seasonal correction factors would amplify this underestimation (PMID 37019261).

    To do things correctly, do five things. ① Report minute-level rise slopes and peak heights; use return τ only to cross-verify air exchange rates, not as a “water fingerprint.” ② Calibrate instrument response time first using step or pulse injection, requiring τ_instrument ≤ 1/5 τ_room. ③ Use flow meters to record water usage signals precisely to the minute, not relying on self-reported logs; randomize water usage times if conditions allow. ④ Sample cold and hot water separately as external constraints for source terms—note that approximately 92% of radon is lost at the hot water end during showering (PMID 24803512). ⑤ Report dose via daughters and unattached fractions, and set a negative control: during periods with no water use, radon should return to the soil gas baseline.

    Regarding Feng Shui, Zang Shu states “Qi scatters when wind blows, but stops when water bounds it,” treating water as a boundary—this is a spatial criterion for site selection on an annual scale. Conversely, the physical action of water operates temporally in the opposite direction: it is not just Qi; it repeatedly releases Qi. These two statements do not conflict; the conflict arises when “annual-scale site selection judgments” and “minute-scale event pulses” are compressed into the same average concentration for comparison. Most debates are not about facts, but about scale.

    Evidence and Limitations

    Verified: Indoor radon levels rise significantly during water usage (bathrooms 510%–1200%, bedrooms 110%–410%) (PMID 21382658); local transfer factors in bathrooms reach 1/2300, with approximately 92% radon loss during showering (PMID 24803512); zone mass balance can fit effective removal parameters, but source terms remain unidentifiable independently (PMID 42790328); continuous monitors and online extraction devices exhibit response lag (PMID 528213, 22835087).
    Partially Verified: The correspondence between air exchange rate magnitude and decay constants is clear (PMID 22894100), but individual variation in residential air exchange rates is extreme, limiting representativeness of single-point measurements; the mechanism of pressure-driven “breathing” is clear (PMID 29908412), but the degree of confounding between water usage and temperature cycles lacks empirical decomposition.
    To Be Verified: The “contribution rate” and “time constant” of single shower events have never been deconvoluted in real residences; whether pulse-period dose is significantly lower than estimates extrapolated from peaks lacks direct measurement; applicability of seasonal correction factors in extreme pulse-type residences remains unknown.

    Related Reading

    Evidence Level: B
    References: PMID 21382658, PMID 24803512, PMID 42790328, PMID 22894100, PMID 528213, PMID 22835087, PMID 29908412, PMID 27283160, PMID 23146814, PMID 27171653, PMID 6094394, PMID 30747753, PMID 37019261


    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.

  • 「控制通风之后评级就失效了」:这句话既能证明风水没用,也能证明风水说对了

    Evidence level: B

    归因不是做减法。把中介变量放进模型让评级系数掉下来,这一步在统计上谁都做得到;难的是判断掉下来的那部分到底该记在谁头上——而这件事,输出表本身不会告诉你。

    正文

    风水里最实在的一句话是「藏风聚气」:藏风是不让风吹散,聚气是让气留下。翻成现代语言就是通风。所以问「风水评级凭什么预测室内氡」,第一个该问的不是风水是不是迷信,而是它自己说路径是什么——它说走风、走气。这给出可检验的中介路径。但检查中介比检查关联难一个量级:难的不是算,是算出的数还配不配叫那个名字。

    第一层:中介不是一个数,是一族数。

    因果中介的定义预设了中介能被「设成某个值」(PMID 22475830)。可「把通风率设成 0.5 次每小时」有无数种做法——开窗、开抽风机、堵门缝、换窗型,改的是不同通道,对氡影响不同。算出的间接效应确有解释,但解释的是某一种设值方式,不是「通风」本身。设值不可行得换成「有机效应」(PMID 27229743)。更前置一步是把因果问题写成可核对的估计目标:处理、对照、干预策略、汇总度量(PMID 41974452)。不做这步,「中介比例」是空的。

    第二层:通风、温差、气压差,是同一个物理量的三张脸。

    三者物理上不是并列,是串联。室内外温差驱动烟囱效应,烟囱效应决定渗入率:实测土壤气扩散进入 5.0±0.8 Bq·m⁻³·h⁻¹,平流可达 ±4.0,纯烟囱效应下渗入率仅 0.7(±0.5)×10⁻¹ h⁻¹;通风率低或断层活动时,土壤氡进入能抬一个数量级(PMID 38372915)。高层住宅无控通风下温差显著影响换气率(PMID 33711618);节能建筑换气率极低,直接导致氡累积(PMID 25977350);同一住宅群六栋房就分出三种通风行为(PMID 32291452)。

    把三样一起当中介,等于把一个变量拆成三个互相解释的副本。六类调整方法给出六套结果,多数研究连该不该相互调整都没想清(PMID 40038753)。比例被三张脸瓜分,分给谁看测量精度,不看物理。

    第三层:误分类专挑中介那条路打。

    模拟研究显示:暴露误分类让总效应与直接效应都向零偏,但「超额 OR 的减少百分比」几乎不受影响;反过来,中介误分类会迅速把间接效应压向下偏(PMID 23386673)。另一条路反向:当中介本身是暴露的生物标志物时,暴露误分类会低估直接效应、高估间接效应(PMID 28234025)。连续暴露误差下偏倚两边走(PMID 36869626)。

    翻译成本案:通风测不准,中介比例就必然被压低;这低值像「通风只能解释一部分」,其实只是「通风没测准」。 通风率标准测法是示踪气体衰减,贵且只能做几次(PMID 41273441);同一户氡的年际变异系数就有 17%(PMID 33418274)。只测几次的量去解释年年抖的量,精度账先就亏了。

    第四层:中间混淆让自然效应识别不了,而「过度校正」与「归因」在数据上长得一样。

    自然直接与间接效应的识别条件明确排除「存在中间混淆」;有中间混淆时只能退回结构方程那套带参数假设的做法(PMID 25504026)。本案中间混淆几乎一定存在:有没有人常住、装修史、地基改没改过,同时决定通风行为与氡的进入。不建模,比例就是错的。

    更麻烦的是相反方向。如果评级效力本来就在于「选了个通风好的地形」,那「控制通风后评级失效」不是评级没用,是你把路径掐了。这是过度校正——不一定向零偏,可远离零、甚至反号(PMID 42821507)。175 篇空气污染队列里 69.7% 有潜在过度校正,68.6% 调了中介(PMID 41671745)。

    于是出现一个尴尬局面:「控制后消失」和「归因于第三变量」在输出表上是同一行数,因果含义却相反。 分辨只能靠事前画图,不能靠事后看 p 值。

    第五层:比例是最不该报的统计量,而「藏风」本来就不是越大越好。

    四种效应量比较说得很清楚:标准化效应量在单、双中介模型里相对无偏且有效,比例型、比值型在功效、一类错误率、覆盖率上全面更差(PMID 28342072)。「中介比例 38%」恰恰是最不稳的写法。

    还有一层标准中介抓不到:风水讲的从来不是「通风越大越好」。通风太大,气散了;太小,气浊了。藏风要的是有边界的通风——有最优区间的量,对应非线性中介(阈值或倒 U)。线性模型把曲线压成一条斜率,然后说「不显著」。

    判决五件事。 ① 先写估计目标,四要素齐(PMID 41974452);中介落成可实施操作(窗开时长、机械通风时长、换气率),三套分开跑。② 三变量单跑与联合跑各一遍,比比例怎么被瓜分(PMID 40038753)。③ 中间混淆显式建模,用不依赖「排除中间混淆」的方法(PMID 25504026),并算 E-value——抹掉间接效应要多强的未测混淆(PMID 28693043)。④ 报标准化间接效应加 bootstrap 区间,不报比例。⑤ 加阴性对照:室内 CO₂ 是通风的忠实代理却无致癌通路,评级应当能预测它(PMID 42731630、42578165)。

    这套设计不是来拆穿风水的。 恰恰相反——风水从不声称绕过风直接作用于氡,它的路径就是藏风聚气。那么「控制通风后评级失效」读作——信息全部经由通风传递,中介比例接近 100%。这不是否定,是把风水写下的机制说明逐条对上。真正会否掉风水的,是另一种结果:评级预测氡,但通风、温差、气压都解释不了,减掉后还剩一大块。那时才须承认——要么风水看到了我们还没测的东西,要么这关联根本是别的变量在冒充。

    证据与局限

    已验证:
    – 中介定义预设可设值的干预,多版本中介下间接效应只对应某一种设值方式(PMID 22475830、27229743)
    – 中介误分类会把间接效应压向下偏,暴露误分类则可能高估间接效应(PMID 23386673、28234025)
    – 通风率、室内外温差、渗入率在物理上串联,实测渗入率与烟囱效应量级已量化(PMID 38372915、33711618、25977350)
    – 自然效应识别排除中间混淆,存在中间混淆须改用其他估计量(PMID 25504026)
    – 过度校正方向不限于向零,且在文献中普遍存在(PMID 42821507、41671745、42826927)
    – 比例型效应量性能劣于标准化效应量(PMID 28342072)
    – 阴性对照结局可用于校准残余混淆与检测偏倚(PMID 42731630、42578165)

    部分验证:
    – 通风率可用示踪气体衰减法实测,但只能做有限次 campaign(PMID 41273441)
    – 同一住宅长期氡的年际变异系数约 17%,测量不确定性可估(PMID 33418274);但该值来自单一地区样本,外推须谨慎
    – 同一住宅群内通风行为差异可分出类型(PMID 32291452),但类型划分依赖住户自报与短期观测

    待验证:
    – 本设计中「评级→微气候→氡」的中介比例到底是多少(本设计核心)
    – 中介是否为非线性(阈值或倒 U),线性模型会否系统性低估中介效应
    – 微气候三变量各自的中介份额能否被数据分开,还是天然共线不可分

    相关阅读

    证据等级: B
    参考文献: PMID 22475830, PMID 27229743, PMID 41974452, PMID 41273441, PMID 36869626, PMID 23386673, PMID 28234025, PMID 38372915, PMID 33711618, PMID 25977350, PMID 40038753, PMID 25504026, PMID 28342072, PMID 42821507, PMID 41671745, PMID 42826927, PMID 42731630, PMID 42578165, PMID 28693043, PMID 32291452, PMID 33418274


    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。

    \”Rating Becomes Invalid After Controlling Ventilation\”: This Statement Proves Feng Shui Is Useless Yet Also Validates Feng Shui

    Evidence level: B

    Attribution is not subtraction. Inserting a mediator variable into the model to reduce the rating coefficient is statistically trivial; the difficulty lies in determining whether the reduced portion should be attributed to the mediator or other factors—something the output table itself cannot tell you.

    Main Text

    The most substantive statement in Feng Shui is “Cangfeng Juyi” (Hide Wind, Gather Qi): Cangfeng prevents wind from dispersing, while Juyi retains the Qi. Translated into modern language, this is ventilation. Therefore, when asking “On what basis does Feng Shui rating predict indoor radon,” the first question should not be whether Feng Shui is superstition, but what mechanism it proposes—it claims the path involves wind and Qi. This provides a testable mediation pathway. However, checking mediation is an order of magnitude harder than checking association: the difficulty is not calculation, but whether the calculated numbers justify the label.

    Layer 1: Mediation is not a single number, but a family of numbers.

    The definition of causal mediation presupposes that the mediator can be “set to a specific value” (PMID 22475830). However, “setting the ventilation rate to 0.5 times per hour” can be achieved in infinitely many ways—opening windows, running exhaust fans, sealing gaps, changing window types. These alter different channels with varying impacts on radon. While the calculated indirect effect offers an explanation, it explains a specific set-valued approach, not “ventilation” itself. Setting values is not feasible without converting the causal question into a verifiable estimation target: treatment, control, intervention strategy, and summary measure (PMID 41974452). Without this step, the “mediation proportion” is empty.

    Layer 2: Ventilation, temperature difference, and pressure difference are three faces of the same physical quantity.

    Physically, these three are not parallel but serial. Indoor-outdoor temperature difference drives the chimney effect, which determines infiltration rates: measured soil gas diffusion infiltration is 5.0±0.8 Bq·m⁻³·h⁻¹, advection reaches ±4.0, while under pure chimney effect infiltration is only 0.7(±0.5)×10⁻¹ h⁻¹; with low ventilation or fault activity, soil radon entry can increase by an order of magnitude (PMID 38372915). In high-rise apartments without controlled ventilation, temperature difference significantly affects air exchange rates (PMID 33711618); energy-efficient buildings have extremely low air exchange rates, directly leading to radon accumulation (PMID 25977350); within the same residential complex, six buildings exhibit three distinct ventilation behaviors (PMID 32291452).

    Treating all three as mediators simultaneously is equivalent to splitting one variable into three mutually explanatory copies. Six adjustment methods yield six sets of results, with most studies failing to clarify whether they should adjust for one another (PMID 40038753). The proportion is divided among the three faces, allocated based on measurement precision rather than physics.

    Layer 3: Misclassification specifically targets the mediation pathway.

    Simulation studies show: exposure misclassification biases both total and direct effects toward zero, but the “percentage reduction in excess OR” remains largely unaffected; conversely, mediator misclassification rapidly biases the indirect effect downward (PMID 23386673). The reverse pathway occurs when the mediator itself is a biomarker for exposure: exposure misclassification underestimates the direct effect but overestimates the indirect effect (PMID 28234025). Under continuous exposure error, bias moves in both directions (PMID 36869626).

    Translated to this case: If ventilation is measured inaccurately, the mediation proportion will inevitably be suppressed; this low value appears as “ventilation explains only part of the effect,” but is actually just “ventilation was not measured accurately.” Ventilation rate standard measurement uses tracer gas decay, which is expensive and limited to few campaigns (PMID 41273441); meanwhile, the interannual coefficient of variation for radon in a single household is 17% (PMID 33418274). Using a measured quantity with few observations to explain a fluctuating annual quantity incurs a precision deficit from the start.

    Layer 4: Intermediate confounding prevents identification of natural effects, while “over-adjustment” and “attribution” look identical in data.

