Написание статей · Начальный
A complete worked paper using synthetic data
Follow a complete educational paper with a verified synthetic paired dataset.
24 мин чтения · 26 окт. 2026 г.
Эта страница ещё не переведена, поэтому вы читаете английскую версию.
Главное
- Trace the question through every section.
- Reproduce the reported calculations.
- Keep teaching values distinct from empirical evidence.
Содержание
Follow a complete educational paper with a verified synthetic paired dataset.
A paper becomes easier to write when you can see how its parts depend on one another. The question determines the comparison. The comparison determines the unit and analysis. The results establish what was observed. The discussion explains what those observations can and cannot support. A review follows the same need for a defensible argument, but its evidence comes from existing studies rather than newly collected observations.
This tutorial contains a complete engineering teaching paper, annotated adaptations for natural science and social science, and a substantial review example. All domain data, scenarios and study summaries are synthetic or hypothetical. No experiment, survey or literature search on these domain questions was performed. Real references support the writing and statistical methods only. They do not verify the invented domain findings.
Read the model sections as examples of reporting, then read the teaching notes to understand the choices. Do not copy a model into a submission and replace a few nouns. Your own manuscript must describe the actual protocol, observations, permissions, analysis and literature.
Before drafting: the reporting contract
For the controller example, the reporting contract is: estimate the mean difference between two measurements paired within each of twelve constructed cases. The outcome is positioning error in millimetres. The direction is controller A minus controller B, so a positive difference means a smaller B value. The dataset is deliberately simple enough to inspect completely.
The example uses an Introduction, Methods, Results and Discussion structure. This organisation is common for empirical papers and is described in ICMJE guidance for medical journals, although requirements differ by discipline and publication. Review articles may use thematic sections. Whatever the headings, readers still need the question, evidence, reasoning and boundaries. S01
Model paper begins
Title
Reporting a paired comparison of controller error: a synthetic teaching example
Abstract
Comparisons between navigation controllers require an explicit evaluation unit, outcome and treatment of matched observations. This educational analysis illustrates a paired comparison using twelve constructed layout cases. Each case contains a positioning error value for controller A and controller B, labelled in millimetres. We calculated the within case difference as A minus B and estimated its mean with a conventional t based 95% confidence interval. Mean error was 20.33 mm for A and 16.75 mm for B. The mean paired difference was 3.58 mm, with a 95% interval from 2.95 to 4.22 mm. All twelve constructed differences were positive. The example demonstrates consistent reporting of pairing, individual observations, effect magnitude and statistical uncertainty. Because no simulator or physical controller generated these values, the analysis provides no empirical evidence of controller performance or generalisability.
Introduction
A comparison between two systems is interpretable only when the outcome and evaluation unit are clear. Consider an evaluation in which both controllers are assessed on the same layout cases. Difficult cases may produce larger errors for both controllers. Matching the measurements within each case preserves information about the comparison under shared conditions. Analysing the two columns as unrelated groups would answer the question less directly and could misrepresent the relevant uncertainty.
The purpose of this paper is educational. We demonstrate the reporting of a paired mean difference using a small constructed dataset whose observations can all be inspected. The analysis concerns a calculation and its communication, rather than an unresolved technical question about navigation. The paired calculation follows the procedure described in the NIST statistical handbook. S28
Our objective was to calculate the mean A minus B difference across the twelve teaching cases and to report its variability and conventional uncertainty interval. We also calculated a paired t statistic for a null mean difference of zero. The constructed cases do not constitute a random sample of physical environments. Consequently, calculations involving sampling uncertainty are demonstrations under an assumed model, not validated statements about a real target population.
Methods
Data provenance and analysis unit. The dataset contains twelve case identifiers and two error values per case. The values were created for teaching; they are neither measurements from a robot nor outputs from an implemented simulator. Each row is treated as one independent case for the calculation. The two observations in that row are paired. There are twelve paired differences, not twenty four independent experimental units.
