We created a nice graph and a table of optimal conditions, but we can’t reproduce the same results three months later. This is because we didn’t record what raw data we used, which response row was linked to which design row, which terms we excluded and why, what the seed and objective function were.

A package is not a collection of files, but a network of connections.

A reproducible analysis package is not simply putting a PDF report, a data table, and a script in one folder. Each result must be linked to the input and judgment from which it came.

Question → Design → Data → Model → Diagnostics → Confirmation Experiment → Reporting

Each arrow has an identity and a reason for the change. If the run ID of the design table and the observed data do not match, or if we cannot tell which model the final graph came from, the provenance chain is broken.

1. First, fix the question and the experimental unit.

The first line of the analysis receipt is not a statistical technique, but a question.

• primary response and unit
• factor to change and range
• independent experimental unit
• summary rule for technical repeats
• block, batch, and execution order
• success criteria and prohibited conclusions

“24 rows” is less reproducible than “take 3 readings from each of 8 independent culture vessels and analyze the 8 vessel-level means.”

2. The design receipt leaves a record of the agreement before execution.

Items changed after observing the response are saved as amendments, not overwriting the original receipt.

3. Data receipt separates raw data and analysis tables

Raw data preserves the measured values, units, sample ID, and timestamp. The analysis table may be derived data that has undergone exclusion, transformation, and summarization of technical repeats. The transformation between the two should be reproducible by script or explicit rules.

At a minimum, the following should be saved:

• source file hash and the location of the immutable original
• data dictionary and missing code
• exclusion and outlier decisions and reasons
• key linking design rows and response rows
• seed, simulator version, and indication of synthetic data
• environmental information that can affect the results, such as time zone, locale, and software version

4. Model receipt does not hide the selection process

Leaving only the final coefficient table does not allow us to understand why that model was chosen.

• Prior candidate models and hierarchy
• Coding and actual unit equations
• Full model and reduced/alternative models
• Timing and reasons for adding/deleting terms
• ANOVA, LOF, residuals, influential point evidence
• Prediction range and no-extrapolation line
• Goal, limit, and desirability importance

If automatic selection was used, record the method, threshold, candidate terms, and limitations of selection bias. The mere presence of a p-value in the final table does not constitute a prior validation.

5. Figures and tables must also carry model identity

Each contour plot, profiler, residual plot, and simulation histogram must be traceable to the following:

• Source dataset ID
• Model/analysis ID
• Factor settings and fixed values
• Axis scale, unit, and filter
• Generation script or output receipt
• Date of generation and software version

If an image attached to a report is an outdated result that differs from the latest model, it cannot be used, even if the numbers are similar.

6. Confirmation experiments are a separate evidence package

The confirmation experiment has a new run ID, distinct from the fitting data. It records the candidate conditions, prior success criteria, number of independent replicates, block, and observed values. It shows the difference between prediction and confirmation in a table, and if the model was updated, it creates a new version.

The foldover in screening, the RSM optimum confirmation, and the DSD augmentation all have different purposes. They should not be combined into a single "additional experiment" line.

7. The conclusion should include the decision boundary.

A good conclusion states both what the current data says and what it does not say.

Within the specified candidate region and prior quadratic model, x=0.55, y=0.35 was the composite desirability maximum candidate. The average of three new independent confirmations was consistent with the prediction interval. The Monte Carlo pass rate is conditional on the specified input distribution and LSL. Beyond the studied range, long-term capability, and actual product performance are not conclusions of this package.

Avoid using words that erase the scope, such as "optimization complete," "process robust," or "validated."

In-Silico Lab: Recreate the receipt with the same seed.

1. Change the receipt seed and see if the screening, RSM, confirmation, and robustness values are all updated.
2. Change the LSL and confirm whether the decision boundary changes even with the same simulation data.
3. Check if all the design, model, diagnostics, confirmation, and limitation fields are present.
4. Write down why this table alone cannot be used to approve the validity of the actual analysis.

The Lab summarizes the fixed fixture for the U21–U28 engines as a synthetic receipt. The values on the screen are reproducible from the same version, input, and seed, but it does not replace an actual JMP Project, regulatory validation, or organizational electronic record system.

Distinguish the roles of JMP artifacts

The Data Table and table script connect the input and analysis definition, and the model report bundles the numerical, diagnostic, and Profiler outputs. Artifacts such as a Journal or Project can preserve the analysis context, but the validity of the links and the consistency of the execution environment must be verified separately. The focus is on what evidence supports which conclusion, rather than the method of use.

Minimum handoff checklist

1. Are the question, primary endpoint, and independent sample size specified?
2. Are the design matrix, execution sequence, block, and factor with actual units provided?
3. Can the raw data hash and analysis table transformation be reproduced?
4. Are the prior model and all changes preserved?
5. Are the ANOVA, LOF, residual, and prediction ranges linked?
6. Are the optimization goals, limits, and weights specified?
7. Is the confirmation experiment separated from the fitting data?
8. Are the simulation distribution, correlation, seed, and version provided?
9. Do the figures and tables have the source data and model ID?
10. Are the limitations of the conclusions and the next steps documented?

Concluding the section

• Reproducibility is the completeness of the provenance chain, not the number of files.
• Separate and connect the design, data, model, and simulation receipts.
• Do not hide selections, exclusions, and changes behind the final result.
• Confirmation has a separate run and predefined success criteria.
• The decision boundary should clearly state both acceptable conclusions and prohibited extrapolations.
• Documentation and actual system validation are different.

This concludes the current Learning Spine of the KO core course. The next step is not to arbitrarily add new modules, but to translate and thoroughly review the approved structure into EN and JA after validation and publication of U21～U29 KO.

Official Supplementary Materials

• NIST/SEMATECH · Process Modeling
• JMP Help · Design of Experiments Guide
• JMP Help · Enhanced Log in JMP

This article and the Lab are educational synthetic materials and are not evidence or a validated system for actual research, process, quality, or regulatory decisions.