    Conditions for identifying natural direct and indirect effects explicitly exclude the existence of intermediate confounding; when intermediate confounding exists, one must revert to structural equation approaches with parametric assumptions (PMID 25504026). Intermediate confounding almost certainly exists in this case: factors such as occupancy history, renovation history, and foundation modifications simultaneously determine ventilation behavior and radon entry. Without modeling, the proportion is incorrect.

    More麻烦的是 the opposite direction. If rating validity inherently relies on “selecting terrain with good ventilation,” then “rating becomes invalid after controlling for ventilation” does not mean the rating is useless, but that you have severed the pathway. This is over-adjustment—not necessarily biased toward zero, but potentially far from zero or even reversing the sign (PMID 42821507). In 175 air pollution cohorts, 69.7% had potential over-adjustment, and 68.6% adjusted for mediators (PMID 41671745).

    Thus arises an awkward situation: “Disappearing after control” and “Attribution to a third variable” appear as the same row in the output table, yet have opposite causal meanings. Differentiation relies only on prior plotting, not post-hoc p-values.

    Layer 5: Proportion is the least desirable statistic to report, and “Cangfeng” is not inherently better with greater magnitude.

    Four types of effect sizes clearly demonstrate: standardized effect sizes are relatively unbiased and efficient in single and dual mediator models, whereas proportion-based and ratio-based metrics perform worse across power, Type I error rates, and coverage (PMID 28342072). The phrase “mediation proportion of 38%” is precisely the least stable formulation.

    Another standard mediation is missed: Feng Shui never claims “greater ventilation is always better.” Excessive ventilation disperses Qi; insufficient ventilation makes it turbid. Cangfeng requires bounded ventilation—an optimal interval corresponding to nonlinear mediation (threshold or inverted U). Linear models flatten the curve into a single slope, then declare it “non-significant.”

    Five judgments. ① First, define the estimation target with four elements aligned (PMID 41974452); ensure mediation is operationalizable (window opening duration, mechanical ventilation duration, air exchange rate), running three distinct sets. ② Run single-variable and joint-variable analyses separately to compare how the proportion is divided (PMID 40038753). ③ Explicitly model intermediate confounding using methods not relying on “excluding intermediate confounding,” and calculate the E-value—how strong unmeasured confounding would need to be to nullify the indirect effect (PMID 28693043). ④ Report standardized indirect effects with bootstrap intervals, not proportions. ⑤ Add a negative control: indoor CO₂ is a faithful proxy for ventilation but lacks carcinogenic pathways; ratings should predict it (PMID 42731630, 42578165).

    This design is not intended to debunk Feng Shui. On the contrary—Feng Shui never claims to bypass wind to directly act on radon; its mechanism is precisely Cangfeng Juyi. Therefore, “rating becomes invalid after controlling for ventilation” reads as: information flows entirely through ventilation, implying a mediation proportion near 100%. This is not a denial, but a verification of each mechanism written by Feng Shui. What would truly invalidate Feng Shui is a different result: ratings predict radon, but ventilation, temperature difference, and pressure difference fail to explain it; after subtracting these, a large residual remains. Only then must one admit—either Feng Shui observed something we have not yet measured, or the association is merely a variable impersonating another.

    Evidence and Limitations

    Verified:
    – Mediation definitions presuppose intervenable set-values; indirect effects under multiple mediation versions correspond only to specific set-valued approaches (PMID 22475830, 27229743)
    – Mediator misclassification biases indirect effects downward, while exposure misclassification may overestimate indirect effects (PMID 23386673, 28234025)
    – Ventilation rate, indoor-outdoor temperature difference, and infiltration rates are physically serial; measured infiltration rates and chimney effect magnitudes are quantified (PMID 38372915, 33711618, 25977350)
    – Natural effect identification excludes intermediate confounding; existence of intermediate confounding requires alternative estimators (PMID 25504026)
    – Over-adjustment direction is not limited to bias toward zero and is prevalent in literature (PMID 42821507, 41671745, 42826927)
    – Proportion-based effect sizes perform inferiorly compared to standardized effect sizes (PMID 28342072)
    – Negative control outcomes can be used to calibrate residual confounding and detect bias (PMID 42731630, 42578165)

    Partially Verified:
    – Ventilation rates can be measured via tracer gas decay methods but are limited to finite campaigns (PMID 41273441)
    – The interannual coefficient of variation for radon in a single household is approximately 17%, with estimable measurement uncertainty (PMID 33418274); however, this value derives from a single regional sample and requires caution in extrapolation
    – Ventilation behavior differences within the same residential complex can be categorized (PMID 32291452), but classification relies on self-reported data and short-term observation

    To Be Verified:
    – The actual mediation proportion in the proposed design (“Rating → Microclimate → Radon”) (Core of this study)
    – Whether mediation is nonlinear (threshold or inverted U); whether linear models systematically underestimate mediation effects
    – Whether data can separate the mediation shares of the three microclimate variables, or if they are naturally collinear and inseparable

    Related Reading

    Evidence Level: B
    References: PMID 22475830, PMID 27229743, PMID 41974452, PMID 41273441, PMID 36869626, PMID 23386673, PMID 28234025, PMID 38372915, PMID 33711618, PMID 25977350, PMID 40038753, PMID 25504026, PMID 28342072, PMID 42821507, PMID 41671745, PMID 42826927, PMID 42731630, PMID 42578165, PMID 28693043, PMID 32291452, PMID 33418274


    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.

  • 两把尺子量同一口井:峦头派和理气派,谁的评级更接近你家的氡

    Evidence level: B

    头对头比较两派的预测力,考的不是”谁更准”,而是”在低信度、低天花板、高共线这三重条件下,这两套尺子到底能不能被数据分开”。

    正文

    风水内部吵了上千年:看地该看形还是看方向。峦头派说山形水势决定一切,理气派说坐向与元运才是关键。要把这场争论从”信谁”变成”测谁”,最直接的设计是:同一批住宅,两派风水师各自独立盲评,同时测室内氡,看谁的评级更能预测氡。

    但一动手就会发现五个坑。

    一、两派测的可能不是同一个东西。 峦头派的输入是地形:坡向、坡度、山体走势、距水体远近——这条路径通向地质与土壤气,因为地形决定了母岩、覆盖层和地下水。理气派的输入是方位:坐向、门向、楼层——这条路径通向风压与日照,进而影响换气次数,再影响氡在室内的累积。两条路径终点都是”室内氡”,但中间隔着一整套不同物理过程。头对头比较因此不是”两个同题答案谁分高”,而是”两条因果路径在同一终点上的相对强度”。

    二、问”谁更准”之前,得先问”它们是不是同一把尺”。 如果两派评出来的吉凶高度一致,那这场比较其实没有意义——你只是在给同一个变量取两个名字。如果不一致,那”哪个更准”就依赖样本里哪一派更”对口”,换个地区结论可能就翻。麻烦的是,衡量一致性的 kappa 本身会被扭曲:同一批评级,只因为”凶宅”比例变了,kappa 就能显著漂移 [PMID 8501467]。所以判一致性不能只报 kappa,还须同时报偏倚与患病率指数。

    三、NRI 会自己变正。 很多人比较增量预测力会用净重分类改善(NRI)。但 Pepe 等人已经证明:即便新加入的标志物毫无预测信息,在大测试集上算出的 NRI 也常常是正的——因为它对基线模型拟合不佳极其敏感 [PMID 26504496]。Kerr 的系统梳理和 Leening 的临床指南给出了同一个提醒:NRI 的解释门槛比想象中高得多 [PMID 24240655][PMID 24592497]。这意味着,拿 NRI 宣布”峦头派增量更大”,很可能只是基线模型没校准好。

    四、信度不等,衰减就不等。 两派的评分误差不会一样大。地形判断留给评者的自由度更大,方位读数则接近机械可复现——这就导致峦头派的评级信度可能更低。而信度低的变量,其真实关联会被系统性地稀释,这叫回归稀释偏倚;UK Biobank 对 2858 个重复测量变量的分析显示,ICC 从 0.29 到 0.85 不等,校正后关联普遍变强 [PMID 37336529]。更麻烦的是,当协变量本身带测量误差且与预测变量相关时,简单校正反而会引入新的正偏 [PMID 38721135]。换句话说,“哪派更准”可能只是”哪派更好测”。

    五、天花板会压扁一切差异。 室内氡能被地理或建筑变量解释的比例本来就不高。瑞士用核估计做全国室内氡制图,模型只解释了 28% 的变异,且不同建筑类型(独立混凝土基础与土基农舍)会给出系统性不同的图 [PMID 25314691]。葡萄牙 416 间教室研究里,约一半室内水平可归因于房间与建筑特征——位置、主建材、楼层、管理类型 [PMID 37084905]。宾州 86 万条实测里,地层是最强预测因子(中位数 365 对 93 Bq/m³),但井水(高 21%)、社区类型、天气各有独立贡献 [PMID 25856050]。两派评级的 ΔAUC 之差,很可能整个落在这片残余变异的噪声里。室内氡本身空间变异又很大 [PMID 30645746][PMID 33130895],季节修正因子还存在争议 [PMID 26093853];测量协议不统一,比较就退化成协议比较。

    所以正确的判决不是”哪派 AUC 大”,而是四格表里的哪一种。 在基线模型(地质、建筑、季节)已校准、独立测试集上,做两次交叉增量:加峦头后理气是否还有增量,加理气后峦头是否还有增量。

    • 互加对方后各自增量都消失 → 两派冗余,测的是同一个底层变量,只是命名不同;
    • 只有一方保留增量 → 那一方确实捕捉了另一条路径;
    • 两者各自保留增量 → 真的存在两条独立通道(地形→土壤气、方位→通风),这才是有价值的结果;
    • 对基线两者都无增量 → 在这个终点上,两套评级都不携带地质与建筑之外的信息。

    技术上还有三件必须做的事:用 DeLong 检验比较相关 ROC 曲线(同一批住宅、两套评级,曲线相关)[PMID 3203132];用决策曲线净收益代替单纯的重分类指标 [PMID 26810254];样本量按”能分辨的最小 ΔAUC”反推,并考虑空间自相关把有效样本量打折。风水实证研究的系统综述已经指出,这个领域从未把评级与有健康后果的物理暴露做过头对头盲法实测 [PMID 37809664]——两派之争,一直停在口头上。

    古人对这件事其实有过一个说法:峦头为体,理气为用。翻译成统计语言,”体”是时间不变的骨架——地形、母岩、覆盖层,是基线模型里最粗的那几根柱子;”用”是随时间与方向变化的修正项——坐向决定风怎么走,元运决定什么时候该调。这句话的含义不是”两派都重要”之和稀泥,而是:它们不是并列的两个候选变量,而是一个主效应加一个调节项。若真如此,正确检验就不是比谁 AUC 高,而是检验”用”的增量是否只在特定的”体”上出现——山形逼仄处,风向的修正空间很小;地势开阔处,坐向才真正决定换气。这是一个交互项,不是一场擂台赛。

    证据与局限

    已验证:
    – 室内氡可由地质、井水、社区、天气等多因素预测,单一地理变量解释力有限(宾州 86 万条)[PMID 25856050]
    – 地理与建筑变量对室内氡的解释力存在明显天花板(瑞士模型 28%)[PMID 25314691]
    – 建筑与房间特征可解释约一半室内氡水平(416 间教室)[PMID 37084905]
    – NRI 在无信息标志物下也可能为正,不宜单独用于增量判断 [PMID 26504496][PMID 24240655][PMID 24592497]
    – 比较相关 ROC 曲线需用 DeLong 方法 [PMID 3203132]
    – kappa 受偏倚与患病率影响,须同时报告 [PMID 8501467]
    – 测量误差导致的关联衰减可校正,但校正方法在非经典误差下会引入新偏倚 [PMID 37336529][PMID 38721135]

    部分验证:
    – 风水评级的评分者间信度有系统综述支持,但信度不等于效度 [PMID 37809664]
    – 室内氡空间变异大、季节修正因子存在争议,协议统一是前提 [PMID 30645746][PMID 33130895][PMID 26093853]
    – 方位→通风→氡这条路径有物理基础,但缺针对”坐向”这一具体变量的实测

    待验证:
    – 两派评级的交叉增量结构(冗余 / 单方独立 / 双方独立 / 双无效)
    – 峦头与理气的信度差是否会系统性低估前者
    – “峦头为体、理气为用”是否表现为显著的流派×地形交互项

    相关阅读

    证据等级: B
    参考文献: PMID 25856050, PMID 25314691, PMID 37084905, PMID 26504496, PMID 24240655, PMID 24592497, PMID 3203132, PMID 8501467, PMID 37336529, PMID 38721135, PMID 30645746, PMID 33130895, PMID 26093853, PMID 26810254, PMID 37809664, PMID 26706761, PMID 39292674, PMID 27343029


    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 Rulers Measuring the Same Well: The Luantou School and the Liqi School, Whose Rating Is Closer to Your Home’s Radon

    Evidence level: B

    A head-to-head comparison of predictive power does not test “who is more accurate,” but rather asks: “Under conditions of low reliability, low ceiling explanatory power, and high collinearity, can these two frameworks be statistically distinguished by data?”

    Main Text

    There has been a debate within Feng Shui for over a millennium: should site assessment prioritize form or direction? The Luantou school argues that landforms dictate everything, while the Liqi school asserts that orientation and elemental cycles are paramount. To transform this debate from “who to trust” into “what to measure,” the most direct design is to use a cohort of identical residences where two independent Feng Shui practitioners rate each site blindly, while simultaneously measuring indoor radon levels. The goal is to see which rating system better predicts radon concentrations.

    However, initiating this analysis reveals five critical pitfalls.

    1. The two schools may be measuring fundamentally different constructs. Luantou inputs are topographic: slope aspect, gradient, mountain走势 (trend), and distance to water bodies. This pathway links to geology and soil gas, as terrain dictates parent rock, cover layers, and groundwater levels. Liqi inputs are directional: facing, door orientation, and floor level. This pathway links to wind pressure and solar radiation, which influence air exchange rates and subsequently affect radon accumulation indoors. While both pathways converge on “indoor radon,” they are separated by distinct physical processes. Therefore, a head-to-head comparison is not “which ruler gives the higher score for the same question,” but rather “the relative strength of two causal pathways converging at the same endpoint.”