Outcome and comparison. The supplied column labels describe mean positioning error in millimetres for controllers A and B. No underlying trajectory or measurement procedure exists in this teaching dataset. The analysis therefore uses the provided case summaries directly. We defined each difference as A minus B. This definition fixes the interpretation of its sign before calculation.
Data checking. We included all twelve rows. Case identifiers were unique, and every case contained both values. There were no missing observations or excluded cases. We verified that the supplied reduction column equalled A minus B in every row. No additional outcomes, subgroups or controller variants were analysed.
Descriptive analysis. We calculated the sample mean and sample standard deviation for each controller and for the paired differences. The standard deviation of the differences uses eleven degrees of freedom. Its standard error is the standard deviation divided by the square root of twelve.
Interval and hypothesis calculation. We calculated the conventional t based 95% confidence interval as the mean difference plus or minus the 97.5th percentile of a t distribution with eleven degrees of freedom multiplied by the standard error. The paired t statistic compares the mean difference with zero. The interval and test require an appropriate independence and sampling model; for exact small sample t inference, normally distributed differences are assumed. Here they illustrate the calculations on constructed values. The formula and interpretation of confidence limits are described by NIST. S14
Scope of inference. The dataset is small, its differences are discrete, and the numbers were chosen rather than randomly generated under a specified population model. We did not regard the t interval as an empirically validated coverage statement for controller performance. No sample size justification, random allocation, calibration, controller tuning or simulation validation was performed because no domain experiment was conducted.
Data availability. All observations appear in Table 1 and in data/synthetic_controller_data.csv supplied with this collection. The associated calculation record is data/verified_calculations.json. Readers can reproduce the summary directly from the table.
Results
Table 1. Complete constructed controller dataset. Error values are labelled in millimetres. Each row is one case, and the difference is A minus B. The identifiers indicate pairing and are not measured quantities.
| Case | A error, mm | B error, mm | A minus B, mm |
|---|---|---|---|
| 01 | 18 | 14 | 4 |
| 02 | 21 | 18 | 3 |
| 03 | 17 | 13 | 4 |
| 04 | 24 | 20 | 4 |
| 05 | 20 | 17 | 3 |
| 06 | 19 | 15 | 4 |
| 07 | 22 | 19 | 3 |
| 08 | 16 | 14 | 2 |
| 09 | 25 | 20 | 5 |
| 10 | 23 | 18 | 5 |
| 11 | 18 | 16 | 2 |
| 12 | 21 | 17 | 4 |
Mean error was 20.33 mm for A, with a sample SD of 2.84 mm, and 16.75 mm for B, with a sample SD of 2.38 mm. Every constructed case had a smaller B value. The paired differences ranged from 2 to 5 mm and had a mean of 3.58 mm, SD of 1.00 mm and standard error of 0.29 mm after rounding.
The conventional t based 95% interval for the mean difference was 2.95 to 4.22 mm. The paired t statistic was 12.46 with eleven degrees of freedom. The computed two sided p value was approximately 7.89 × 10⁻⁸. These values describe the mathematical result obtained from the constructed data under the specified calculation.
Figure 1. Individual pairs and their mean difference. Each line joins A and B values from the same constructed case. The right panel shows the mean paired difference and its conventional t based 95% confidence interval. Positive differences indicate smaller B values. The analysis contains twelve paired cases. All data are synthetic; the display is not evidence of an actual controller improvement.
Discussion
This example demonstrates how a manuscript can retain the structure of its comparison from question to figure. The difference was defined within each case, the calculation used those differences, and the reporting stated both the magnitude and uncertainty. The individual pairs reveal that all constructed cases have the same direction of difference. The mean alone would not show that pattern.
