📄 other

Auditable Authenticity: A Four-Domain Conjunctive Rule for Chinese White Porcelain, Benchmarked Against Machine Learning on Published Archaeometric Data

This paper proposes an auditable, non-compensatory conjunctive rule for authenticating Chinese white porcelain across four evidence domains that, while slightly less accurate than black-box machine learning models, achieves comparable performance when restricted to calibrated high-confidence predictions, thereby offering a transparent alternative that exposes the corroboration structure of every judgment.

Original authors: Shidong Li

Published 2026-10-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shidong Li

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine standing in a museum, looking at a bowl of pure white porcelain. To the untrained eye, it is simply a beautiful object. To a specialist, however, it is a document written in clay and fire. The story of the bowl is not just in its shape or how long ago it was made, but in the very atoms that compose it. Scientists have long known that the chemical makeup of a ceramic piece—the specific mix of minerals in its body and glaze—can reveal where and when it was created. This field, known as archaeometry, uses tools like X-ray fluorescence to read these chemical signatures. Yet, a persistent problem remains: the most powerful computer programs that analyze this data are "black boxes." They can tell you with high confidence that a bowl is from a specific kiln, but they cannot explain why. They offer a single answer without showing the evidence, making it impossible for a human expert to verify the logic or understand which part of the object's history led to the conclusion. In the world of heritage, where a wrong guess can mean misidentifying a priceless artifact or failing to spot a clever forgery, knowing the reasoning is just as important as the answer itself.

This is the challenge Shidong Li set out to solve in a new study focused on Chinese white porcelain. The goal was to create a method that is both accurate and transparent, one that mimics the way human experts think without losing the precision of a computer. Experts traditionally authenticate objects by looking for agreement across four distinct lines of evidence: the material (what the clay is made of), the process (how it was formed and fired), the form (its shape and decoration), and time (its historical context). If the chemistry says one thing but the shape says another, a human expert knows to be suspicious; they do not let a strong chemical match cancel out a weak physical one. Li translated this careful, cross-checking logic into a formal rule for a computer. Instead of letting a computer weigh all the evidence together and produce a single score, the new system requires that every available line of evidence points to the same conclusion. It acts like a strict gatekeeper: if even one piece of evidence disagrees or is too weak, the system refuses to make a final claim, flagging the object for further review rather than forcing a guess.

To test this idea, the researcher applied the rule to two real-world collections of porcelain that had already been measured by other scientists. The first group consisted of sixty-nine pieces from the famous Ding kiln, spanning three different historical periods. The second group included thirty-seven pieces from Jingdezhen, divided into four groups based on their era and whether they were made for common use or imperial courts. The computer was asked to sort these pieces into their correct groups using only the chemical data available. The results showed that the new, transparent rule performed remarkably well. On the Ding kiln collection, it correctly identified the period for about ninety percent of the pieces. On the more difficult Jingdezhen collection, it was correct about eighty-two percent of the time. While the most powerful "black box" computer programs were slightly more accurate overall, they could not explain their reasoning. The new rule, however, did something the others could not: it identified a specific subset of predictions where it was absolutely certain. When the system was allowed to say "I am not sure" and skip those difficult cases, the accuracy of the remaining, confirmed answers jumped even higher, reaching over ninety-two percent for the Ding kiln pieces.

Crucially, the study also proved that the method does not invent patterns where none exist. The researcher tested the system on two sets of data where the science suggested there should be no clear answer to be found. In one case, the pieces came from different pits at the same archaeological site; in another, they came from three different kilns that were very close to each other and made at the same time. In both scenarios, the chemical differences were too small to tell the groups apart. The new rule, along with the other computer programs, correctly failed to separate these groups, performing no better than random chance. This is a vital finding because it shows the system is not just guessing or finding false connections in the noise. It respects the limits of the evidence. If the chemistry does not hold the answer, the system admits it, rather than confidently giving a wrong answer.

The study concludes that while this transparent method is not the absolute most powerful predictor for every single problem, it offers a vital advantage for authentication: it exposes the structure of the judgment. It allows an expert to see exactly which lines of evidence agreed and which ones were missing. The researcher notes that the current tests relied mostly on chemical data because detailed records of the object's shape and firing process are rarely published alongside the chemical results. The ultimate goal is a future where every analyzed object comes with a complete package of data—chemistry, physical measurements, shape descriptions, and historical context—allowing this four-part check to be used in its full power. For now, the work demonstrates that it is possible to build a computer system that does not just give an answer, but shows its work, bringing the rigor of scientific verification back into the hands of human experts.

Technical Summary: Auditable Authenticity in Chinese White Porcelain

Problem Statement
Scientific authentication of heritage objects faces a fundamental tension between two requirements: predictive accuracy and auditability. High-performing machine learning classifiers (e.g., kernel SVMs, random forests) often function as "black boxes," providing a single class label without revealing the specific lines of evidence supporting that decision. Conversely, traditional domain-expert connoisseurship is inherently auditable, relying on the convergence of independent evidence domains (material, process, form, and time), but its decision thresholds are tacit, difficult to quantify, and resist rigorous validation. The paper addresses the challenge of formalizing expert practice into a computable rule that maintains auditability without sacrificing too much accuracy, while ensuring the method does not manufacture false signals in data where none exist.