    2. Before asking “who is more accurate,” ask “are they using the same ruler?” If the auspicious/inauspicious ratings from both schools show high consistency, the comparison is meaningless—you are merely assigning two names to the same variable. If they differ, “accuracy” depends on which school aligns better with the specific sample; conclusions may flip across regions. Complicating this is the measurement of consistency itself: the kappa coefficient can be significantly distorted by changes in the proportion of “inauspicious” cases within the same rating set [PMID 8501467]. Thus, reporting consistency should not rely solely on kappa; one must also report bias and prevalence indices.

    3. NRI can become positive even without new information. Many comparisons of incremental predictive power utilize Net Reclassification Improvement (NRI). However, Pepe et al. have demonstrated that even if a new marker possesses no predictive information, the calculated NRI on a large test set is often positive due to extreme sensitivity to poor baseline model fit [PMID 26504496]. Kerr’s systematic review and Leening’s clinical guidelines offer a similar warning: the interpretive threshold for NRI is far higher than assumed [PMID 24240655][PMID 24592497]. Consequently, declaring “Luantou provides greater incremental value” based on NRI may simply reflect a poorly calibrated baseline model.

    4. Reliability varies, and thus attenuation varies. Rating errors will not be equal across schools. Topographic judgment allows raters greater freedom, whereas directional readings approach mechanical reproducibility. This suggests Luantou ratings may have lower reliability. Variables with low reliability suffer from systematic dilution of true associations, known as regression dilution bias. Analysis of 2,858 repeated measures in the UK Biobank showed ICC values ranging from 0.29 to 0.85, with corrected associations generally strengthening [PMID 37336529]. More problematic is that when covariates carry measurement error and are correlated with the predictor, simple correction can introduce new positive bias [PMID 38721135]. In short, “which school is more accurate” may simply mean “which is easier to measure reliably.”

    5. The ceiling effect flattens all differences. Indoor radon is inherently difficult to explain via geographic or architectural variables alone. Switzerland used nuclear estimates for national indoor radon mapping; the model explained only 28% of the variance, with different building types (independent concrete foundations vs. earth-sheltered farmhouses) yielding systematically different maps [PMID 25314691]. In a study of 416 classrooms in Portugal, approximately half the indoor levels were attributable to room and building characteristics—location, primary materials, floor level, and management type [PMID 37084905]. Among 860,000 measured entries in Pennsylvania, geology was the strongest predictor (median 365 vs. 93 Bq/m³), but well water (high in 21%), community type, and weather each contributed independently [PMID 25856050]. The difference in ΔAUC between the two schools likely falls entirely within the noise of this residual variance. Indoor radon itself exhibits high spatial variability [PMID 30645746][PMID 33130895], and season adjustment factors remain controversial [PMID 26093853]; without standardized measurement protocols, the comparison devolves into a comparison of protocols.

    Therefore, the correct verdict is not “which school has the higher AUC,” but which cell in the 2×2 table applies. On an independent test set with a calibrated baseline model (geology, architecture, seasonality), perform two cross-incremental analyses: does adding Luantou to Liqi yield further increment, and vice versa?

    • If increments vanish for both after adding the other → The schools are redundant, measuring the same underlying variable with different labels.
    • If only one retains an increment → That school captures a distinct pathway.
    • If both retain increments → True independent channels exist (terrain→soil gas, orientation→ventilation); this is a valuable result.
    • If neither adds increment to the baseline → Neither rating carries information beyond geology and architecture for this endpoint.

    Technically, three additional steps are required: use DeLong’s test to compare correlated ROC curves (same residences, two rating sets, assessing curve correlation) [PMID 3203132]; use net benefit from decision curves instead of purely reclassification metrics [PMID 26810254]; and calculate sample size based on the minimum resolvable ΔAUC, adjusting for spatial autocorrelation to reduce effective sample size. Systematic reviews of empirical Feng Shui research have noted that this field has never conducted head-to-head blind实测 (measurement) against physical exposures with health consequences [PMID 37809664]—the debate remains purely theoretical.

    Historically, there was a formulation regarding this matter: “Luantou is the substance (Ti), Liqi is the function (Yong).” Translated into statistical language, “substance” refers to the time-invariant skeleton—terrain, parent rock, cover layers, forming the coarsest pillars in the baseline model; “function” refers to correction terms varying with time and direction—orientation dictates wind flow, elemental cycles dictate when adjustments are needed. The meaning is not a muddying of waters by summing “both are important,” but rather: “They are not two parallel candidate variables, but a main effect and a moderator.” If true, the correct test is not comparing AUCs, but verifying whether the “function” increment appears only on specific instances of “substance”—where terrain is constricted, the space for wind correction is small; where terrain is open, orientation truly determines air exchange. This is an interaction term, not a knockout tournament.

    Evidence and Limitations

    Verified:
    – Indoor radon can be predicted by multiple factors (geology, wells, community, weather), but single geographic variables have limited explanatory power (Pennsylvania 860k records) [PMID 25856050]
    – Geographical and architectural variables exhibit a clear ceiling in explaining indoor radon (Swiss model 28%) [PMID 25314691]
    – Building and room characteristics explain approximately half of indoor radon levels (416 classrooms) [PMID 37084905]
    – NRI can be positive even with non-informative markers and should not be used alone for incremental judgments [PMID 26504496][PMID 24240655][PMID 24592497]
    – Comparing correlated ROC curves requires the DeLong method [PMID 3203132]
    – Kappa is influenced by bias and prevalence and must be reported alongside these metrics [PMID 8501467]
    – Measurement error causes association attenuation which can be corrected, but correction methods under non-classical errors may introduce new bias [PMID 37336529][PMID 38721135]

    Partially Verified:
    – Inter-rater reliability of Feng Shui ratings has support from systematic reviews, but reliability does not equate to validity [PMID 37809664]
    – Indoor radon exhibits high spatial variability and season adjustment factors are controversial; protocol standardization is a prerequisite [PMID 30645746][PMID 33130895][PMID 26093853]
    – The orientation→ventilation→radon pathway has a physical basis, but lacks direct measurement specifically for the “orientation” variable

    To Be Verified:
    – The cross-incremental structure of ratings from both schools (redundancy / single independence / double independence / double invalidity)
    – Whether the reliability gap between Luantou and Liqi systematically underestimates the former
    – Whether “Luantou as substance, Liqi as function” manifests as a significant School × Terrain interaction term

    Related Reading

    Evidence Level: B
    References: PMID 25856050, PMID 25314691, PMID 37084905, PMID 26504496, PMID 24240655, PMID 24592497, PMID 3203132, PMID 8501467, PMID 37336529, PMID 38721135, PMID 30645746, PMID 33130895, PMID 26093853, PMID 26810254, PMID 37809664, PMID 26706761, PMID 39292674, PMID 27343029


    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.

  • 「富铀的石头更该躲」:可氡在土里一步只挪一两米,那条公里级的距离量的是别的东西

    Evidence level: B

    把花岗岩体按铀含量和风化程度分成两堆,各堆里量”穴到边界的距离”,再看两堆的距离分布是否不同——这个设计看起来只是把检验做得更细。但它默认了一件事:氡会走,而且是按公里走。氡不走那么远。

    正文

    第一层:距离在最该量它的尺度上已经归零。

    氡-222 在土壤里有两条迁移路径:扩散和载体气流。扩散有明确的长度尺度,孔隙材料里用”浓度衰减到 1/e 的厚度”来定义,实验室对混凝土、石膏这类材料测出来是厘米到分米量级(PMID 15083146)。土壤孔隙率更高时能长一些,但仍是米级。

    这意味着:站在距岩体边界一公里处,岩体产出的氡在扩散项上早已衰减到零。 能在公里尺度搬运氡的只剩载体气流——压力差、温度梯度、风、裂隙通道。而气流走的是裂隙和构造,不是”到边界的直线距离”。

    所以当研究把 1/5/10/20 公里设成扰动档位、把距离当连续变量放进模型时,它量到的不是”氡浓度随距离衰减”,而是”这块地是不是那块石头的风化残积物或坡积物”。距离在这里是替身,替的是岩性,不是浓度。 替身的噪声有多大?用不精确的位置代理替代真实位置,暴露错分的量级足以改变结论方向(PMID 22469491)。

    第二层:铀含量不等于析出强度,中间隔着两个系数。

    就算承认”距离只是岩性的替身”,下一步还要假设”富铀岩体析出更多氡”。这一步也不成立——铀和氡之间隔着析出系数(能逸出矿物晶格的份额)与产生率。

    葡萄牙 Panasqueira 矿区的一组实测把这事说得很直白:随物理化学蚀变推进,析出系数从约 0.1 升到 0.4,产生率从 40 升到 160 Bq·m⁻³·h⁻¹(PMID 29223909)。也就是说,一块铀含量中等但蚀变严重的岩石,析出可以是一块铀更高但新鲜致密岩石的四倍。研究还指出物理蚀变与化学蚀变是解耦的,必须分别量化。

    粒级和含水率同样是主导项:同一批土样,<0.1 mm 粒级的饱和析出系数是 0.47,>0.5 mm 只有 0.23;饱和含水率分别是 16% 和 4%(PMID 32217199)。同一种花岗岩,风化成砂和风化成黏土,差两倍。

    于是”按铀含量分层”这个设计,很可能把真正高析出的样本分进了”低铀组”。

    第三层:地表的铀可能比母岩还低。

    这一层最反直觉。西班牙西北部花岗岩上七个风化-成土剖面的实测显示:铀和钍从表层土淋失,在深部 C 层富集,黏粒级中铀可达母岩的 4 倍、钍 5 倍;砾石含量超过 20% 的层位淋失更强烈(PMID 15923024)。

    也就是说,你站在富铀花岗岩体上,脚下 30 厘米的土可能比下面的石头贫铀得多。 而室内氡的主通道是深部经裂隙向上的气流,不是地表土。用”岩体铀含量”当暴露代理,在剖面上方向可能是反的:地表贫、深层富。

    这也解释了一个常被当成噪声的现象:欧洲几份地生氡潜力制图里,最高值并不出现在铀含量最高处,而出现在坡麓碎屑堆积带(PMID 24412775)——因为那里是”高析出+高渗透+深部补给”三者同时成立的位置,而这三项没有一项由”到边界多远”决定。

    第四层:分层之后,随机基线必须跟着重画。

    这是纯技术点,漏了整套检验就作废。原设计是”风水穴距离分布 vs 随机基线”。一旦按富铀/贫铀分层,每一层里的随机点必须在该层范围内生成,而不是仍在全县撒点。否则富铀区面积小,随机点落进去的期望数量就少,两个分布连支撑集都不一样。这是点格局分析的标准要求,也是边缘效应校正要处理的问题(PMID 11318150)。

    配套还有两个代价。一是切点从哪来:按数据中位数切,等于用同一批数据既选切点又做检验,效应量的区间会系统性偏窄(PMID 15160403);按地质图阈值切,那个阈值是给岩性分类定的,不是给析出强度定的。二是二分本身丢信息:把连续变量二分,即使计入组内相关被削弱带来的补偿,净效应仍是功效下降(PMID 22733454)。

    第五层:避让和得水,可能是同一个方向。

    最后一层不是技术,是解释。如果检验真的显示”风水穴离富铀蚀变岩体更远”,能不能说古人知道氡?

    不能。风水择地第一原则是”得水为上”,而”得水”与”低氡”在多数地貌下指向同一个地方:河谷冲积层——平坦、有水、可耕、低氡。反过来,花岗岩风化壳常常”有泉水但高氡”。所以”远离花岗岩”这个结果,可以由”靠近水”单独解释,不需要任何氡知识。

    宾夕法尼亚 86 万个首层与地下室实测给出了参照:基岩地层是室内氡的强预测因子(某地层中位数 365 Bq/m³,另一地层 93 Bq/m³),但同一模型里,井水、社区类型、天气各有一份独立贡献(PMID 25856050)。地层能解释的那部分,距离替代不了。

    证据与局限

    已验证: 铀镭在风化剖面中分离、表层淋失深层富集(PMID 15923024);蚀变使析出系数与产生率数倍上升(PMID 29223909);粒级与含水率主导析出与扩散(PMID 32217199);多孔材料扩散长度在分米级(PMID 15083146);基岩地层与室内氡强关联(PMID 25856050)。

    部分验证: “距离作为岩性替身”是本篇的推断,未在中国风水点数据上实测;暴露错分量级(PMID 22469491)来自位置代理研究,迁移到地质距离需再验证。

    待验证: 在控制”到河流距离/坡度/到耕地距离”后,风水穴对富铀蚀变岩体的避让是否仍然存在;以及”是否落在岩体残积物上”这一二分变量是否吞掉距离变量。

    相关阅读

    证据等级: B(分层检验设计与各单项证据均有 PubMed 支撑,但”风水穴避让富铀蚀变岩体”本身无实测先例,属研究设计提案)
    参考文献: PMID 15083146, PMID 22469491, PMID 29223909, PMID 32217199, PMID 15923024, PMID 24412775, PMID 11318150, PMID 15160403, PMID 22733454, PMID 25856050


    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。

    \”Stone with enriched uranium should be avoided more\”: Radon moves only one or two meters in soil per step, but the kilometer-scale distance measures something else

    Evidence level: B

    Dividing granite bodies into two groups based on uranium content and weathering intensity, then measuring the “distance from the site to the boundary” within each group to compare distance distributions—this design appears merely to refine the test. However, it presupposes a fact: radon travels, and it travels kilometers. Radon does not travel that far.

    Main Text

    Layer 1: Distance is already zero at the scale where it should be measured.

    Radon-222 in soil has two migration pathways: diffusion and carrier airflow. Diffusion has a well-defined length scale, defined as the thickness where concentration decays to 1/e; laboratory measurements on materials like concrete and gypsum yield centimeters to decimeters (PMID 15083146). While higher soil porosity can extend this range, it remains meter-scale.