The calculation does not establish that controller B performs better in practice. The values were constructed with smaller B errors, and there was no physical sampling, simulation implementation or independent test environment. A small p value cannot repair those absences. More generally, statistical significance does not establish the size or importance of an effect, as emphasised in the ASA principles. S13
A real evaluation would need a target set of environments and a justified method for selecting cases from it. It would also need a precise error definition, a reference trajectory, controller implementations, tuning procedures, sensor assumptions, computational resources and a rule for handling failures. If multiple runs were performed within each layout, the analysis would need to distinguish variation between layouts from variation between runs. Treating every recorded time point as an independent case would not preserve that structure.
Practical value also requires a benchmark. A 3.58 mm reduction might matter in one application and be negligible in another. This example supplies no application specific requirement, cost information, failure probability or energy measurement. It therefore cannot support a claim that B is preferable overall. A future real study would specify which performance differences matter before choosing an analysis and would evaluate the relevant tradeoffs.
Conclusion
A transparent paired comparison identifies the case, defines the subtraction direction, reports the individual data, and gives an effect estimate with an appropriate account of uncertainty. This constructed example demonstrates that reporting sequence. It does not validate any controller.
Declarations
Research status: This is educational material, not an empirical engineering study. Data: All values are constructed and disclosed. People and animals: No participants or animals were involved in creating the numerical example. Authorship and support: No named researcher, institution, grant or ethical approval is implied. Production: The teaching text and figures were prepared with AI assistance and computational checking. These declarations describe this example and must not be copied into an actual study without verification.
Model paper ends: inspect the decisions
The abstract names the synthetic status at the beginning and the inferential boundary at the end. The introduction explains the methodological problem without inventing a literature gap about robotics. The Methods state what was actually done with the data, including what was not available. The Results describe observations and calculations. The Discussion explains their relevance and limits.
Notice the distinction between an omitted detail and a nonexistent procedure. A real engineering Methods section should report a simulator version if a simulator was used. This example must not invent a version merely to look complete. Transparent reporting sometimes means acknowledging that an example cannot demonstrate a particular part of the research workflow.
Reproduce the central calculation
The twelve differences sum to 43 mm. Dividing by twelve gives 3.583333 mm. Their sample SD is 0.996205 mm, so the standard error is 0.996205 / √12 = 0.287580 mm. Multiplying this standard error by the t critical value 2.200985 gives a margin of approximately 0.632958 mm. Adding and subtracting that margin gives 2.950375 and 4.216292 mm.
The t statistic is 3.583333 / 0.287580 = 12.460319. The two sided tail calculation with eleven degrees of freedom yields the reported p value. Retain extra digits during calculation and round only for presentation. If a calculator displays a very small probability as zero, report an appropriate scientific notation or inequality instead of claiming a probability of exactly zero.
A 95% confidence interval is produced by a procedure with a stated repeated sampling coverage under its assumptions. It is not the interval containing 95% of individual cases, and it is not automatically a 95% probability statement about the fixed population mean after observing these data. Here, the missing real sampling process is an additional reason to describe the interval as a teaching calculation. S14
What changes in a real engineering manuscript?
Replace the constructed data section with actual system and evaluation details. Define whether error is a mean absolute error, a root mean squared error, a final position error or another quantity. Specify units and the evaluation interval. Explain whether failed runs contribute a penalty, a separate failure outcome or another planned measure. Reporting only successful runs can produce an attractive but incomplete performance picture.
Document verification of the implementation separately from validation against reality. A program can calculate its equations correctly while the equations represent the physical system poorly. Report how baselines were tuned, how test cases were protected from tuning decisions, and how random seeds or repeated runs were handled. See Article 29 for that workflow. The short teaching dataset cannot substitute for those design decisions.
Adaptation 1: a natural science paper with independent units
This section uses a fictional seedling experiment to show why changing the subject also changes the unit and analysis. No plants were grown. The table contains constructed dish means, not original seedling measurements.
Define the hypothetical design before drafting
Imagine that twelve separate dishes would each receive one lighting condition, with six dishes per condition. Seedlings within a dish would share the treatment and local environment. The experimental unit for that allocation would therefore be the dish. A protocol would specify species, seed source, number of seedlings, medium, light exposure, temperature, random allocation of dishes, positional control, measurement rules and exclusions. None of those procedures is claimed to have been performed here.
