同一份文件里两句话互相打架:不靠任何外部数据,也能判死一个衰减数值

Written by

in

Language

Evidence level: B

先说结论

一份讲”山体能不能挡住极低频电磁场”的文章里,有两句话是这么写的:

  • 正文说:1 km 厚花岗岩,衰减约 0.27 dB;
  • 同文表格说:100 m 厚花岗岩,衰减 27 dB。

厚度相差 10 倍,衰减却相差 100 倍。而电磁波在导体里走,衰减和厚度是成正比的——厚 10 倍,衰减就该大 10 倍。所以这两句话里必有一句错,且不是”大概差一点”,是差了一千倍。

关键不在于”算出了什么”,而在于”怎么发现的”:判死它,不需要查任何文献,不需要知道花岗岩的电导率到底该取多少,不需要任何地下实测。只需要这一份文件里的两个数字,和一个比例关系。

这条检查现在已经写进审计工具(delta-table-scan.py v1.4 的 R5 项),全库自动跑。


为什么这条检查值钱:它绕开了”参数争议”

前面几轮审计一直在跟一个纠缠不清的问题打架:到底是谁错?

δ/A 表里出现不自洽,有两种可能:

  1. δ 或 A 算错了(数值问题);
  2. σ 或 μr 本来就该取另一个值(参数问题)。

第 2 种情况下,”不自洽”其实不是错——只是表格本意用了一个没写出来的等效参数。这就是前几轮反复出现”判了又推翻”的原因:第 12 轮曾判定”A 档 δ 被高估 37 倍”,第 13 轮自我证伪,发现真错在 C 档。

R5 绕开了整个争议:它不比较”你写的值”和”我算的值”,它比较你自己写的两个值之间的比例。只要同一份文件里、同一材料、同一频点下出现两个”厚度–衰减”对,比例关系就必须成立——无论 σ 取多少。

这是审计里很少见的一种判据:零外部依赖。它的误报率天然低,因为它不问”你对不对”,只问”你前后一致吗”。


那道比例关系有多硬

对于良导体中的平面波,趋肤深度

$$\delta = \sqrt{\frac{2}{\omega \mu_0 \mu_r \sigma}}$$

而厚度 $d$ 上的衰减(以 dB 计)近似为

$$A_{dB} \approx 8.7 \cdot \frac{d}{\delta}$$

把两式对各参数求偏导就能看出:在 δ 固定时,$A$ 对 $d$ 是严格一次齐次的,即

$$\frac{A_1}{A_2} = \frac{d_1}{d_2}$$

这个比例与 σ、μr、频率全部无关——它们都在分子分母里约掉了。这就是为什么这条检查不需要任何参数。

顺便说一句,这个”一次齐次”性质也是本站另一篇(第 14 轮)关于”端点方向”的物理来源:因为 $A \propto \sqrt{\sigma}$,所以 σ 增大时 A 增大、δ 减小,两者方向相反。这次用的是它的厚度维度,上次用的是它的电导率维度——同一条式的两个侧面。


实际跑出来的结果:全库只有 1 处

工具在 780 全库(37 个文件、400 条数值断言)里找出 43 个可参与比例检验的点,从中构成 4 组”同材料、同频点、厚度至少差 2 倍”的配对,最终报出 1 处违反:

项 内容
文件 skin-depth-basement-building-mu-metal-table.md
材料 / 频点 花岗岩 @ 7.83 Hz
点 A L76:厚 100 m,A = 27 dB
点 B L80:厚 1000 m,A = 0.27 dB
厚度比 10×
衰减比 100×
偏差 10×(判据容差 2×)

这正是第 13 轮靠人工对读才看见的那一处。当时是靠”散文 vs 表格”逐字对照发现的,属于工具盲区边缘;现在工具能自动复现,且只报这一处,没有顺带误报一片。

一个反直觉的地方

这份文件的立论是”山体对极低频几乎透明,所以靠山屏蔽不成立“。按复算(花岗岩 100 m ≈ 0.048 dB、玄武岩 100 m ≈ 1.53 dB),真实值比表里的 27 / 270 dB 更透明——也就是说,这批错把正确结论写得对它自己不利。