    This means: At a distance of one kilometer from the granite body boundary, radon generated by the rock has already decayed to zero via diffusion. Only carrier airflow—driven by pressure differences, temperature gradients, wind, and fracture channels—can transport radon over kilometers. Airflow moves through fractures and structures, not the “straight-line distance to the boundary.”

    Therefore, when a study sets 1/5/10/20 km as disturbance tiers and treats distance as a continuous variable in models, it is not measuring “radon concentration decay with distance,” but rather “whether this land is weathered residual or slope deposit of that specific rock.” Here, distance is a proxy; it stands in for lithology, not concentration. How large is the noise of this proxy? Using imprecise location surrogates for true locations can shift the magnitude of misclassification enough to reverse conclusions (PMID 22469491).

    Layer 2: Uranium content does not equal emanation intensity; two coefficients intervene.

    Even if one accepts “distance is a proxy for lithology,” the next step assumes “uranium-rich bodies emit more radon.” This also fails—between uranium and radon lie the emanation coefficient (the fraction escaping the mineral lattice) and production rate.

    A set of field measurements from the Panasqueira mine in Portugal makes this explicit: as physicochemical alteration progresses, the emanation coefficient rises from ~0.1 to 0.4, and production rates jump from 40 to 160 Bq·m⁻³·h⁻¹ (PMID 29223909). In other words, a rock with moderate uranium but severe alteration can emit four times as much radon as a rock with higher uranium but fresh, dense lithology. The study also notes that physicochemical and chemical alteration are decoupled and must be quantified separately.

    Grain size and water content are also dominant factors: in the same soil sample set, saturated emanation coefficients are 0.47 for <0.1 mm grain size versus 0.23 for >0.5 mm; saturated water contents are 16% and 4%, respectively (PMID 32217199). The same granite can emit twice as much radon when weathered into sand compared to when weathered into clay.

    Thus, the design of stratifying by uranium content likely misclassifies samples with high actual emanation into the “low uranium” group.

    Layer 3: Surface uranium may be lower than in the parent rock.

    This layer is counterintuitive. Field measurements on seven weathering-profile sections in northwest Spain show that uranium and thorium leach from surface soils, enriching in the deep C layer; uranium in clay fractions can reach four times the parent rock level, and thorium five times; leaching is more intense in layers with >20% gravel content (PMID 15923024).

    In other words, standing on a uranium-rich granite body, the soil beneath your feet (30 cm deep) may be significantly poorer in uranium than the rock below. Since the primary pathway for indoor radon is deep airflow through fractures upward, not surface soil, using “parent rock uranium content” as an exposure proxy may yield inverse results across profiles: surface-poor, deep-rich.

    This also explains a phenomenon often treated as noise: in several European maps of terrestrial radon potential, the highest values do not appear at locations with the highest uranium content but rather in scree accumulation zones at slope toes (PMID 24412775)—because these are locations where “high emanation + high permeability + deep recharge” all hold simultaneously, none of which are determined by “how far from the boundary.”

    Layer 4: After stratification, the random baseline must be redrawn.

    This is a purely technical point; omitting it invalidates the entire test. The original design was “Feng Shui site distance distribution vs. random baseline.” Once stratified by high/low uranium, random points must be generated within the range of that specific stratum, not scattered across the entire county. Otherwise, if the high-uranium area is small, the expected number of random points falling into it decreases, causing the support sets of the two distributions to differ entirely. This is a standard requirement for point pattern analysis and an issue addressed by edge-effect correction (PMID 11318150).

    Two associated costs accompany this. First, the cut-point source: using the median of the data selects the cut-point while simultaneously performing the test, causing the effect size interval to be systematically narrowed (PMID 15160403); using a geological map threshold sets the cut-point based on lithology classification, not emanation intensity. Second, dichotomization loses information: even accounting for compensation from reduced within-group correlation, the net effect remains a loss of statistical power (PMID 22733454).

    Layer 5: Avoidance and water-seeking may point in the same direction.

    The final layer is not technical but interpretive. If a test truly shows “Feng Shui sites are farther from uranium-rich altered granite bodies,” can we claim ancients knew about radon?

    No. The first principle of Feng Shui site selection is “water-seeking is paramount,” and “water-seeking” and “low radon” point to the same location under most geomorphologies: alluvial layers in river valleys—flat, water-bearing, arable, and low radon. Conversely, granite weathering crusts often have “spring water but high radon.” Therefore, the result of being “far from granite” can be explained solely by being “near water,” requiring no radon knowledge.

    Measurements of 860,000 first-floor and basement units in Pennsylvania provide a reference: bedrock geology is a strong predictor of indoor radon (median 365 Bq/m³ in one formation, 93 Bq/m³ in another), but within the same model, well water, community type, and weather each contribute independently (PMID 25856050). The portion explained by geology cannot be substituted by distance.

    Evidence and Limitations

    Verified: Uranium-radium separation in weathering profiles with surface leaching and deep enrichment (PMID 15923024); alteration causes multi-fold increases in emanation coefficient and production rate (PMID 29223909); grain size and water content dominate emanation and diffusion (PMID 32217199); diffusion length in porous materials is decimeter-scale (PMID 15083146); strong association between bedrock geology and indoor radon (PMID 25856050).

    Partially Verified: “Distance as a proxy for lithology” is an inference in this paper, not empirically tested on Chinese Feng Shui site data; the magnitude of exposure misclassification (PMID 22469491) comes from location proxy studies and requires re-validation when applied to geological distances.

    To Be Verified: Whether avoidance of uranium-rich altered granite bodies by Feng Shui sites persists after controlling for “distance to river / slope / distance to farmland”; and whether the binary variable of “falling on residual rock” swallows the distance variable.

    Related Reading

    Evidence Level: B (Stratified test design and individual evidence items are supported by PubMed citations, but “Feng Shui site avoidance of uranium-rich altered granite bodies” itself has no prior empirical precedent, constituting a research design proposal)
    References: PMID 15083146, PMID 22469491, PMID 29223909, PMID 32217199, PMID 15923024, PMID 24412775, PMID 11318150, PMID 15160403, PMID 22733454, PMID 25856050


    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.

  • 把断层分成几类,「避让」就变成了几句不同的话

    Evidence level: B

    分层不是在原有结论上磨细刻度,它是把一个问题换成几个问题——而这几个问题能不能合回一句,取决于一个从没人算过的比值。

    正文

    一、分层前后,问的不是同一句话

    不分层时,问题只有一个:风水穴到最近断层的距离,是否系统性地大于随机点。分层之后,问题变成几个:在活动断层附近,穴位是否更远;在死断层附近,是否也更远;在压性断层附近,是否同样。

    这不是同一句话的细化,是几句话。只有各层方向一致、按样本量加权,才会合回原来那一句。任何一层反向,加权平均就可能被拉到零甚至翻转——这就是辛普森悖论。PMID 26854260 在一项多中心 HIV 队列里实测过这种翻转:忽略点位聚类,聚合数据里的趋势会在分组后反向。PMID 22246415 是同一件事的另一面:同一批肺癌数据,聚合分析与分区分析给出互相矛盾的结论。

    二、”类型”是制图者的选择,不是氡的属性

    活动断层与死断层,本质是一条时间窗的划分——”晚更新世以来有没有活动”,窗口是人定的。张性、压性、走滑是应力状态的划分,而同一条断层沿走向可以转换性质,分段后每段的力学属性都可能不同。倾角、破碎带宽度、断层泥厚度,这些都没进标签。

    标签是离散的,机制是连续的,中间还隔着一层制图者的取舍。真正的机制变量是渗透率,它决定氡能不能上来。而渗透率恰恰不是”类型”的函数。PMID 37020109 在北亚平宁两条结构复杂的断层带上做了系统的原位露头渗透率测量,结论是”强烈的空间非均质”:同一条带内部,不同脆性结构相差异极大。PMID 23373845 对一条成熟走滑断层带做水压试验加数值模拟,测出的是各向异性渗透率叠加内部分区。PMID 31290141 更直接:断层既可以当通道,也可以当屏障。

    所以”活动断层是氡的高速通道”在类型层面成立,在点位层面经常不成立。

    三、组内差是三个数量级,这是本篇最要紧的数字

    北安纳托利亚断裂带西段 12 个测点、连续观测 12 个月,土壤氡 29 到 7059 Bq/m³,均值 1930(PMID 26903173)——极差两百多倍。印度 Kachchh 2001 年地震震区 277 个点覆盖约 8250 km²,11.3 到 15,319 Bq/m³,均值 1331(PMID 42631768)——三个数量级。印尼 Lembang 断裂带平均 14,117 Bq/m³,异常点超 20,000,且西段系统性高于东段(PMID 36442382)。希腊北部五个活动断层,断层轴附近信号/背景比从 2 到 13(PMID 12941512)——同类断层,增益差六倍。

    换言之,同一个”活动断层”标签下的点位之间,本身就能差三个数量级。

    分层的收益是”组间差除以组内差”。组内差跨三个数量级时,方差缩减微乎其微,切掉的样本量却是实打实的。这不是统计学推论,是算术。

    四、切碎样本的代价,在置换检验上格外明显

    样本分三四层,每层可能只剩二三十个穴位点,最小可检距离差随之抬升——原来能分辨的 150 米与 400 米,分层后未必分得出来,”不显著”于是成了”这一层没有避让”。这不是结论,是分辨率不够。

    再叠上”逐层各做一次检验”:三层各出一个 p 值,名义 5% 的实际族错误率接近 15%,还多给了一个自由度——哪层显著就报哪层。PMID 17599075 把这套做法直接叫亚组分析的谬误,药方三条:必须用正规的交互检验、必须预注册、必须有足够功效。PMID 32820592 补技术细节:分层稀疏时,分层置信区间的覆盖率会明显失真,多数方法在这个区间不能信。

    五、分层还会塞进新的混杂

    活动断层多在山地,死断层多在平原与盆地边缘;风水穴选址又同时受水源、坡度、耕地驱动。于是”穴位远离活动断层”很可能只是”穴位不在陡山上”。分层没解决这个问题,只把地形效应关进层内,让人以为控制过了。这与 #468 的坐标精度是两件事:那里错的是位置,这里错的是归组。

    公道的另一半:不分层也不对。若效应修饰真存在——比如压性、走滑挤压侧氡反而偏低——不分层会把方向相反的两层平均成零,得出”无关联”的错误结论。争点不是”分不分”,而是”分完怎么报”。

    六、判决三件事

    其一,分层变量尽量按连续或多分类处理,层间差异用一个交互项检验,不逐层扫 p。其二,先算组内/组间方差比(ICC);ICC 低就承认分层不划算,改用把断层距离当连续变量、断层段当随机效应的地统计模型。其三,每层都报”最小可检出距离差”,层太小就合并,或明确写”该层不可判”。三件事做完,”避让”才有确定的意思。

    易学落点:风水自己就是分层的,但它分层的依据不是类型,是通塞。《雪心赋》讲龙有真假、有过峡束气入首,同一条山脉分段看,不同段落吉凶不同——这已经是”断层分段”的古代版本。区别在于:古人按”气通不通”分,现代人按”活动/死””张/压”分。前者直接对着输运效率,后者只是一个粗糙代理,而且常常失效。用标签的标签去追机制,是分层最容易犯的错。更值得记的一句是”石山不可葬”——岩石裸露、裂隙过通,气反而留不住;这与氡的乘积形式(源浓度×渗透率)在结构上同构:既要有源,也要留得住。真正的分层变量,应该是这个”留”字。

    证据与局限

    已验证: 断层带内部渗透率呈强烈空间非均质、且可同时充当流体通道与屏障(PMID 37020109、23373845、31290141);同类断层上不同测点的土壤氡可跨两到三个数量级(PMID 26903173、42631768、36442382、12941512);聚合与分区分析可给出方向相反的结论(PMID 26854260、22246415);亚组分析须用交互检验、须预注册、须有功效(PMID 17599075),分层稀疏时置信区间覆盖率失真(PMID 32820592)。

    部分验证: “活动断层供气更强”在类型-区域层面有实测支持(PMID 30425320、32339146、42631768),但从未有人把”断层类型”作为分层变量,系统检验过风水穴距离分布的组间异质性;”断层的渗透性随断层泥与矿物沉淀自我封闭、故活动性与通透性并非单调对应”是基于断层带结构研究的合理推论(PMID 37020109、23373845),尚无直接的时序封闭速率实测。

    待验证: 分层前后效应量比值的实际分布;”穴位远离活动断层”在控制地形与水源后是否仍存在;断层段(而非断层类型)作为分析单元时结论是否改变;中国南方湿润区与西北干旱区是否因分类判据保存条件不同而呈现完全不同的分层结果。

    相关阅读

    证据等级: B(断层带渗透率非均质、土壤氡的组内变异幅度均有 PubMed 实测支撑;”按断层类型分层后风水穴避让结论是否稳健”本身无先例,属研究设计裁决)
    参考文献: PMID 12941512, PMID 17599075, PMID 22246415, PMID 23373845, PMID 26854260, PMID 26903173, PMID 30425320, PMID 31290141, PMID 32339146, PMID 32820592, PMID 36442382, PMID 37020109, PMID 42631768


    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。

    Categorizing Faults Turns \”Avoidance\” into Several Different Statements

    Evidence level: B

    Stratification is not merely refining the scale on an existing conclusion; it transforms a single question into several. Whether these can be reconciled into one statement depends on a ratio that has never been calculated before.

    Main Text

    I. The Question Changes Before and After Stratification

    Without stratification, there is only one question: Is the distance from a Feng Shui site to the nearest fault systematically greater than that of random points? After stratification, the question becomes multiple: Is the site farther near active faults? Is it also farther near dead faults? Is it similarly distant near compressive faults?