Synthetic dish means for writing practice:
| Dish index within condition | Lamp A mean shoot length, mm | Lamp B mean shoot length, mm |
|---|---|---|
| 1 | 18 | 22 |
| 2 | 21 | 24 |
| 3 | 19 | 21 |
| 4 | 22 | 25 |
| 5 | 20 | 23 |
| 6 | 20 | 23 |
The rows align values to save space; they do not create pairing. Dish 1 under A is not a matched version of dish 1 under B. This distinction must be stated because the visual arrangement could otherwise invite the wrong analysis.
Model Methods and Results
For this synthetic demonstration, each value represents one dish mean at day seven. We treated the six A values and six B values as independent groups and estimated the mean difference as B minus A. We used a Welch interval to avoid requiring a common population variance. All supplied values were included. The calculation assumes independence of dishes and a suitable model for the distribution of dish means; no actual experimental allocation or environmental control was conducted.
Mean shoot length was 20.0 mm under A and 23.0 mm under B. The sample SD was 1.41 mm in each group. The estimated B minus A difference was 3.0 mm, with a conventional 95% Welch interval from 1.18 to 4.82 mm. The analysis used six independent values per condition. These results describe the constructed dish means and provide no evidence about a real lighting intervention.
The standard error is √(2/6 + 2/6) = 0.816497 mm, because each sample variance equals 2 mm². The Welch degrees of freedom happen to equal ten for these balanced values with equal sample variances. This equality is a feature of the example, not a reason to assume equal variances in every study. NIST documents the independent means calculation and Welch degrees of freedom. S15
Model Discussion and teaching note
The synthetic comparison illustrates a difference in shoot length, not a difference in overall plant health. Even in a real experiment, a longer shoot would not by itself establish greater biomass, improved survival or better crop yield. Those outcomes would require their own definitions and measurements. If all dishes under one condition shared a single chamber while the other condition used a different chamber, chamber conditions could be confounded with treatment. The number of dishes would then overstate the independent information about the lighting contrast.
The paragraph connects each limitation to a specific claim. It does not merely say “the sample was small”. To write your own version, identify exactly what the outcome measures and which features of the setup could create alternative explanations. Do not count many seedlings within the same treated dish as many independent treatment allocations. See Article 3.
Adaptation 2: a social science paper about an association
A different question needs a different interpretation. Suppose the teaching question is whether reported study time is associated with assessment score. The following eight rows are invented; no students supplied them. The exercise demonstrates an unadjusted linear association and cannot establish whether additional study causes higher scores.
| Constructed case | Weekly study hours | Assessment score out of 100 |
|---|---|---|
| 1 | 1 | 45 |
| 2 | 2 | 60 |
| 3 | 3 | 50 |
| 4 | 4 | 68 |
| 5 | 5 | 55 |
| 6 | 6 | 70 |
| 7 | 7 | 63 |
| 8 | 8 | 69 |
Model Methods and Results
We fitted an ordinary least squares regression of score on weekly study hours to eight constructed cases. The model contained an intercept and one slope. We used all rows, performed no variable selection and interpreted the slope as an unadjusted association. A conventional interval was calculated using six residual degrees of freedom. Its interpretation assumes a suitable linear mean relationship and independent errors with the required distribution and variance conditions. These assumptions were not established by a real sampling process.
The fitted slope was 2.74 score points per additional reported weekly study hour, with a conventional 95% confidence interval from 0.12 to 5.36 points per hour. The fitted intercept was 47.68 points. The model R² was 0.52, describing the fraction of sample variation around the mean represented by this fitted linear model. The intercept corresponds to zero hours, which lies outside the observed range of one to eight hours; it should not be presented as a validated prediction for students reporting no study.