所以按”看是否产生错误结论,不看倍数”的原则,不推翻任何结论,但数值必须改。这一点很重要:审计的价值不是”抓错显得厉害”,而是分清”错了但结论稳”和”错了且结论塌”。


顺带修掉的一类更隐蔽的缺口:静默不入账

做这条检查时撞上一个更麻烦的问题。

工具的覆盖率一直是这么报的:”含 δ/A 列的表格 68 张,成功解析 45 张,未解析 23 张”。但实际情况是:有一批表既不算”已解析”,也不算”未解析”,它们根本没有进入分母。

原因是表头写法。工具认 A(dB)、认 1 km A(dB),但不认 f=7.83 Hz (A, dB) 这种带逗号的写法。而全库唯一那处最严重的错表(就是上面那张),表头恰好就是这个写法。于是它:

  • 不进”已解析”(解析不出断言);
  • 也不进”未解析”(因为它连”这是 δ/A 表”都没被认出来,在分类那一步就被跳过了);
  • 报告里一个字都不提。

这比”未解析”危险得多。”未解析”至少留下了一条可见的缺口;”静默不入账”连缺口都看不见。 修复后这类表被单独计数并列出(当前为 0),同时英文镜像里 $\delta$(7.83 Hz)、$\sigma$ (S/m) 这类 LaTeX 表头也一并纳入识别。

修复的代价是:断言数 355 → 400(+45),表格分母 60 → 68,丢失 0 条。覆盖率从 68.3% 降到 66.2%——这不是退步,是因为分母里补进了原来被藏起来的表。覆盖率口径跨版本不可直接比较,这条规矩第 14 轮已经立过,这次又验证了一遍。


修的时候自己造的三个假阳性(如实记)

必须写下来,因为这是同一个失败模式的第 N 次复发:放宽匹配范围时没有同步收紧排除。

# 我做了什么 造出了什么假阳性 怎么修的
1 归一 LaTeX 表头后,$\\sigma\mu_r$ / $\\sigma\mu_r$ 这类比值列被认成 σ 列 取到 0.0996 这种”比值”当电导率,整行 δ 被判错 1 万倍 排除 implied / ratio / 比值 / 反推 / 推算
2 散文中”抓第一个长度当厚度” 回看到了上一行表格里的厚度,或把 ASCII 图里标注的 δ 值当厚度 取消”第一个长度”启发式,改为三类显式绑定
3 HAS_DB = r"dB" 无词边界 sandbox 里的 db 命中,一份无关文档被列为”含 dB 的表” 加词边界 (?<![A-Za-z])dB(?![A-Za-z])

三个都不是读代码发现的,全部靠”① selftest 边界用例 ② 与上版基线逐条 diff 新出现的误报”。第 14 轮立的规矩”改审计工具前先留基线快照”,这一轮又救了三次。


一个方法学副产品:判据”充分性”与判据”必要性”要分开

这次还澄清了一件事,值得单独写下来。

R5 这类内部一致性检查,只能判”不成立”,不能判”成立”:

  • 两处比例对得上 → 不能推出这两个数都对。它们可能用同一个错的 δ 算出来,错错相消。
  • 两处比例对不上 → 必然至少有一处错。

本文那张 μ-metal 行就是活例子:它在 δ 表里被判”OK”,恰恰因为它用的是同一个错的 δ(0.57 cm)——§2 与 §3 错错相消。所以”通过 R5″不等于”没问题”,只等于”没自相矛盾”。

这条区分对应审计里的一个老问题:必须把”证伪能力”和”验证能力”分开报。一个检查项如果只有证伪能力,就不能拿它的”通过数”当质量指标。


证据与局限

证据等级 E3(结构类比 + 库内文本核算),不是 E1。理由与边界:

  1. 全部为解析复算,无任何实测数据。本文 4 个 DOI 均为地物/材料类期刊(Studia Geophysica et Geodaetica、Geophysics、Water Resources Research、Clays and Clay Minerals),不在 PubMed 收录范围,故无 PMID;已逐条 Crossref 回查(标题/期刊/年份/作者全部一致)。
  2. “自洽值”不是”真值”。$A \approx 8.7d/\delta$ 与 $\delta=\sqrt{2/(\omega\mu\sigma)}$ 均假设均匀半空间、无位移电流、μr 与 σ 为常数。真实山体分层、各向异性、σ 随频率弥散(上述四篇均支持频散存在),所以任何”正确值”都只是在给定假设下的自洽值。
  3. “错”的判定范围很窄:仅指”同一文件内两处数字违反了 A ∝ d”。若原文本意是用某个未写出的等效 σ,问题性质从”算错”变为”没写清 σ”——这一点本文无法区分。
  4. 样本很小:43 个可判点里只构成 4 组有效配对,判据容差 2× 也是人为设定(选 2× 是因为它比观测到的 10× 偏差小一个量级,留了余量)。4 组配对不能外推为全库的自相矛盾率。
  5. 只覆盖等式断言。不等式断言(A ≥ 10 dB)不参与比例检验——因为不等式两端的最不利端语义会混进来,那是另一套判据(第 14 轮 R7/R8 已处理)。全库 400 条断言里有 23 条不等式,被本条检查有意排除。
  6. 工具只认 markdown 表格 + 特定散文句式。ASCII 图、纯叙述型描述仍在盲区;本文唯一那处发现是”表格 + 散文”跨形态对照,这恰好在工具能力的边缘。
  7. 覆盖率口径不可跨版本比:v1.3 的 68.3% 与 v1.4 的 66.2% 分母不同(补进了原来静默跳过的表),数字下降不代表质量下降。

下一步

  1. R5 扩到”跨文件同材料”:目前只在单文件内互检。若两个文件对同一材料给出不同的”厚度–衰减”关系,同样违反 A ∝ d,但目前查不到。
  2. R5 与 R7 合并成一张”断言充分性矩阵”:R7 管端点方向(σ 区间),R5 管厚度比例,两者都是”零参数内部检查”,合并后可以统一报”这条断言的可判定性有几成”。
  3. 已发布正文的数值更正仍等 JOE 批(D1′/D1/D2/D7 一组),本文不改任何已发布正文。

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 contradictory sentences in the same document: without any external data, you can still determine a decay value.

Evidence level: B

Conclusion

A document discussing “whether a mountain can block extremely low-frequency electromagnetic fields” contains two contradictory statements:

  • The main text states: 1 km thick granite attenuates approximately 0.27 dB;
  • The same document’s table states: 100 m thick granite attenuates 27 dB.

The thickness differs by a factor of 10, yet the attenuation differs by 100 times. For electromagnetic waves propagating in conductors, attenuation is proportional to thickness—increasing thickness by 10 times should increase attenuation by only 10 times. Therefore, one of these statements must be incorrect, and not merely “off by a little,” but off by a factor of 100.

The key issue is not “what was calculated,” but “how it was discovered”: determining this inconsistency requires no external literature, no knowledge of the exact conductivity value for granite, and no underground measurements. It relies solely on two numbers within this single document and their proportional relationship.

This check is now implemented in the audit tool (delta-table-scan.py v1.4, item R5) and runs automatically across the entire database.


Why This Check Is Valuable: It Bypasses “Parameter Controversy”

Previous audit rounds struggled with an unresolved question: Who is wrong?

When the δ/A table shows self-inconsistency, there are two possibilities:

  1. δ or A was calculated incorrectly (numerical error);
  2. σ or μr should have taken a different value (parameter issue).

In the second case, “self-inconsistency” is not necessarily an error—it may simply reflect that the table intended to use an equivalent parameter not explicitly stated. This explains why previous rounds repeatedly “judged and then overturned” results: Round 12 concluded that “A-grade δ was overestimated by 37 times,” but Round 13 self-refuted this, finding the real error lay in C-grade.

R5 bypasses the entire controversy: it does not compare “your stated value” with “my calculated value.” Instead, it compares the ratio between two values you yourself stated. As long as the same document, same material, and same frequency point contain two “thickness–attenuation” pairs, the proportionality must hold—regardless of σ.

This is a rare criterion in auditing: zero external dependency. Its false-positive rate is naturally low because it does not ask “are you right?” but only “are you consistent with yourself?”


How Strong Is This Proportionality?