    This is not a refinement of the same sentence; it is several sentences. Only if all layers align directionally and are weighted by sample size will they reconcile back into the original statement. If any layer reverses direction, the weighted average may be pulled to zero or even flip—this is Simpson’s Paradox. PMID 26854260 demonstrated this flip in a multi-center HIV cohort: ignoring point clustering caused trends in aggregated data to reverse after grouping. PMID 22246415 shows the other side of the same coin: the same lung cancer dataset yielded contradictory conclusions from aggregated versus stratified analyses.

    II. “Type” Is a Cartographer’s Choice, Not a Property of Radon

    Active versus dead faults are essentially a temporal window division—”has there been activity since the late Pleistocene?”—a window defined by humans. Normal, reverse, and strike-slip classify stress states, yet a single fault can change its nature along its trace; mechanical properties of each segment may differ. Dip angle, fault zone width, and clay thickness are not included in the labels.

    Labels are discrete; mechanisms are continuous, separated by a layer of cartographer discretion. The true mechanism variable is permeability, which determines whether radon can rise. Permeability is not a function of “type.” PMID 37020109 conducted systematic in-situ outcrop permeability measurements on two structurally complex fault zones in the North Apennines, concluding in “strong spatial heterogeneity”: different brittle structural phases within the same zone differed vastly. PMID 23373845 performed hydraulic testing combined with numerical simulation on a mature strike-slip fault zone, measuring anisotropic permeability superimposed with internal zoning. PMID 31290141 is even more direct: faults can act as conduits or as barriers.

    Thus, “active faults are high-speed conduits for radon” holds at the typological level but frequently fails at the point level.

    III. Intra-group Variation Spans Three Orders of Magnitude; This Is the Most Critical Number in This Piece

    In the western North Anatolian Fault Zone, 12 measurement points observed continuously for 12 months showed soil radon ranging from 29 to 7059 Bq/m³, with a mean of 1930 (PMID 26903173)—a range of over two hundred times. In the Kachchh earthquake zone in India covering ~8250 km² with 277 points, values ranged from 11.3 to 15,319 Bq/m³, with a mean of 1331 (PMID 42631768)—three orders of magnitude. In the Lembang fault zone, Indonesia, the average was 14,117 Bq/m³ with anomalies exceeding 20,000, and the western segment was systematically higher than the eastern (PMID 36442382). In five active faults in northern Greece, the signal-to-background ratio near the fault axis ranged from 2 to 13 (PMID 12941512)—a six-fold difference among similar faults.

    In other words, points under the same “active fault” label can differ by three orders of magnitude.

    The benefit of stratification is the ratio of “between-group difference divided by within-group difference.” When intra-group variation spans three orders of magnitude, variance reduction is negligible, yet the discarded sample size is real. This is not a statistical inference; it is arithmetic.

    IV. The Cost of Fragmenting Samples Is Especially Visible in Permutation Tests

    Dividing samples into three or four layers may leave only twenty to thirty sites per layer, raising the minimum detectable distance difference. What was once resolvable at 150 meters versus 400 meters may no longer be distinguishable after stratification, rendering “not significant” effectively meaning “no avoidance in this layer.” This is not a conclusion; it is insufficient resolution.

    Compounding this with “performing one test per layer”: three layers yield three p-values, pushing the nominal 5% family-wise error rate close to 15%, while also consuming an extra degree of freedom—report significance only where a layer is significant. PMID 17599075 explicitly labels this practice as the fallacy of subgroup analysis, prescribing three remedies: must use formal interaction tests, must preregister, and must have sufficient power. PMID 32820592 adds technical details: when stratification is sparse, coverage of stratified confidence intervals becomes significantly distorted; most methods cannot be trusted in this interval.

    V. Stratification Also Introduces New Confounders

    Active faults are often in mountainous areas, while dead faults are typically at the edges of plains and basins. Feng Shui site selection is simultaneously driven by water sources, slope, and arable land. Thus, “sites far from active faults” may simply mean “sites not on steep mountains.” Stratification does not solve this; it merely confines the topographic effect within layers, leading one to believe control has been achieved. This is distinct from #468 regarding coordinate precision: there, the error lies in location; here, it lies in grouping.

    The fair counterpoint: not stratifying is also incorrect. If effect modification truly exists—for example, if compressive or reverse faults on the squeeze side show lower radon—stratification averages opposing directions into zero, yielding the erroneous conclusion of “no association.” The debate is not “whether to stratify,” but “how to report after stratifying.”

    VI. Three Things Must Be Decided

    First, treat stratification variables as continuous or multi-class whenever possible; test inter-layer differences with one interaction term, not by sweeping p-values layer by layer. Second, calculate the ratio of within-group to between-group variance (ICC); if ICC is low, admit stratification is not cost-effective and instead use geostatistical models treating fault distance as continuous and fault segments as random effects. Third, report the “minimum detectable distance difference” for each layer; merge layers if too small, or explicitly state “unjudgable for this layer.” Once these three things are done, “avoidance” has a definite meaning.

    Yijing Studies Application: Feng Shui itself is inherently stratified, but its basis is not type, but flow and obstruction. Xue Xin Fu discusses true versus false dragons, overpasses, and qi gathering at entrances; viewing the same mountain range in segments yields different auspiciousness for each segment—this is already an ancient version of “fault segmentation.” The difference lies in how ancients stratified by “whether qi flows,” whereas moderns stratify by “active/dead” or “normal/compressive.” The former directly addresses transport efficiency; the latter is a rough proxy that often fails. Chasing mechanisms with labels is the most common error in stratification. A more memorable principle is “do not bury on exposed rock”—where rocks are bare and fissures are overly permeable, qi cannot be retained. This is structurally isomorphic to radon’s multiplicative form (source concentration × permeability): there must be a source, but also retention. The true stratification variable should be this “retention.”

    Evidence and Limitations

    Verified: Fault zones exhibit strong spatial heterogeneity in permeability and can simultaneously act as fluid conduits and barriers (PMID 37020109, 23373845, 31290141); soil radon at different points on similar faults can span two to three orders of magnitude (PMID 26903173, 42631768, 36442382, 12941512); aggregated and stratified analyses can yield opposite conclusions (PMID 26854260, 22246415); subgroup analysis requires interaction tests, preregistration, and sufficient power (PMID 17599075), with confidence interval coverage distortion when stratification is sparse (PMID 32820592).

    Partially Verified: “Active faults provide stronger radon supply” has empirical support at the typological-regional level (PMID 30425320, 32339146, 42631768), but no one has ever used “fault type” as a stratification variable to systematically test between-group heterogeneity in Feng Shui site distance distributions. The inference that “permeability self-seals with fault clay and mineral precipitation, so activity and permeability are not monotonically correlated” is a reasonable deduction based on fault zone structural studies (PMID 37020109, 23373845), but lacks direct empirical measurement of temporal sealing rates.

    To Be Verified: The actual distribution of effect size ratios before and after stratification; whether “sites far from active faults” persists after controlling for topography and water sources; whether conclusions change when fault segments (rather than fault types) serve as analysis units; and whether southern humid regions and northwestern arid regions in China show completely different stratification results due to differing classification criteria preservation conditions.

    Related Reading

    Evidence Level: B (Spatial heterogeneity of fault zone permeability and intra-group variation magnitude in soil radon are supported by PubMed measurements; “whether Feng Shui avoidance conclusions remain robust after stratification by fault type” itself has no precedent, belonging to a research design judgment)
    References: PMID 12941512, PMID 17599075, PMID 22246415, PMID 23373845, PMID 26854260, PMID 26903173, PMID 30425320, PMID 31290141, PMID 32339146, PMID 32820592, PMID 36442382, PMID 37020109, PMID 42631768


    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.

  • 把欧洲的图换成中国的图:文化点 × 地生氡潜力的跨文化复制该怎么判

    Evidence level: B

    复制之前得先确认两件事:你要复制的是哪个方向,以及你手里那把尺子和对方是不是同一把。

    正文

    上一篇讲透了定位误差。这一篇往下走一步:假设坐标准了、本土的地生氡数据也拿到了,把欧洲的结果在中国再做一遍,算不算复制?

    先说结论:不能直接做。因为”复制”这个词里藏着三个没定的前提:方向、变量、尺子。

    一、先定方向:你要复制的是”聚气”还是”避害”

    风水讲”藏风聚气”,按字面读,吉地应当是”气”更足的地方;若把”气”对应到氡,吉地就该落在高 GRP 上。可现代健康语境里氡是肺癌第二诱因,同一批研究便不自觉把”吉地”写成”低氡”。

    这两条方向相反,p 值不可能同时支持。中国全国室内氡综述汇总 37,886 条测点、覆盖 31 省,加权均值 29.4 Bq/m³(PMID 39968227);31 个省会 2,875 户实测算术均值 65、几何均值 55,13.6% 超 100、0.6% 超 300(PMID 39667966)。这些数字不带方向——它们只回答”多不多”,不回答”该多还是该少”。

    所以跨文化复制的第一步不是跑检验,是预注册方向:”聚气”读法预测吉地 GRP 偏高,”避害”读法预测偏低。选一个写死再去复制,两个都报等于哪个方向都能事后解释。

    二、数据不是没有,是错层的

    中国不缺室内氡:2023—2025 年 20 个城市的调查给出农村中位 72、城市 65 Bq/m³(PMID 42146003),样本量够做省级分层。

    但 GRP 不是室内氡。它的定义是土壤气氡浓度 × 土壤气渗透率的联合量(PMID 33254926 明确写出这个乘积形式)。中国的土壤气氡测量是点状的:深圳做过土壤气氡加渗透率加铀含量的三联测(PMID 27594727),陕北做过不同岩性不同深度的析出率,量级 0.3—0.6 Bq/m²/s(PMID 35948128),广东测过土壤气 220Rn(PMID 22923249)。没有全国统一的 GRP 栅格。

    也就是说:室内氡有全国网但被建筑污染,土壤气氡干净但只有点。 跨文化复制缺的正是后者。

    三、GRP 是乘积,换代理变量换的是分子不是分母

    渗透率必须单独测,现在有工具能同步测渗透率与 CO₂(PMID 39751708)。马来西亚的 GRP 图范围 0.04 到 154,最高值全在花岗岩上(PMID 34320010);瑞士用伽马剂量率加地质加断层加表土渗透率做有序 logistic(PMID 27343029);环境伽马剂量率也被当作替代指标(PMID 33172630)。

    问题是,渗透率这一半在中国几乎空白。匈牙利的研究里 GRP 最高值出现在坡麓碎屑堆积,山地丘陵普遍高于平原(PMID 24412775)——说话的其实是渗透率,不是铀含量。黄土的孔隙结构对含水率极敏感,初始含水率改变会重排孔隙(PMID 42860066),氡析出率则随温度线性上升到 400℃ 后才指数下降(PMID 35367261)。同一块黄土,含水率差几个百分点,GRP 就能差一截。

    四、室内氡这一层会盖住地质信号,且盖法与房子共变

    英国西南部的经典研究给出一个刺眼的数字:土壤气氡与室内氡只有极弱相关,仅在高土壤气浓度端相关变强,土壤气通常是室内的 10 到 1000 倍(PMID 9600304)。瑞士 15 万栋房、24 万条测量,最好的模型也只解释 28% 的变异(PMID 25314691)。

    盖住它的东西恰好与中国的地域共变:农村中位高于城市,且低楼层、砖木结构、节能建筑更高(PMID 42146003);高层公寓的通风率显著改变浓度(PMID 35368088);建材贡献还可能被系统性低估(PMID 39920154)。中国农村的”砖木加低层加无地下室”三件事,与”南方北方、山区平原”高度绑定。若文化点集中在水网平原,测出低氡可能只是”砖木房通风好”,不是”选址避氡”。

    五、复制判的是斜率能不能搬,不是这次显不显著

    欧洲那张 GRP 图本身就是拼图:各国输入数据的定义、采样协议、代理变量都不同(PMID 23295644),室内氡图覆盖 24 国(PMID 25063783)。在这样的底图上比较,先得证明”同一把尺子量的是同一个东西”——测量不变性要求先验形态、再验载荷、最后验截距(PMID 33119375、25499982),用到氡上就是检测器类型、季节窗、暴露时长、楼层必须一致。

    再是外部验证的对象。地理可迁移性要用留一中心交叉验证:16 家医院、11,599 例,AUC 在医院间从 0.59 到 0.79,7 家校准失效(PMID 36822443);时间与地理可迁移性必须分开检验(PMID 17907249)。

    所以复制判据分三层:① 符号一致;② 效应量落进原研究的置信区间,而非”自己显著”;③ 校准与阈值能否搬。最常见的结局是前两条过、第三条不过——关系可迁移,阈值不可搬。中国室内氡几何均值 55,与欧洲基线不同,同一个”高氡分位”对应的绝对剂量根本不是一回事。

    六、还有一层:机会结构

    中国聚落本就长在黄土塬、河谷冲积平原、江南水网——恰恰是低 GRP 单元,高 GRP 的花岗岩山地很少是人口密集点。若抽样空间里高 GRP 格占比本就小,”避开高氡”在统计上不可识别,因为缺对照格。

    这不是文化差异,是地理机会结构差异。所以报告必须写一件事:可用高 GRP 格的比例。若低于 5%,正确结论是”在本土空间结构下不可判”,而不是”复制失败”。

    收尾:《葬书》说”气乘风则散,界水则止”。这句话在物理上分成两半:一半讲源,一半讲留——正好是 GRP 的乘积形式。含水率升高、渗透率下降,对应”界水则止”(PMID 42860066)。所以”藏风聚气”译成 GRP 语言,不是”氡浓度高”,而是”源足且封闭”。

    这也说明:古人要”聚”,现代要”避”。把这两件事混成一个”吉地低氡”的假设去做跨文化复制,无论显著与否,都不是在检验风水。

    证据与局限

    已验证: 中国室内氡有全国量级的实测基线(37,886 条测点覆盖 31 省,加权均值 29.4 Bq/m³,PMID 39968227;31 省会 2,875 户算术均值 65、几何均值 55,PMID 39667966;20 城农村中位 72 高于城市 65,PMID 42146003);GRP 的定义确为土壤气氡浓度与土壤气渗透率的联合量(PMID 33254926),且已有同步测渗透率与 CO₂ 的工具(PMID 39751708);中国已有区域性土壤气氡—渗透率—铀含量三联测(PMID 27594727)与分岩性析出率数据(PMID 35948128);室内氡与土壤气氡仅弱相关、最佳模型仅解释 28% 变异(PMID 9600304、25314691);地理可迁移性可用留一中心交叉验证量化,且性能跨中心波动很大(PMID 36822443、17907249);测量不变性的分步验证程序成熟(PMID 33119375、25499982)。