Ordinary least squares estimates coefficients by minimising the sum of squared residuals. A residual is the observed score minus its fitted value. NIST explains this fitting principle and cautions about unusual observations and extrapolation. S308
Model Discussion
The positive synthetic slope illustrates an association between two recorded quantities. It does not estimate the effect of assigning an additional study hour. Prior attainment, course difficulty, access to resources and motivation could influence both study time and score. Reported study time could also differ from actual study time, and the relationship may not remain linear outside the observed range. A real study would need to define its population, recruitment strategy, measurement validity and approach to these alternative explanations before drawing broader conclusions.
The model uses eight cases only to keep the arithmetic inspectable. Eight observations do not become adequate evidence because a regression package returns a coefficient and interval. A real sample size rationale would depend on the intended precision, anticipated variation, model complexity and design. If students were clustered within schools, independence would need further consideration. If measurements were repeated within students, the data would need that structure represented. See Articles 4 and 17.
Adaptation 3: a worked narrative review
A review contributes an interpretation of existing evidence. It should answer a question such as why results differ, which mechanisms are supported, or which conditions limit an apparent benefit. Its value is not the number of papers mentioned.
The following review is a writing simulation about educational feedback. H1, H2, H3 and H4 are fictional evidence cards, not publications. They have no authors, journals or DOIs. They must not be cited as real evidence. Real review methods are discussed in Articles 26 and 27.
Begin with an evidence matrix
| Fictional card | Design and setting | Outcome and timing | Constructed evidence summary | Main interpretive constraint |
|---|---|---|---|---|
| H1 | Random allocation of introductory course students to feedback formats | Score on a similar task immediately afterwards | More specific feedback favoured immediate performance | Immediate task resembles practice |
| H2 | Random allocation in a different introductory course | Score on a different task four weeks later | Estimate was imprecise and compatible with benefit or little difference | Uncertain retention and transfer |
| H3 | Voluntary survey across several courses | Satisfaction with feedback | Students reporting detailed feedback also reported higher satisfaction | Selection, self report and confounding |
| H4 | Interviews with students from one course | Experiences of interpreting comments | Some described useful action guidance; others described information overload | Contextual interpretation, not prevalence |
The matrix separates design, outcome and limitation. It prevents the writer from treating satisfaction as achievement or treating an interview theme as a numerical effect estimate. Because no effect estimates are supplied, the model must not invent them or count “significant” studies as a substitute for synthesis.
Model review title and abstract
Title: Feedback specificity, immediate performance and transfer: a fictional narrative review for writing practice
Abstract: This teaching review examines how different outcomes affect interpretations of detailed educational feedback. Four fictional evidence cards represent immediate task performance, delayed transfer, reported satisfaction and interview accounts. We compare what each design could contribute to a review and identify claims that require additional evidence. The cards illustrate that a benefit on a similar immediate task would not, by itself, establish retention or transfer. Satisfaction and perceived usefulness address complementary questions but cannot replace a performance outcome. A coherent synthesis would therefore distinguish learning, experience and implementation, rather than combining all favourable findings into one claim of effectiveness. The evidence cards are invented and provide no substantive conclusion about educational feedback.
Model introduction and scope
The phrase “effective feedback” can refer to several outcomes: completing the current task, applying knowledge later, feeling supported or understanding what to do next. A review that treats these outcomes as interchangeable may produce a confident conclusion that no individual study actually tested. The purpose of this teaching synthesis is to show how an explicit outcome framework helps organise evidence and prevent that overstatement.
We use four fictional evidence cards to compare the contributions of experimental, survey and interview designs. No database search, eligibility screening or appraisal of real education studies was conducted. The cards were selected to illustrate differences in outcome, timing and inference. This is therefore a narrative writing exercise, not a systematic review of the education literature.