For plane waves in good conductors, the skin depth is:

$$\delta = \sqrt{\frac{2}{\omega \mu_0 \mu_r \sigma}}$$

and the attenuation over thickness $d$ (in dB) is approximately:

$$A_{dB} \approx 8.7 \cdot \frac{d}{\delta}$$

Taking partial derivatives of both equations with respect to their parameters reveals: when δ is fixed, $A$ is strictly linearly homogeneous with respect to $d$, i.e.,

$$\frac{A_1}{A_2} = \frac{d_1}{d_2}$$

This ratio is independent of σ, μr, and frequency—they all cancel out in the numerator and denominator. This is why this check requires no parameters at all.

Incidentally, this “linear homogeneity” property is also the physical origin of another article on this site (Round 14) concerning “endpoint direction”: because $A \propto \sqrt{\sigma}$, increasing σ increases A while decreasing δ, causing them to move in opposite directions. This time the check uses the thickness dimension of this relationship; previously it used the conductivity dimension—two complementary aspects of the same formula.


Actual Results: Only One Violation Found

The tool identified 43 points eligible for proportionality testing across the entire 780-document corpus (37 files, 400 numerical assertions). From these, it constructed 4 pairs of “same material, same frequency, thickness differing by at least 2×,” ultimately reporting 1 violation:

Item Content
File skin-depth-basement-building-mu-metal-table.md
Material / Frequency Granite @ 7.83 Hz
Point A L76: thickness 100 m, A = 27 dB
Point B L80: thickness 1000 m, A = 0.27 dB
Thickness ratio 10×
Attenuation ratio 100×
Deviation 10× (tolerance is 2×)

This is precisely the one location detected manually in Round 13. At that time, it was found through careful word-by-word comparison of “prose vs. table,” lying at the edge of the tool’s blind spot; now the tool can automatically reproduce it, and reports only this single violation without any spurious false positives.

A Counterintuitive Aspect

The document’s argument is that “mountains are nearly transparent to extremely low frequencies, so shielding by mountains does not work.” Recalculating (granite 100 m ≈ 0.048 dB, basalt 100 m ≈ 1.53 dB), the true values are actually more transparent than the table’s 27 / 270 dB—meaning this batch of errors wrote the correct conclusion in a way that contradicts itself.

Therefore, following the principle of “judging whether an error leads to a wrong conclusion, not just a wrong ratio,” no conclusions are overturned, but the numerical values must be corrected. This distinction is crucial: the value of auditing lies not in “catching errors to appear impressive,” but in distinguishing between “errors that do not affect the conclusion” and “errors that collapse the conclusion.”


A More Subtle Gap Fixed: Silent Tables Excluded from Denominator

During this check, a more problematic issue emerged.

The tool’s coverage reporting has always been: “68 tables containing δ/A columns, 45 successfully parsed, 23 unparsed.” But in reality: a batch of tables is neither counted as “parsed” nor “unparsed”—they never entered the denominator at all.

The reason lies in header formatting. The tool recognizes A(dB) and 1 km A(dB), but does not recognize headers like f=7.83 Hz (A, dB) that include commas. The most serious error table in the entire corpus (the one above) happens to use exactly this format. Consequently, it:

  • Does not enter “parsed” (cannot extract assertions);
  • Does not enter “unparsed” (because it wasn’t even recognized as a δ/A table during the classification step, so it was skipped entirely);
  • Appears nowhere in the report.

This is far more dangerous than “unparsed.” “Unparsed” at least leaves a visible gap; “silent exclusion” hides the gap entirely. After fixing, such tables are now counted separately and listed (currently 0), and the English mirror now also recognizes LaTeX headers like $\delta$(7.83 Hz) and `$\sigma$ (S/m).

The cost of this fix: assertion count increased from 355 to 400 (+45), table denominator from 60 to 68, with zero lost entries. Coverage dropped from 68.3% to 66.2%—this is not a regression, because the denominator now includes tables previously hidden. Coverage metrics across versions cannot be directly compared; this rule was established in Round 14 and verified again here.


Three False Positives Created During Fixing (Recorded Honestly)

Must document these, as this represents the Nth recurrence of the same failure mode: expanding match scope without tightening exclusion criteria.