    部分验证: 黄土的孔隙结构与含水率、温度对氡析出的控制有实验数据(PMID 42860066、35367261、41855605),但尚未在真实村落尺度上与 GRP 栅格耦合;中国部分城市与区域已做出局部氡风险图(PMID 27594727),但没有全国统一 GRP 栅格;欧洲 GRP 图的跨国异质性已被官方文献承认(PMID 23295644、25063783),这为”跨文化复制需要先做尺度对齐”提供了先例,但从未有人把该框架用在中国文化点上。

    待验证: 中国本土 GRP 栅格的可得性与粒度;”聚气”与”避害”两种读法各自的先验方向与效应量;室内氡对地质信号的掩盖比例在中国南北、城乡、结构类型上的分层估计;以及本土空间结构中高 GRP 格的可及比例——这一项直接决定复制是”成功””失败”还是”不可判”。

    相关阅读

    证据等级: B(中国室内氡基线、GRP 定义、室内外相关强度、外部验证方法均有 PubMed 直接支撑;”跨文化复制判据”属方法移植,本场景无先例)
    参考文献: PMID 39968227, PMID 39667966, PMID 42146003, PMID 33254926, PMID 39751708, PMID 34320010, PMID 27343029, PMID 33172630, PMID 24412775, PMID 27594727, PMID 35948128, PMID 22923249, PMID 42860066, PMID 35367261, PMID 41855605, PMID 9600304, PMID 25314691, PMID 35368088, PMID 39920154, PMID 23295644, PMID 25063783, PMID 33119375, PMID 25499982, PMID 36822443, PMID 17907249


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

    How to Judge Cross-Cultural Replication of Cultural Points and Geological Hydrogen Sulfide Potential: Replacing European Maps with Chinese Maps

    Evidence level: B

    Before replication, two things must be confirmed: which direction you are replicating, and whether your ruler matches theirs.

    Main Text

    The previous article clarified localization errors. This one moves a step further: assuming coordinates are fixed and local geogenic radon data is available, can we simply re-run European results in China to count as replication?

    The conclusion is clear: No direct replication is possible. The term “replication” hides three undefined premises: direction, variables, and the ruler.

    I. First, define the direction: Are you replicating “Gathering Qi” or “Avoiding Harm”?

    Feng Shui emphasizes cang feng ju qi (hiding wind and gathering Qi). Literally, auspicious sites should have more “Qi.” If mapped to radon, auspicious sites should correspond to high GRP values. However, in modern health contexts, radon is the second leading cause of lung cancer; consequently, studies naturally frame “auspicious sites” as “low radon.”

    These two directions are opposite; p-values cannot simultaneously support both. The national review of indoor radon in China aggregates 37,886 measurement points across 31 provinces, with a weighted mean of 29.4 Bq/m³ (PMID 39968227). In 31 provincial capitals, measurements of 2,875 households show an arithmetic mean of 65 and a geometric mean of 55, with 13.6% exceeding 100 and 0.6% exceeding 300 (PMID 39667966). These figures are directionless—they only answer “how much,” not “should it be more or less.”

    Therefore, the first step in cross-cultural replication is not running a test but pre-registering the direction: the “Gathering Qi” reading predicts higher GRP at auspicious sites, while the “Avoiding Harm” reading predicts lower GRP. Selecting one direction rigidly before replication prevents post-hoc justification for both.

    II. Data exists but is misaligned in layers

    China lacks no indoor radon data: surveys from 2023–2025 in 20 cities provide medians of 72 for rural areas and 65 for urban areas (PMID 42146003), with sufficient sample sizes for provincial stratification.

    However, GRP is not indoor radon. Its definition is the joint quantity of soil gas radon concentration × soil gas permeability (PMID 33254926 explicitly defines this product). China’s soil gas radon measurements are point-based: Shenzhen conducted a tri-measurement of soil gas, permeability, and uranium content (PMID 27594727); Northern Shaanxi measured emission rates across different lithologies and depths with magnitudes of 0.3–0.6 Bq/m²/s (PMID 35948128); Guangdong measured soil gas at 220Rn (PMID 22923249). There is no unified national GRP grid.

    In short: Indoor radon has a national network but is polluted by buildings; soil gas radon is clean but only point-based. Cross-cultural replication lacks precisely the latter.

    III. GRP is a product; changing proxies changes the numerator, not the denominator

    Permeability must be measured separately; tools now exist to measure permeability and CO₂ simultaneously (PMID 39751708). Malaysia’s GRP maps range from 0.04 to 154, with maximum values entirely on granite (PMID 34320010); Switzerland uses gamma dose rate plus geology plus faults plus surface soil permeability in an ordered logistic model (PMID 27343029); environmental gamma dose rate is also used as a proxy (PMID 33172630).

    The issue is that permeability data in China is almost blank. In Hungarian studies, the highest GRP values appeared on slope scree deposits, with mountains and hills generally higher than plains (PMID 24412775)—what was actually speaking was permeability, not uranium content. The pore structure of loess is extremely sensitive to water content; initial water content changes rearrange pores (PMID 42860066), and radon emission rates rise linearly with temperature until 400°C before dropping exponentially (PMID 35367261). The same loess can have significantly different GRP values differing by just a few percentage points of water content.

    IV. The indoor radon layer masks geological signals, and the masking mechanism covaries with housing types

    A classic study from southwestern England provides a striking figure: soil gas radon and indoor radon show only a very weak correlation, becoming stronger only at high soil gas concentrations, with soil gas typically being 10 to 1000 times higher indoors (PMID 9600304). Switzerland’s dataset of 150,000 houses and 240,000 measurements shows that even the best model explains only 28% of the variance (PMID 25314691).

    What masks it covaries with China’s regional patterns: rural medians are higher than urban ones, and low floors, brick-wood structures, and energy-efficient buildings show higher levels (PMID 42146003); ventilation rates in high-rise apartments significantly alter concentrations (PMID 35368088); and material contributions may be systematically underestimated (PMID 39920154). China’s rural “brick-wood + low floor + no basement” triad is tightly bound to “South/North and Mountain/Plain.” If cultural points concentrate on water network plains, finding low radon might simply mean “brick-wood houses ventilate well,” not “site selection avoids radon.”

    V. Replication judges whether the slope can be transferred, not whether significance appears this time

    The European GRP map itself is a puzzle: definitions, sampling protocols, and proxy variables differ across countries (PMID 23295644); indoor radon maps cover 24 countries (PMID 25063783). Comparing against such a baseline requires first proving “the same ruler measures the same thing”—measurement invariance requires prior validation of form, then verification of loadings, finally verification of intercepts (PMID 33119375, 25499982). Applied to radon, this means detector types, seasonal windows, exposure duration, and floor levels must be consistent.

    Next is external validation of the object. Geographical transferability requires leave-one-center cross-validation: across 16 hospitals and 11,599 cases, AUC varied between 0.59 and 0.79 across centers, with 7 failing calibration (PMID 36822443); temporal and geographical transferability must be tested separately (PMID 17907249).

    Thus, replication criteria are three-fold: ① Sign consistency; ② Effect sizes falling within the original study’s confidence interval, not just “significance”; ③ Whether calibration and thresholds can be transferred. The most common outcome is passing the first two but failing the third—the relationship transfers, but thresholds do not. China’s indoor radon geometric mean is 55, differing from European baselines; the same “high radon percentile” corresponds to a completely different absolute dose.

    VI. One more layer: Opportunity Structure

    Chinese settlements naturally grow on loess plateaus, river valley alluvial plains, and Jiangnan water networks—precisely low GRP units. High GRP granite mountains rarely coincide with dense population centers. If the sampling space already has a small proportion of high GRP grids, “avoiding high radon” is statistically unidentified due to a lack of control grids.

    This is not a cultural difference but a difference in geographical opportunity structure. Therefore, reports must explicitly state the proportion of usable high GRP grids. If below 5%, the correct conclusion is “unjudgable under local spatial structure,” not “replication failure.”

    Conclusion: Zang Shu states, “Qi scatters when caught by wind and stops when bounded by water.” Physically, this splits into two halves: one speaks of sources, the other of retention—exactly the multiplicative form of GRP. Increased water content and decreased permeability correspond to “bounded by water” (PMID 42860066). Thus, translating “Gathering Qi” into GRP language means not “high radon concentration,” but “sufficient source and closed retention.”

    This also clarifies: ancients sought to “gather”; moderns seek to “avoid.” Merging these into a single hypothesis of “auspicious sites have low radon” for cross-cultural replication, regardless of significance, is not testing Feng Shui.

    Evidence and Limitations

    Verified: China possesses a national-level empirical baseline for indoor radon (37,886 points across 31 provinces, weighted mean 29.4 Bq/m³, PMID 39968227; 2,875 households in 31 capitals with arithmetic mean 65 and geometric mean 55, PMID 39667966; rural median 72 higher than urban 65 in 20 cities, PMID 42146003); GRP is confirmed as the joint quantity of soil gas radon concentration and soil gas permeability (PMID 33254926), with tools existing for simultaneous permeability and CO₂ measurement (PMID 39751708); China already has regional tri-measurements of soil gas radon, permeability, and uranium content (PMID 27594727) and lithology-specific emission data (PMID 35948128); indoor radon and soil gas radon show only weak correlation with best models explaining 28% variance (PMID 9600304, 25314691); geographical transferability can be quantified via leave-one-center cross-validation with large performance fluctuations across centers (PMID 36822443, 17907249); measurement invariance validation procedures are mature (PMID 33119375, 25499982).

    Partially Verified: Experimental data exists on loess pore structure and water content/temperature controlling radon emission (PMID 42860066, 35367261, 41855605), but it has not yet been coupled with GRP grids at the true village scale; partial radon risk maps exist for some Chinese cities and regions (PMID 27594727), but no unified national GRP grid exists; European GRP map cross-national heterogeneity is acknowledged in official literature (PMID 23295644, 25063783), providing precedent for “cross-cultural replication requires scale alignment,” yet no one has applied this framework to Chinese cultural points.

    To Be Verified: Availability and granularity of domestic GRP grids in China; prior directions and effect sizes for both “Gathering Qi” and “Avoiding Harm” readings; stratified estimates of indoor radon masking geological signals across North/South, urban/rural, and structural types in China; and the accessibility proportion of high GRP grids in local spatial structures—this directly determines whether replication is “successful,” “failed,” or “unjudgable.”

    Related Reading

    Evidence Level: B (China’s indoor radon baseline, GRP definition, indoor-outdoor correlation strength, and external validation methods are directly supported by PubMed; “cross-cultural replication criteria” is a methodological transplant with no precedent in this context)
    References: PMID 39968227, PMID 39667966, PMID 42146003, PMID 33254926, PMID 39751708, PMID 34320010, PMID 27343029, PMID 33172630, PMID 24412775, PMID 27594727, PMID 35948128, PMID 22923249, PMID 42860066, PMID 35367261, PMID 41855605, PMID 9600304, PMID 25314691, PMID 35368088, PMID 39920154, PMID 23295644, PMID 25063783, PMID 33119375, PMID 25499982, PMID 36822443, PMID 17907249


    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.

  • 当坐标模糊到和网格一样大:文化点 × 地生氡潜力叠加检验的定位误差敏感性分析

    Evidence level: B

    一个”县”级坐标落在 10 公里网格的哪一格,很可能由测绘本身体决定,而不是由风水决定。

    正文

    上一篇把实验摆上了桌:中国风水穴、穴位地名、名医故里数字化成点,投到欧洲地生氡潜力(GRP)栅格上做蒙特卡洛置换检验。那篇留了个尾巴——方志里的”县”只有公里级精度,而 GRP 图是 10 公里网格,两者同量级。这一篇专治这个尾巴。

    先说结论:如果文化点的定位误差与栅格尺度同量级,”落在哪一格”这个动作本身就成了随机抽取;置换检验做得再漂亮,检验的也可能只是”我们怎么测坐标”,而不是”古人怎么选址”。

    一、定位误差有多大,有硬数字

    不是”可能有点偏”。街道级地理编码实测误差中位数 41 米、99 分位 273 米,而且这些误差不是随机的——250 米以内它系统性高估暴露人数,假阳假阴同时大量出现(PMID 17367533)。农村更差:不做地址规范化,25% 的点偏离真位超过 1000 米(PMID 20406495);法国 E3N 队列 4247 条记录里,城市地址能落到门牌或街道段一级的占 74.4%,农村只有 10.5%(PMID 28235407);美国农业健康研究两轮复测,农村与低密度区误差系统性更大(PMID 33572119)。

    换算一下:方志”县治”级坐标精度通常几公里到十几公里,村一级一到三公里。欧洲 GRP 图是 10×10 公里网格(PMID 23295644、25063783)。误差半径约等于网格边长。 一个点相当大概率落在它”应该”在的那一格之外。

    二、为什么这次不能用”大样本”救

    经典测量误差会衰减效应量,样本够大仍能检出衰减后的真效应。空间定位误差是另一种东西。系统综述把它分成四类,从单纯的点位误差一路排到协变量与结局的空间错位(PMID 27368370)。误差一旦非随机,就会造出关联,而不只是削弱关联。而地理编码误差恰恰是结构性的:越偏远越糙,越靠路越容易被”吸附”到路段中点。这与”名医故里集中于江南水乡”这类聚集结构同向——误差方向和被检验的空间结构耦合,样本放大只会让偏倚更”显著”,不会让它消失。