Methodology
The author proposes a Four-Domain Conjunctive Rule that formalizes the ceramics-specialist practice of "joint confirmation."

The Four Domains: Evidence is partitioned into four causal stages of production:
1. Material (M): Chemical and mineralogical composition (e.g., body and glaze oxides).
2. Process (P): Physical traces of forming and firing (e.g., water absorption, porosity, kiln furniture impressions).
3. Form (F): Morphological features (e.g., vessel shape, foot treatment, decoration).
4. Time (T): Stratigraphic context and chronological positioning.
The Conjunctive Logic: Unlike weighted-sum approaches that allow a strong signal in one domain to compensate for a weak or contradictory signal in another, this model uses a non-compensatory minimum operator. For a candidate class $k$ , the confidence $C(k)$ is the minimum of the confidences $c_d(k)$ across all populated domains $d$ . This enforces the logic that "all four must hold."
Audited Prediction: A prediction is flagged as "fully audited" only if:
1. Agreement: Every populated domain independently points to the same class.
2. Sufficient Strength: The conjunctive confidence exceeds a calibrated threshold $\delta$ .
  Objects failing these conditions are abstained from (flagged for further evidence) rather than forced into a class.
Data-Driven Calibration: The audit threshold $\delta$ and distance scales are not set by hand. They are learned via nested cross-validation and global optimization (differential evolution) to maximize macro-F1 while maintaining a minimum coverage of audited predictions (0.6).
Benchmarking: The rule is tested against six baselines (Majority Class, LDA, Regularized Logistic Regression, kNN, RBF-SVM, Random Forest) on two real-world datasets:
1. Ding Kiln (N=69): Three production phases (Five Dynasties, Northern Song, Jin) classified via LA-ICP-AES body composition.
2. Jingdezhen (N=37): Four groups (Yuan common, Yuan imperial, Ming imperial, Qing imperial) classified via EDXRF body composition.
Null Tests: Two deliberately difficult cases (Wanquansi Jin-dynasty deposit contexts and Hebi Yuan-dynasty kiln glazes) are used to verify that the method does not overfit or manufacture separation where archaeology predicts none.

Key Results

Performance on Signal Tasks:
- Ding Kiln: The conjunctive rule achieved a macro-F1 of 0.905 (balanced accuracy 0.909), significantly outperforming the majority baseline (0.202) and approaching the best black-box classifier, the SVM (0.941).
- Jingdezhen: The rule achieved a macro-F1 of 0.823 (balanced accuracy 0.793), again far exceeding the baseline (0.175) but trailing the SVM (0.983). The gap is attributed to the rule's inability to model fine non-linear boundaries between small, similar classes (e.g., Yuan imperial vs. common).
Audit Efficacy: The data-calibrated audit gate successfully filtered for high-reliability predictions.
- For Ding, 87% of predictions passed the audit, with accuracy rising from 0.900 (all samples) to 0.923 (audited subset).
- For Jingdezhen, 97% passed, with accuracy rising from 0.892 to 0.894.
Null Test Performance: On the Wanquansi and Hebi tasks, where archaeological evidence suggests no chemical signal exists for the target classes, all methods (including the conjunctive rule and black boxes) correctly returned chance-level performance (balanced accuracy $\approx$ 0.50). This confirms the method does not hallucinate structure.

Significance and Claims
The paper makes a modest but specific claim: An interpretable, corroboration-exposing classifier can approach black-box accuracy on real archaeometric tasks while explicitly exposing the structure of every judgment.

Not a Complete System: The author explicitly states they do not claim this is a complete four-domain authentication system. Due to data availability, the primary empirical tests relied almost exclusively on the Material (M) domain, with the rule degrading gracefully to a single-channel nearest-prototype classifier when other domains were missing.
Value of Auditability: The primary contribution is not superior raw accuracy (where black boxes like SVMs still win), but the ability to identify a high-reliability subset of judgments where independent evidence lines converge. This allows experts to see why a decision was made and to identify which specific domain (e.g., form or process) caused a failure in corroboration.
Future Requirement: The paper concludes that the next necessary step is not further algorithmic simulation, but a change in publication standards: the shared release of joinable per-object data covering Material, Process, Form, and Time domains simultaneously. Only with such data can the "genuine four-dimensional hypothesis" be empirically tested.

In summary, the work demonstrates that formalizing expert logic into a non-compensatory, data-calibrated rule provides a transparent alternative to black-box classification, offering a practical mechanism for "auditable authenticity" in heritage science.

Technical Summary: Auditable Authenticity in Chinese White Porcelain

More like this