Correction removes what it monitors: a structural model of engineering realisation and the right-truncation of quality records

In the world of large-scale engineering, from the cleanrooms that build microchips to the massive plants that refine fuel, success depends on a chain of events that is often invisible to the naked eye. It begins with a set of requirements, a vision of what a machine or building must do, and ends with a physical object standing in the real world. Between the drawing board and the finished structure lies a process of translation, where ideas are broken down into parts, connected, and assembled. This process is governed by rules of quality: if a part does not fit perfectly, it is fixed, measured again, and the new, corrected measurement is written down. For decades, the industry has operated on the assumption that these records provide a true picture of how well the work is being done. The logic seems sound: if every item that fails is fixed until it passes, then the final list of measurements should show a world where everything is within acceptable limits. But this assumption hides a structural flaw in how we understand the data we collect.

A new study by independent researcher Yang Liu challenges this fundamental belief, proposing that the very act of fixing a mistake erases the evidence of the mistake's existence. The research suggests that when engineers correct a part that is out of tolerance, they are not just improving the product; they are also deleting the most important information about the process that created it. By building a structural model of how engineering projects are realized and testing it with computer simulations, the author demonstrates that the standard way of recording quality data creates a blind spot. The records we keep are not a random sample of reality; they are a filtered version where the most extreme errors have been surgically removed. This removal distorts our understanding of risk, making unstable processes look stable and hiding the true nature of the problems that plague construction and manufacturing.

The paper begins by mapping the journey of a project from a requirement to a physical object using a framework of six distinct steps, or registers. These steps move from the static world of design, where bodies and their connections exist only as ideas, to the evolving world of construction, where those ideas take shape in time and space, and finally to the realm of constraints, where safety and contract rules prune away impossible outcomes. The author identifies a specific gap in this process: a design can show a system that is perfectly continuous on paper, yet impossible to assemble in the real world. In mathematical terms, this is described as a topological defect, a situation where local connections work fine, but the global loop fails to close. This is not a matter of poor workmanship or a lack of checking; it is a structural impossibility inherent in the design itself. The study uses a method of breaking down fields into three components—those that represent structure, those that represent change, and those that represent unchangeable constraints—to show that this failure mode is a precise, measurable property of the design, not just a coordination error.

The core discovery of the paper, however, lies in what happens after the quality control check. In a typical industrial setting, when a measurement falls outside the allowed tolerance, the item is reworked. It is adjusted, measured again, and the new value is stored in the database. The original, failed measurement is discarded. The author argues that this practice creates a "right-truncated" record. Imagine a histogram of all the measurements taken during a project; the standard practice cuts off the right side of the graph where the large errors live and replaces them with values that sit just inside the limit. Because the tail of the distribution—the part that shows how often and how badly things go wrong—is removed, any statistical analysis performed on the remaining data is fundamentally flawed. The data no longer represents the population of errors; it represents only the population of errors that were small enough to pass, or were successfully fixed.

This distortion has a specific and dangerous consequence: it inflates the perceived stability of the process. The study focuses on a statistical measure known as the tail index, which describes how heavy the tail of a distribution is. A lower index means the process is prone to large, unpredictable deviations, while a higher index suggests a more stable, predictable process. In the simulations run for this paper, a process with a true tail index of 1.8, indicating a high risk of large errors, was recorded as having an index of 7.26 after the rework and re-measurement cycle. This is a massive shift. The system that is actually the most unstable appears to be the most stable. The error is not a random scatter of numbers; it is a systematic bias that makes the worst processes look the best. This reversal of ranking means that a quality system relying on these records would inadvertently reward the most dangerous processes and penalize the most stable ones, simply because the dangerous ones have had their worst data points edited out.

The research further explores why simply collecting more data does not solve the problem. Even with a large number of samples, the tail index remains impossible to recover accurately if the original, pre-correction values are lost. The study shows that the interval of uncertainty for the tail index remains so wide that it cannot be used to make critical decisions, such as whether a process is safe enough to continue. The data is not just noisy; the information is missing. The paper argues that the solution is not a better mathematical formula or a larger sample size, but a change in what is recorded. The author proposes a simple schema remedy: for every measurement, the system should store two values that cannot overwrite each other. The first value is the initial measurement taken before any correction is made, and the second is the value after rework. By preserving the initial, failed measurement, the record retains the true shape of the error distribution, allowing engineers to see the actual risk.

The paper also addresses the role of quality codes and standards. Often, these codes are assumed to act as gates that improve the quality of work. The study suggests that comparing work done under a strict code to work done without one is misleading because the two groups are not comparable; strict codes are usually applied to the most critical and well-resourced projects. Instead, the author proposes testing the effectiveness of a code by looking at the rate of exceedance—the proportion of items that fail to meet the tolerance—before and after a rule change, or by varying the intensity of enforcement on the same type of work. The simulations indicate that detecting a real difference in these rates requires a much larger sample size than typically assumed, roughly two thousand items per group, rather than the few hundred often used in industry studies.

Finally, the study connects these engineering findings to a broader question of how machines and models understand the world. The author draws a parallel between the structural gaps in engineering design and the way artificial intelligence models process information. Just as a design can look continuous on paper but fail to assemble, a machine learning model can predict the next frame of a video without truly understanding the underlying structure of the space it is observing. The paper suggests that a true understanding of three-dimensional space requires the ability to separate a field of data into its structural, evolutionary, and constraint components, much like the decomposition used in the engineering model. While the paper does not claim to have solved this for AI, it establishes a testable criterion: a model that can separate these components is likely to have a deeper, more robust understanding of the world than one that cannot.

The implications of this work are practical and immediate. The author concludes that the current practice of overwriting failed measurements is not just a minor administrative habit; it is a structural flaw that prevents organizations from seeing their own risks. The remedy is not complex. It requires a change in the data schema to ensure that the first measurement is preserved alongside the corrected one. Without this simple step, the record cannot answer the question it was collected to answer: how reliable is this process? The study does not claim that the industry is broken, but rather that the lens through which it views its own performance is distorted. By correcting the record-keeping process, engineers can move from a state of false optimism to a clear, accurate understanding of the challenges they face, allowing them to manage risk with the precision that modern engineering demands.