    三、抖动检验:四步,两只臂

    第一步,给每个点加随机扰动,半径取 1/5/10/20 公里,每档重复千次。第二步,每次重跑完整的置换检验——不只重算 p,而要重跑零分布,因为点位变了,聚簇结构也跟着变。第三步,记录 p 值、效应方向,以及 GRP 值的实际变化率。第四步最容易被跳过——同时开一只”零效应臂”:把点位随机撒在同一研究区内(真实关联人为设零),走完全相同的流程,看它有多大比例在抖动后变显著。这是假翻转率,减掉它剩下的才是真损失。

    判据写死:若 5 公里扰动下翻转率超过 20%,或零效应臂假显著率超过 10%,则该检验在此坐标精度下不可判,须降级为村级子样本探索分析,或改做区域级聚合(并接受 MAUP 的另一套代价);若 5 公里扰动下 GRP 值变化率中位数低于 10% 且翻转率低于 5%,方可报告主结论。

    四、三个容易骗自己的地方

    其一,”抖动后还显著”常常是假稳健。 GRP 是空间自相关场,相邻格高度相似。抖动 5 公里往往还落在同一高值斑块里,GRP 值几乎不变,检验当然不翻。所以要报的是 GRP 值变化率,不是 p 值稳不稳。地统计里早有对应做法:用条件模拟生成一组等概率风险图,逐张送进同一套聚类算子,看分类在多大比例的模拟里成立——结果某些”显著低风险聚集”只有少数县在九成以上模拟中显著,另一处高值聚集则在六成模拟图中出现(PMID 16469095)。这才叫”结论的稳定度”。

    其二,尺度本身会改结论。 空间聚合的尺度与分区(MAUP)能让回归结果整体偏移:同一份渥太华 NO₂ 与呼吸系统疾病数据,换三种空间结构,OLS 估计显著不同,而三种结构最终都测不到效应(PMID 22040001);”对 MAUP 有多敏感”本身也已可做成统计检验(S-maup,PMID 30481193),给出”最多聚合到哪一级还不失真”的答案。县级坐标是一种聚合,10 公里网格是另一种聚合,两层叠在一起,误差不是相加,而是相乘。

    其三,地图的”分辨率”不等于”精度”。 同一地区换数据源做 GRP 图,预测力差别很大:挪威航空伽马能谱把解释方差做到约 70%,纯地质图只有约 40%,且前者能给每个估计配上 95% 置信限(PMID 27105766);爱尔兰的详细制图同时输出预测不确定性(PMID 36497982);另一项研究里仅用地物理数据的模型无法准确预测土壤氡活度,必须整合地质信息才解释约 75% 的变异(PMID 37343860)。这些不确定性图层正是抖动检验最好的输入——扰动幅度应按每格的预测方差抽样,而非全局固定半径。

    还有地形。三维地表投到平面坐标时面积与距离都不守恒,这种投影失真会系统性把点判成”聚集”(PMID 28452047)。名医故里扎堆江南,可能只是江南本来就点密。

    收尾:这个检验最诚实的表述,不是”古人避开了高氡带”,而是”在我们能测到的坐标精度之内,古人的选址与高氡带的关系是可判或不可判”。把”不可判”当结论写出来,比一个漂亮的 p 值更有价值——它明确告诉你下一步该补的是族谱与方志的村级定位,而不是再换个更花的模型。

    证据与局限

    已验证: 地理编码位置误差的量级与城乡差异有实测数据(PMID 17367533、20406495、28235407、33572119);空间测量误差可系统分为四类并被统一建模(PMID 27368370);MAUP 的尺度与分区敏感性可量化、可检验(PMID 22040001、30481193、30917821);GRP 制图的分辨率与预测不确定性有成熟方法(PMID 23295644、25063783、27241368、36497982、37343860、27105766);条件模拟能把空间不确定性传播到聚类分类结论上(PMID 16469095)。

    部分验证: “抖动-翻转率”作为稳健性指标,在生态点过程(环形位移零模型、地形校正零模型,PMID 28452047)与疾病制图(PMID 16469095)中已有成熟工具,空间分辨率对时空聚集检测的影响也做过系统模拟(PMID 23110167);但它从未被用在”跨文化坐标叠加”这类场景里。

    待验证: 中国文化点县/村级坐标的真实误差分布(须以 GPS 实测抽样标定,不能直接套用欧美地理编码数据);误差与本底空间聚集的耦合方向(是系统抹平还是系统造出信号);以及”20% 翻转率”这类阈值本身,仍需通过零效应模拟标定后才能定稿。

    相关阅读

    证据等级: B(定位误差、MAUP、GRP 制图不确定性各有 PubMed 直接支撑;”抖动-翻转率”用于本场景属方法移植,无先例)
    参考文献: PMID 17367533, PMID 20406495, PMID 28235407, PMID 33572119, PMID 27368370, PMID 22040001, PMID 30481193, PMID 30917821, PMID 16469095, PMID 23295644, PMID 25063783, PMID 27241368, PMID 36497982, PMID 37343860, PMID 27105766, PMID 28452047, PMID 23110167


    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。

    Positioning Error Sensitivity Analysis When Coordinates Blur to Grid Scale: A Superposition Test of Cultural Sites and Hydrogen-Rich Groundwater Potential

    Evidence level: B

    A “county-level” coordinate falling into which cell of a 10 km grid is likely determined by the cartographer’s body, not by Feng Shui.

    Main Text

    The previous article placed the experiment on the table: Chinese Feng Shui sites, acupoint place names, and hometowns of famous physicians were digitized into points and subjected to Monte Carlo permutation testing against European Geogenic Radon Potential (GRP) raster grids. That article left a tail—county-level gazetteer coordinates have kilometer-level precision, while GRP maps are 10 km grids; the two are of the same order of magnitude. This article addresses that tail specifically.

    Conclusion first: If the localization error of cultural points is commensurate with grid scale, the act of “falling into which cell” itself becomes random sampling. No matter how beautiful the permutation test is, it may only test “how we measured coordinates” rather than “how ancients selected sites.”

    1. How Large Is the Localization Error? Hard Numbers

    Not “possibly slightly off.” In street-level geocoding, the median实测 error is 41 meters with a 99th percentile of 273 meters, and these errors are not random—within 250 meters they systematically overestimate exposure counts, producing both false positives and false negatives (PMID 17367533). Rural areas are worse: without address normalization, 25% of points deviate from the true location by over 1000 meters (PMID 20406495); in the French E3N cohort of 4,247 records, urban addresses could be mapped to the door number or street segment level in 74.4% of cases, whereas rural addresses only achieved this in 10.5% (PMID 28235407); two rounds of re-measurement in the US Agricultural Health Study showed systematically larger errors in rural and low-density areas (PMID 33572119).

    Converted: Gazetteer “county seat” coordinates typically have precision of several kilometers to over ten kilometers; village-level precision ranges from one to three kilometers. European GRP maps use a 10×10 km grid (PMID 23295644, PMID 25063783). The error radius approximates the grid side length. A point has a high probability of falling outside the cell it “should” occupy.

    2. Why Large Samples Cannot Save This Time

    Classic measurement errors attenuate effect sizes; with sufficiently large samples, the attenuated true effect can still be detected. Spatial localization errors are a different category. Systematic reviews categorize them into four types, ranging from simple point errors to spatial misalignment between covariates and outcomes (PMID 27368370). Once errors are non-random, they create associations rather than merely weakening them. Geocoding errors are precisely structural: the more remote and rough, the worse; the closer to a road, the more likely they are “adsorbed” into the segment midpoint. This aligns with aggregation structures such as “famous physician hometowns concentrated in Jiangnan water towns”—the error direction couples with the spatial structure being tested. Amplifying the sample size will only make the bias more “significant,” not cause it to disappear.

    3. Jitter Testing: Four Steps, Two Arms

    Step 1: Add random perturbation to each point with radii of 1/5/10/20 km, repeating 1,000 times per tier. Step 2: Re-run the complete permutation testing each time—not just recalculating p-values, but re-running the null distribution, because point locations change and cluster structures follow. Step 3: Record p-values, effect directions, and the actual rate of change in GRP values. Step 4 is most easily skipped—open a second “null-effect arm”: randomly scatter points within the same study area (artificially setting true association to zero) and follow the exact same process to see what proportion becomes significant after jittering. This is the false flip rate; subtract it from the remainder to get the true loss.

    Rule fixed: If the flip rate exceeds 20% under 5 km perturbation, or if the null-effect arm false-significance rate exceeds 10%, then this test is unjudgable at this coordinate precision and must be downgraded to village-level subsample exploration analysis, or shifted to regional aggregation (accepting another set of MAUP costs); only if the median GRP value change rate is below 10% and flip rate is below 5% under 5 km perturbation may the main conclusion be reported.

    4. Three Places That Easily Deceive You

    First, “Significance after jitter” is often false robustness. GRP is a spatially autocorrelated field where adjacent cells are highly similar. A 5 km jitter often still falls within the same high-value patch, leaving GRP values nearly unchanged; thus, the test naturally does not flip. Therefore, one must report the GRP value change rate, not whether p-values remain stable. Spatial statistics already have corresponding methods: use conditional simulation to generate a set of equally probable risk maps, feed each into the same clustering operator, and observe in what proportion of simulations a classification holds—results show that certain “significant low-risk clusters” are significant in over 90% of simulations for only a few counties, while another high-value cluster appears in 60% of simulated maps (PMID 16469095). Only then is this called “conclusion stability.”

    Second, scale itself can alter conclusions. Spatial aggregation scales and zoning (MAUP) can shift regression results overall: the same Ottawa NO₂ and respiratory disease dataset, with three different spatial structures, yields significantly different OLS estimates, yet none of the three structures detect an effect (PMID 22040001); “sensitivity to MAUP” itself can be made into a statistical test (S-maup, PMID 30481193), providing an answer of “up to which aggregation level does not distort truth.” County-level coordinates represent one type of aggregation; the 10 km grid represents another. Stacked together, errors do not add but multiply.

    Third, a map’s “resolution” does not equal “precision.” Switching data sources for GRP maps in the same region yields vastly different predictive powers: Norwegian airborne gamma spectrometry achieves explanatory variance of approximately 70%, while pure geological maps only achieve about 40%, and the former can assign 95% confidence limits to each estimate (PMID 27105766); Ireland’s detailed mapping simultaneously outputs prediction uncertainty (PMID 36497982); in another study, models using only geophysical data cannot accurately predict soil radon activity and must integrate geological information to explain approximately 75% of the variance (PMID 37343860). These uncertainty layers are precisely the best inputs for jitter testing—perturbation amplitude should be sampled from the prediction variance per cell, not a global fixed radius.

    There is also terrain. Projecting a three-dimensional surface onto planar coordinates does not conserve area or distance; such projection distortion systematically misclassifies points as “clusters” (PMID 28452047). Famous physician hometowns clustering in Jiangnan may simply reflect that Jiangnan is denser in points to begin with.

    Wrap-up: The most honest statement of this test is not “ancients avoided high-radon zones,” but “within the coordinate precision we can measure, the relationship between ancient site selection and high-radon zones is either judgable or not.” Writing “not judgable” as a conclusion holds more value than a pretty p-value—it explicitly tells you the next step is to supplement gazetteer data with village-level positioning from genealogies, not to switch to a more flashy model.

    Evidence and Limitations

    Verified: Magnitude of geocoding location errors and urban-rural differences have empirical data support (PMID 17367533, PMID 20406495, PMID 28235407, PMID 33572119); spatial measurement errors can be systematically categorized into four types and unified modeling (PMID 27368370); MAUP scale and zoning sensitivity can be quantified and tested (PMID 22040001, PMID 30481193, PMID 30917821); GRP map resolution and prediction uncertainty have mature methodologies (PMID 23295644, PMID 25063783, PMID 27241368, PMID 36497982, PMID 37343860, PMID 27105766); conditional simulation can propagate spatial uncertainty to clustering classification conclusions (PMID 16469095).

    Partially Verified: “Jitter-flip rate” as a robustness indicator has mature tools in ecological point processes (circular displacement null model, terrain correction null model, PMID 28452047) and disease mapping (PMID 16469095); the impact of spatial resolution on spatiotemporal clustering detection has also been systematically simulated (PMID 23110167); however, it has never been applied to “cross-cultural coordinate overlay” scenarios.

    To Be Verified: True error distribution of Chinese cultural point county/village coordinates (must be calibrated via GPS实测 sampling, cannot directly apply US/European geocoding data); coupling direction between errors and background spatial aggregation (systematically flattening or systematically creating signals); and thresholds like “20% flip rate” themselves still require finalization through null-effect simulation calibration.

    Related Reading

    Evidence Level: B (Localization errors, MAUP, and GRP mapping uncertainty each have direct PubMed support; “jitter-flip rate” application in this scenario is method transplantation with no precedents)
    References: PMID 17367533, PMID 20406495, PMID 28235407, PMID 33572119, PMID 27368370, PMID 22040001, PMID 30481193, PMID 30917821, PMID 16469095, PMID 23295644, PMID 25063783, PMID 27241368, PMID 36497982, PMID 37343860, PMID 27105766, PMID 28452047, PMID 23110167


    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.