# What I Did What False Positive It Created How Fixed
1 After normalizing LaTeX headers, ratio columns like $\sigma\mu_r$ / \sigma\mu_r$ were mistaken for σ columns Used 0.0996 such “ratios” as conductivity, causing δ to be judged wrong by 10,000× Exclude terms containing implied / ratio / 比值 / 反推 / 推算
2 In prose, “grab the first length as thickness” Repeatedly saw thickness from the previous table row, or mistook δ values annotated in ASCII diagrams as thickness Cancel the “first length” heuristic, replacing it with three explicit binding types
3 HAS_DB = r"dB" lacked word boundaries Matched db inside “sandbox,” causing an unrelated document to be flagged as “containing dB tables” Add word boundaries (?<![A-Za-z])dB(?![A-Za-z])

None of these were discovered by reading code; all relied on “① self-test boundary cases ② diff against the baseline from the previous version to catch newly appearing false positives.” The rule established in Round 14—”save a baseline snapshot before modifying the audit tool”—saved us three times again this round.


A Methodological Side Effect: Distinguish “Sufficiency” from “Necessity” of the Criterion

This round also clarifies an important point worth documenting separately.

R5 and similar internal consistency checks can only determine “does not hold,” not “holds”:

  • If two ratios match → cannot conclude both numbers are correct. They might both be derived from the same wrong δ, canceling each other’s errors.
  • If two ratios mismatch → necessarily at least one is wrong.

The μ-metal row in this document exemplifies this: it passed R5 in the δ table precisely because it used the same wrong δ (0.57 cm)—§2 and §3 errors canceled each other out. Thus “passing R5” does not mean “no problem,” only “no self-contradiction.”

This distinction corresponds to a longstanding issue in auditing: must report “falsification capability” and “verification capability” separately. If a check item has only falsification capability, its “pass count” cannot serve as a quality indicator.


Evidence and Limitations

Evidence level E3 (structural analogy + in-database text recalculation), not E1. Reasons and boundaries:

  1. All are parsed recalculations, with no measured data. The four DOIs cited in this article (Studia Geophysica et Geodaetica, Geophysics, Water Resources Research, Clays and Clay Minerals) are all earth/materials journals, not included in PubMed, hence no PMID; each was cross-checked via Crossref (title/journal/year/author all match).
  2. “Self-consistent values” are not “true values”. $A \approx 8.7d/\delta$ and $\delta=\sqrt{2/(\omega\mu\sigma)}$ both assume a uniform half-space, no displacement current, and constant μr and σ. Real mountains are layered, anisotropic, and σ exhibits frequency dispersion (all four cited papers support the existence of dispersion), so any “correct value” is only self-consistent under given assumptions.
  3. The scope of “wrong” is narrow: only refers to “two numbers within the same document violating A ∝ d.” If the original intent was to use an unstated equivalent σ, the nature of the problem shifts from “calculation error” to “unclear σ specification”—a distinction this article cannot resolve.
  4. Sample size is small: among 43 testable points, only 4 pairs formed valid matches; the tolerance of 2× is artificially set (chosen because it is one order of magnitude smaller than the observed 10× deviation, leaving margin). Four pairs cannot be extrapolated to represent the overall self-contradiction rate of the corpus.
  5. Only covers equality assertions. Inequality assertions (e.g., A ≥ 10 dB) do not participate in proportionality testing—because the worst-case semantics of both ends would mix, requiring a separate criterion (R7/R8 already handled this in Round 14). Among the 400 assertions in the corpus, 23 are inequalities and were intentionally excluded by this check.
  6. The tool only recognizes Markdown tables + specific prose patterns. ASCII diagrams and purely narrative descriptions remain blind spots; the single discovery in this article involved “table + prose” cross-format comparison, which lies exactly at the edge of the tool’s capability.
  7. Coverage metrics cannot be compared across versions: v1.3’s 68.3% and v1.4’s 66.2% have different denominators (the latter includes tables previously silently skipped), so the numerical drop does not indicate quality decline.

Next Steps

  1. Expand R5 to “cross-file same material”: currently only checks within single files. If two files assign different “thickness–attenuation” relationships for the same material, this also violates A ∝ d, but cannot yet be detected.
  2. Merge R5 and R7 into a single “assertion sufficiency matrix”: R7 manages endpoint direction (σ intervals), R5 manages thickness ratios; both are “zero-parameter internal checks.” Merging them will uniformly report “to what extent is this assertion determinable.”
  3. Numerical corrections in published prose await JOE approval (D1′/D1/D2/D7 group); this article makes no changes to already-published prose.

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.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *