Open comment window: three papers on reverse-calculated pharmaceutical batch records (closes 2026-09-30)

[lizhuojunx86]

I'm opening a six-week comment window on three working papers about data integrity in pharmaceutical manufacturing records.

The papers go to v1.0 in October whether or not anyone comments, and they will say on their face that they were not peer reviewed. This window is what stands in place of peer review, so I would rather it got used than sat here as a formality.

What the three papers claim

1. Dual reverse calculation — English working paper, SSRN 7099138

Across six batches, yields sit at σ ≈ 0.02 percentage points and hug the specification lower bound. Charging masses are recorded to 0.01 kg, which nobody achieves on a factory floor. The claim: those numbers were computed backwards from the specification, not measured forwards. It is the manufacturing analogue of earnings benchmark-beating, and it follows the Graham/Harvey/Rajgopal pattern where people prefer real actions to book adjustments.

2. Layered compliance narrative — Chinese, docs/methodology/09-layered-compliance-narrative-pattern.md

A GMP batch record is three things stacked. A compliance narrative on the surface, largely back-calculated. An unwritten middle layer of what the operators actually decided. A deep layer of embodied craft that nobody writes down because nobody thinks of it as information. The layers are separated by a structural gap, and any project that wants to improve manufacturing with AI has to state which layer it is touching.

3. Sub-channel stratified detection — Chinese, docs/methodology/10-subchannel-stratified-detection-pattern.md

How hard a number is to fake backwards scales as 1 / (spec value × tolerance). In one incoming-inspection lot, the range ratio across three sub-channels came out around 55×, strictly decreasing with magnitude. The corollary is the uncomfortable part: the numbers that are easiest to manipulate are the ones hardest to catch, and those are not two facts but one.

The three questions I actually want answered

1. On the boundary. The evidence is one enterprise, two products. Does that sink the claim, or is an interview-confirmed mechanism enough to put the detection framework forward as a hypothesis worth testing elsewhere?
2. On the three layers. Is the three-layer split operationalizable as a field-research category, or does it collapse in practice into the familiar formal/informal dichotomy?
3. On transfer. Does "difficulty of back-calculation scales as 1 / (target × tolerance)" hold in other compliance-record domains — clinical trial records, emissions reporting, financial close, safety incident logs? If it fails somewhere, that failure is more useful to me than agreement.

A single paragraph on any one of them is a real contribution. You do not need to read the papers to answer question 3.

What happens to a comment

Every comment gets a line in a public disposition log: taken up and where, or not taken up and why. Attribution by GitHub handle unless you ask otherwise. If the window closes with zero comments, v1.0 will say zero, because a window that reports only its successes is not a window.

Data and licence

De-identified: product names, batch numbers, lots, equipment models, companies and people are all pseudonyms. The σ, range ratios and multipliers keep their original values, so the results recompute. CC BY 4.0.

中文

这是 case-2 三篇方法论文章的公开评议窗口，开到 2026-09-30，之后按 v1.0 定稿并在文中显式标注「未经外部同行评议」。

三篇分别讲：批生产记录的数字是从标准反算出来的（σ ≈ 0.02 个百分点、贴着下限、称量精确到 0.01 kg）；批记录是三层制品（合规叙事 / 未写下的实际执行 / 身体化经验），三层之间有结构性代差；以及反算难度与「标准值 × 允许误差」成反比，所以最好操纵的数恰恰最难查——这两件事是同一件事。

想听的三个问题在上面英文部分。第 3 个问题不需要读论文就能答：这条反比关系在别的合规记录领域（临床试验记录、排放申报、财务结账、安全事件台账）成不成立？如果哪里不成立，那比赞同有用得多。