  • 「越严重反应越大」:一条从没人验证过的直线,和四个不能互相换算的严重度

    Evidence level: B

    「严重度越高、放大越强」听上去像常识,其实是一个从未被检验过的形状假设——而且它默认的那种直线,恰恰是生理系统最不常给出的一种。

    正文

    前两篇把「病理放大」拆到了两处:先问它是不是真的(尺度与假象),再问它长在哪儿(局部还是全身)。剩下最后一问:它随严重度怎么变。这个问法看似温和,其实换了一整套逻辑——前面的问题是「有没有」,这里是「形状」。

    形状问题有个特殊陷阱:它默认了「严重度」是一根轴。但临床上至少有四种东西被叫作严重度,它们互不换算。疼痛 VAS/NRS 是主观报告,是序数量表而非等距尺,VAS 与 Likert 的响应关系本身是非线性的(PMID 12814613),而单维疼痛工具的重测信度与反应度证据质量普遍偏低,该领域 31 项研究里多数未达 COSMIN 方法学标准(PMID 34999588);病程是时间累积,方向可能相反(久病既可能适应,也可能中枢敏化);CRP/IL-6 是全身炎症活性,与局部疼痛并不同步——非特异性腰痛的系统综述结论很干脆:hs-CRP 与腰痛不相关,且证据不足以把生物标志物与疼痛严重度或活动状态联系起来(PMID 37195364);类风湿关节炎中炎症指标只解释 22%–27% 的疼痛方差(PMID 39380043);慢性紧张型头痛患者用非甾体药后疼痛显著下降,TNF-α、IL-1、IL-6 却一动不动(PMID 28293317)。中枢敏化指标(QST) 又是另一套尺度。四根轴若给出不同斜率,「剂量-反应」四个字就没有单一所指,讨论必然各说各话。这四根轴自身的精度也不齐:CRP 的个体内变异显著大于胆固醇(ICC 0.62),基线最高三分位者有一半在随访时落入不一致类别,基线超过 3 mg/L 的人里 69% 后来降到更低风险档(PMID 22846611)——一根自己就会跳动的轴,撑不起精密的斜率估计。

    第二个陷阱更硬:「放大倍数」是个比值,而比值在分母小的时候会自己长出形状。 若把响应写成「患者组变化 ÷ 健康组变化」,健康组响应接近零时,分母既小又抖,比值于是爆炸且方差极大——低严重度那一端恰恰分母最不稳,曲线末端自然翘起。这不是生理,是除法。有一份干净的实证先例:一项 304 例的骨质疏松研究里,用百分比变化做结局时,「低 BMI 患者反应更强」显著成立;换成绝对变化后,这个关联彻底消失(r = −0.064,p = 0.267),作者直接判定百分比指标是数学伪影(PMID 42203994)。同一批数据、两种度量、两个结论。

    第三个陷阱:直线是最不该默认的形状。 生理系统的剂量-反应通常先饱和再平台;阿托品近视防控的剂量-反应 meta 分析给出的正是阈值效应而非线性(PMID 42070753),他克莫司浓度与急性排斥也是明确非线性(PMID 31333109)。中医自己也不认为它单调,《素问·六微旨大论》讲「亢则害,承乃制」——过头会反噬,这是倒 U,不是直线。

    第四层要说句公道话:基线-变化的相关里确实有一份来自数学耦合(基线与变化共享误差项),但这条批评有边界——耦合在个体分数层面成立,组水平统计量并不受其影响(PMID 21803544)。所以在本主题里,耦合只是个体层面的威胁;组水平真正的威胁是分母不稳与回归到均值,而非耦合本身。

    第五层,严重度高的人不是「同一种人的加重版」。严重度与用药、共病、年龄高度共线,这就是适应证混杂——倾向评分也治不了没测到的那部分(PMID 24217206)。可用的探针有三件:负对照结局(PMID 40838594;PMID 38619297)、E-value 量化「需要多强的未测混杂才能解释掉」(PMID 28693043),以及警惕反向的过度校正——校正中介会偏离零而不是趋零(PMID 42821507)。把严重度二分成「重/轻」还会额外损失功效并留下残余混杂(PMID 16217841)。

    把这些落成一个能执行的方案,有四条硬约束。同一人做接地/非接地交叉,用配对内差值当响应量,让每个人的基线和用药史自己给自己当对照——这是唯一能同时削掉组间基线和回归到均值的做法。严重度按连续变量入模型,除主效应外只放「严重度 × 接地」一个交互项,不预设分组切点。四根轴各自跑一遍,把四条斜率与四条 95%CI 并排放在一张表里,同向才叫剂量-反应,不同向就老实写「不同严重度定义下不可比」。曲线用限制性样条,结点数预先定死(3–4 个),报拟合优度与形状检验,不用「斜率显著」代替「形状是什么」。加一个健康基线组不为比较放大倍数,只为定位「零响应点」在哪儿——有了零点,倍数才有分母。

    所以判决只有三句。一是剂量轴先选一根、预注册,别在同一篇里四条轴混着讲。二是绝对变化作主结局,倍数只作敏感性分析,且必须同时报分母的变异。三是形状先验三选一(单调/阈值/平台),用限制性样条拟合后比拟合优度,而不是只报斜率符号。

    易学落点在《素问·通评虚实论》:「邪气盛则实,精气夺则虚。」同一个人症状加重,可能是邪增,也可能是正减——两条路在「严重度」这一个数字上长得一模一样,但对外部干预的响应曲线方向相反。所以传统模型从来不把剂量-反应设成直线:它设的是「实者泻之、虚者补之」的分岔。单调是现代人替它加上去的。

    证据与局限

    已验证: 炎症标志物与疼痛严重度的关联在不同病种中强度差异极大,且 hs-CRP 与腰痛不相关(PMID 37195364);炎症仅解释 RA 疼痛方差的 22%–27%(PMID 39380043);镇痛药可显著降低疼痛而不改变 TNF-α/IL-1/IL-6(PMID 28293317);CRP 个体内变异显著大于胆固醇(ICC 0.62,最高三分位 51% 随访时落入不一致类别,PMID 22846611);百分比变化可系统制造出「低基线反应更强」的伪信号(PMID 42203994);阿托品剂量-反应呈阈值型(PMID 42070753)。

    部分验证: 数学耦合在个体层面成立而组水平不受影响(PMID 21803544),这限定了耦合批评的适用范围;单维疼痛量表的测量学证据质量偏低(PMID 34999588),使「严重度」这条轴本身精度可疑;负对照结局与 E-value 可探测未测混杂(PMID 40838594;PMID 28693043),但均未在接地-阻抗场景中应用过。

    待验证: 同一批数据在 VAS、病程、CRP、QST 四根轴上的剂量-反应斜率是否同向;「放大倍数」在低严重度分位是否系统性偏高(分母效应);限制性样条拟合下曲线形状更接近单调、阈值还是平台。

    相关阅读

    证据等级: B
    参考文献: PMID 37195364, PMID 39380043, PMID 28293317, PMID 22846611, PMID 42203994, PMID 42070753, PMID 34999588, PMID 12814613, PMID 21803544, PMID 31333109, PMID 24217206, PMID 40838594, PMID 38619297, PMID 28693043, PMID 42821507, PMID 16217841


    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。

    \”Greater Severity Equals Greater Reaction\”: An Unverified Straight Line and Four Incommensurable Measures of Severity

    Evidence level: B

    “Higher severity implies greater amplification” sounds like common sense, but it is actually an untested shape assumption—one that defaults to a straight line, which is precisely the form physiological systems rarely produce.

    Main Text

    The previous two articles dissected “pathological amplification” into two parts: first, whether it exists at all (scale vs. illusion), and second, where it occurs (local vs. systemic). The final question remains: How does it change with severity. This framing, while seemingly gentle, shifts the entire logic—previous questions asked “does it exist,” whereas this asks about “shape.”

    The shape question carries a specific trap: it defaults to treating “severity” as a single axis. However, clinically, there are at least four distinct constructs labeled “severity,” which are not interchangeable. Pain VAS/NRS is a subjective report using an ordinal scale rather than an interval scale; the response relationship between VAS and Likert scales is inherently nonlinear (PMID 12814613), and unidimensional pain tools generally show low reliability and responsiveness evidence quality, with most of the 31 studies in this field failing to meet COSMIN methodological standards (PMID 34999588). Duration represents time accumulation, where direction may be opposite (chronic conditions may lead to adaptation or central sensitization); CRP/IL-6 reflect systemic inflammatory activity, which is not synchronized with local pain—systematic reviews on nonspecific low back pain yield clear conclusions: hs-CRP is unrelated to low back pain, and evidence is insufficient to link biomarkers to pain severity or functional status (PMID 37195364); in rheumatoid arthritis, inflammatory markers explain only 22%–27% of pain variance (PMID 39380043); and in patients with chronic tension-type headache, pain significantly decreases after nonsteroidal anti-inflammatory drugs while TNF-α, IL-1, and IL-6 remain unchanged (PMID 28293317). Central sensitization metrics (QST) represent yet another scale. If four axes yield different slopes, the term “dose-response” lacks a single meaning, and discussions inevitably become disjointed. Furthermore, the precision of these four axes is inconsistent: CRP exhibits significantly higher within-individual variability than cholesterol (ICC 0.62); among those in the highest third at baseline, half fall into inconsistent categories at follow-up, and 69% of individuals with baseline levels exceeding 3 mg/L later drop into a lower-risk category (PMID 22846611)—an axis that fluctuates on its own cannot support precise slope estimation.

    A second, harder trap is that “amplification ratio” is a quotient, and quotients explode in shape when the denominator is small. If response is written as “patient group change ÷ healthy group change,” a near-zero healthy group response creates a small, noisy denominator, causing the ratio to explode with extreme variance—this instability occurs precisely at the low-severity end, naturally curving upward at the tail. This is not physiology; it is division. A clean empirical precedent exists: in a study of 304 osteoporosis patients, using percentage change as the outcome showed that “lower BMI patients responded more strongly” significantly; however, when switching to absolute change, this association vanished entirely (r = −0.064, p = 0.267), with authors directly concluding that percentage metrics are mathematical artifacts (PMID 42203994). Same data, two metrics, two conclusions.

    The third trap is that the straight line is the least appropriate default shape. Physiological dose-response curves typically saturate before plateauing; a meta-analysis of atropine myopia prevention dose-response yielded a threshold effect rather than linearity (PMID 42070753), and tacrolimus concentration versus acute rejection is explicitly nonlinear (PMID 31333109). Traditional Chinese Medicine itself does not view it as monotonic; the Su Wen · Liu Wei Zhi Da Lun states “Excess leads to harm, and counter-regulation controls it”—overshoot leads to backlash, forming an inverted U, not a straight line.

    On the fourth layer, we must speak fairly: there is indeed some mathematical coupling in the correlation between baseline and change (sharing an error term at the individual score level), but this critique has boundaries—coupling holds at the individual score level but does not affect group-level statistics (PMID 21803544). Therefore, in this topic, coupling poses a threat only at the individual level; the genuine threats at the group level are denominator instability and regression to the mean, not coupling itself.

    On the fifth layer, individuals with high severity are not simply “heavier versions of the same person.” Severity is highly collinear with medication, comorbidities, and age, constituting indication confounding—propensity scores cannot fix unmeasured portions (PMID 24217206). Three probes are available: negative control outcomes (PMID 40838594; PMID 38619297), E-values quantifying “how strong unmeasured confounding must be to explain it away” (PMID 28693043), and vigilance against reverse overcorrection—correcting for mediators deviates from zero rather than converging to it (PMID 42821507). Dichotomizing severity into “heavy/light” additionally loses power and leaves residual confounding (PMID 16217841).

    To operationalize this, four hard constraints apply. Same individual performing grounded/non-grounded crossover, using paired within-subject differences as the response variable, allows each person’s baseline and medication history to serve as their own control—this is the only method that simultaneously removes between-group baseline effects and regression to the mean. Severity entered as a continuous variable, with only a “severity × grounding” interaction term added beyond the main effect, avoiding arbitrary grouping cutoffs. Run all four axes separately, placing the four slopes and their 95% CIs side-by-side in a table; only if they align can it be called dose-response; otherwise, honestly report “incomparable under different severity definitions.” Use restricted cubic splines, fixing the number of knots in advance (3–4), reporting goodness-of-fit and shape tests, rather than substituting “significant slope” for “what the shape is.” Add a healthy baseline group not to compare amplification ratios, but to locate the “zero-response point”—only with a zero does the ratio have a denominator.

    Thus, the verdict consists of three sentences. First, select one dose axis and preregister it; do not mix four axes in a single paper. Second, use absolute change as the primary outcome, treat ratios only as sensitivity analyses, and must simultaneously report denominator variability. Third, choose one of three shape priors (monotonic/threshold/plateau), fit using restricted cubic splines, and compare goodness-of-fit rather than merely reporting slope signs.

    The Yijing perspective rests on the Su Wen · Tong Ping Shi Xun Lun: “When pathogenic qi is excessive, it becomes solid; when vital essence is deprived, it becomes void.” The same individual experiencing worsening symptoms may have increased pathogenic qi or decreased vital essence—two paths that grow identically on the single number of “severity,” yet produce response curves with opposite directions under external intervention. Therefore, traditional models never set dose-response as a straight line; they establish a bifurcation of “treat excess with reduction, treat void with supplementation.” Monotonicity is a modern imposition.

    Evidence and Limitations

    Verified: The association between inflammatory markers and pain severity varies greatly in intensity across conditions, and hs-CRP is unrelated to low back pain (PMID 37195364); inflammation explains only 22%–27% of RA pain variance (PMID 39380043); analgesics can significantly reduce pain without altering TNF-α/IL-1/IL-6 (PMID 28293317); CRP exhibits significantly higher within-individual variability than cholesterol (ICC 0.62, with 51% of highest third at baseline falling into inconsistent categories at follow-up, PMID 22846611); percentage changes can systematically generate false signals of “stronger response at low baseline” (PMID 42203994); atropine dose-response exhibits a threshold pattern (PMID 42070753).

    Partially Verified: Mathematical coupling holds at the individual level but is unaffected at the group level (PMID 21803544), limiting the scope of coupling critiques; low measurement evidence quality for unidimensional pain scales (PMID 34999588) casts doubt on the precision of the “severity” axis itself; negative control outcomes and E-values can detect unmeasured confounding (PMID 40838594; PMID 28693043), but neither has been applied in grounding-impedance scenarios yet.

    To Be Verified: Whether dose-response slopes across VAS, duration, CRP, and QST axes share the same direction in the same dataset; whether “amplification ratio” is systematically inflated at low severity percentiles (denominator effect); and whether curve shapes under restricted cubic spline fitting approximate monotonic, threshold, or plateau patterns more closely.

    Related Reading

    Evidence Level: B
    References: PMID 37195364, PMID 39380043, PMID 28293317, PMID 22846611, PMID 42203994, PMID 42070753, PMID 34999588, PMID 12814613, PMID 21803544, PMID 31333109, PMID 24217206, PMID 40838594, PMID 38619297, PMID 28693043, PMID 42821507, PMID 16217841


    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.