In a real review, replace that provenance statement with the actual search and selection approach. State databases, dates, major concepts, boundaries and how papers were chosen. Do not claim a comprehensive search because several familiar papers were read. SANRA, developed for narrative review appraisal in a medical publishing context, emphasises a justified aim, description of the search, appropriate referencing and scientific reasoning. It is a useful prompt for transparency, not a universal certification of every review method. S24
Model synthesis: distinguish immediate benefit from transfer
The two experimental cards address related but different learning outcomes. H1 concerns performance soon after feedback on a task resembling the practice activity. H2 concerns a different task assessed four weeks later. A favourable immediate result would support a narrower claim than durable transfer. Conversely, an imprecise delayed estimate would not establish that the immediate finding was false. The contrast may reflect timing, task demands, population differences or uncertainty. Without designs that compare these conditions directly, the review cannot identify which factor explains the contrast.
This distinction changes the structure of the review. Immediate performance and delayed transfer should be discussed separately before considering a common explanation. A useful evidence table would report the tasks, intervals, effect estimates and uncertainty for both outcomes. It would also identify whether the same participants contributed multiple reports. Treating two reports from one sample as independent studies could exaggerate the apparent breadth of evidence.
Model synthesis: integrate experience without replacing effectiveness
The survey and interview cards extend the question from whether performance changes to how feedback is experienced. H3 links reported detail with satisfaction, but voluntary participation and shared self report measurement could contribute to the association. H4 suggests possible processes: comments may help when they identify an actionable next step, while excessive information may make priorities harder to recognise. These accounts can inform hypotheses about implementation. They do not establish how common each experience is or demonstrate a causal effect on achievement.
The four cards therefore support an organising framework rather than a single numerical verdict. Outcome, timing and usability should be treated as separate dimensions. A future experiment could compare the same feedback formats using immediate and delayed tasks while collecting process information about how students interpret the comments. Such a study would test whether an apparent tradeoff between detail and usability predicts learning, rather than assuming that favourable satisfaction automatically implies better performance.
Model limitations and conclusion
This exercise deliberately simplifies the evidence. It contains no actual search, effect estimates, risk of bias assessments or appraisal of education research. It cannot recommend a feedback policy. Its purpose is to demonstrate a synthesis that distinguishes what different studies could establish and avoids substituting a count of favourable findings for an argument.
A useful review conclusion follows from the evidence framework: immediate performance, delayed transfer and satisfaction require separate interpretation. A real review could then state which conclusions are supported, how certain they are and what evidence would resolve the remaining uncertainty. The conclusion should remain narrower than the combined scope of the evidence, rather than becoming broader simply because several disciplines or methods were included.
The model paragraphs compare studies around concepts instead of presenting four disconnected summaries. They also distinguish what an experiment, survey and interview can contribute without ranking every method on one simplistic scale. Different methods answer different questions. A systematic review would additionally require a reproducible selection process and reporting appropriate to its review type; the narrative structure alone would not satisfy that requirement.
Final exercise: trace every claim back to evidence
Choose one model and identify its strongest claim. Write down the observation supporting it, the assumptions needed, and the information that would weaken it. For the controller example, the descriptive claim that every supplied B value is smaller can be checked directly in the table. A claim that B is a better real controller has no supporting experiment. For the review, a claim that the fictional cards illustrate different outcomes is supported by the matrix. A claim that detailed feedback improves real learning is unsupported.
Now draft your own title, aim, one Methods paragraph, one Results or synthesis paragraph, and one Discussion paragraph using Template 7 or Template 8. Keep a list of missing facts beside the draft. Resolve those facts from records or sources before submission. Clear writing makes uncertainty explicit; it does not fill missing evidence with plausible detail.
Sources
Файлы
- Synthetic teaching data: synthetic_controller_data.csv (1 KB)
- Calculation record: verified_calculations.json (1 KB)
- Python script that reproduces the examples: reproduce_examples.py (3 KB)
- Calculation record: worked_paper_calculations.json (1 KB)
- Calculation record: computation_environment.json (1 KB)
- About the teaching data: README.txt (8 KB)
Войдите, чтобы скачать эти файлы.