Essai 1 · Foray 265 · 1 November 2026
Audit the route from evidence to claim
A research map can be robust in one respect and fragile in another. Reconstructing it makes the difference visible.
Executed table reconstruction and sensitivity analysis · exploratory, not preregistered · Codex-assisted analysis with a separate AI verification pass
What the first attempt found
Every one- and two-study omission keeps Process highest and Task lowest in the transcribed matrix. It takes at least 11 adversarial study recodings to remove Process’s sole lead. Yet just one added mark could create a row covering all five categories. Broad emphasis and absolute absence have different robustness.
These are exact properties of a 62-row transcription and specified perturbations. They are not estimates of error probability or evidence that the original studies were coded correctly.
Why this attempt advances the foray
The continuing End is to discover which influential project-research claims deserve reliance. A supplied programme proposed scrutiny of seven papers, but mixed undated citation counts, uncertain references and assumed replication methods. Before investigating the papers, we needed to repair that starting account.
Oraee and colleagues’ review offered an accessible, concrete test. It maps research on collaboration in building information modelling (BIM). A published table connects individual studies to five antecedents: Context, Process, Task, Team and Actor. We can inspect that connection rather than treat a polished summary as the evidence itself.
The test: can another analyst recover the same category frequencies from the published table, and which conclusions survive explicit changes to that table? This is an original analysis of published coding, not a repetition of the authors’ full review.
The source → construction route
Selection counts reported by Oraee et al.; not retrieved again in this essai.
| Source contribution | Our construction | What it cannot establish |
|---|---|---|
| Final paper’s Table 1: study labels and category marks | Literal 62 × 5 binary transcription, row identifiers and page locations | Whether each original study deserves that code |
| Paper’s category summaries | Arithmetic reconciliation against visible marks | Which witness reflects the authors’ intended coding |
| Observed category distribution | All 62 single omissions, all 1,891 pairs, and minimal recoding witnesses | The likelihood of those changes or causal effectiveness in projects |
Source: Oraee, Hosseini, Papadonikolaki, Palliyaguru and Arashpour (2017), IJPM 35(7), 1288–1301. We used the publisher-formatted PDF linked from Eleni Papadonikolaki’s publication page, with its hash recorded in the dataset. The accepted manuscript differs in wording and category labels; we did not silently combine versions.
Reconcile the counts
| Category | Visible marks | Reported count | Recomputed share | Printed share |
|---|---|---|---|---|
| Context | 20 | 22 | 32% | 35% |
| Process | 42 | 41 | 68% | 66% |
| Task | 8 | 8* | 13% | 13% |
| Team | 18 | 18 | 29% | 29% |
| Actor | 13 | 14 | 21% | 23% |
Denominator 62; nearest whole percent. *Task’s count of 8 is inferred from the printed 13%, which uniquely implies 8/62 at this rounding; the other reported counts are explicit in the narrative. Marks may overlap, so percentages should not sum to 100%.
How the transcription was checked
Table 1 spans journal pages 1296–1297 (PDF pages 9–10). Text-position extraction and visual inspection were followed by a separate AI reader’s cold visual count and image-column check. That reader initially missed the Process mark in row 32; enlarged inspection corrected the miss. The retained extraction note records this repair. Both passes then agreed on all five totals. This is not independent human double coding.
Row 43 (Ochieng and Price, 2010) has no visible mark. It remains a row with five zeros, meaning “no visible mark”, not “the study discusses none of these factors”. The narrative uses some studies as examples in categories absent from their visible row. Four cell changes could match all the aggregate totals, but the totals do not uniquely tell us which cells to change.
We preserve the visible matrix and the reported summary separately. No cell was repaired to force agreement. The mismatch does not explain its own cause, establish misconduct, or show that the main finding is unreliable.
Read all 62 study rows
Loading the study rows…
| Row / study | Context | Process | Task | Team | Actor | Page |
|---|
Explore the actual matrix
How hard is it to change the leader?
Choose a competing category and a number of hypothetical study recodings. For each eligible row, the stress test removes its Process mark and adds the competing category. Other marks stay unchanged. It uses the first eligible row numbers so the witness is inspectable.
Temporary exploration: controls change only this page’s calculation. Nothing is saved, sent or written back to the paper or dataset. Reload or Reset restores the published transcription. Source downloads below always contain the unchanged baseline.
Loading the unchanged table…
Two different kinds of fragility
The leader is comparatively stable. Process has 42 visible marks and Context has 20: a gap of 22. One binary cell change closes the gap by at most one. One study recoding can close it by two when it removes Process and adds Context. There are enough eligible rows to construct a tie after 11 such recodings (31 against 31), and an overtake after 12 (30 against 32). This is a worst-case witness, not a claim that those studies should be recoded.
Complete coverage is much more fragile. No row in the visible table contains all five marks, but row 47 contains four. Adding its missing Context mark would create one all-five row. Even that would establish only five-category coverage, not a study of the relationships among all five. The final paper’s description of all-inclusive work as rare must not be inflated into a universal absence claim.
Omission is a different test from recoding. Removing any one or two rows never reverses the most- or least-frequent category in this matrix. That protects a limited ordering statement against small exclusions. It says nothing about a different search, new papers since 2016, a different conceptual framework or systematic miscoding.
Repair the starting programme
The seven papers remain candidate investigations. They are not a verified “top seven”: the original mixed citation counts lacked a consistent dated export, eligible universe and reproducible query. Here is what the primary-source check supports, and how far each check went.
Cui et al. (2018)
Review of studies on public–private partnerships for infrastructure projects · IJPM 36(5), 773–794. DOI 10.1016/j.ijproman.2018.03.004.
Publisher highlights report 754 articles and six topics. The seed’s assumed 300–500 corpus is not a basis for replication.
Publisher metadata/highlights checked; complete analysis and underlying corpus not reproduced.
Locatelli et al. (2017; online 2016)
Corruption in public projects and megaprojects: There is an elephant in the room! · IJPM 35, 252–268. DOI 10.1016/j.ijproman.2016.09.010.
Keep online and issue dates distinct. A comparative case study needs an explicit causal argument; adding a few projects is not enough to identify a corruption effect.
Open publisher copy located; no independent case replication.
Banihashemi et al. (2017)
Critical success factors for integration of sustainability into construction project management practices in developing countries · IJPM 35(6), 1103–1119. DOI 10.1016/j.ijproman.2017.01.014.
The abstract specifies 16 interviews, 101 completed questionnaires and PLS-SEM. Those facts replace the seed’s vague sample and proposed EFA/regression design.
Institutional record and abstract checked; questionnaire/data availability remains unresolved.
Oraee et al. (2017)
Collaboration in BIM-based construction networks: A bibliometric-qualitative literature review · IJPM 35(7), 1288–1301. DOI 10.1016/j.ijproman.2017.07.001.
A mixed-method review with a documented Scopus search and a 1,031 → 271 → 62 selection route. The first essai reconstructs its final published coding table.
Final author-hosted publisher PDF read and table inspected; original retrieval and qualitative coding not rerun.
Aarseth et al. (2017; online 2016)
Project sustainability strategies: A systematic literature review · IJPM 35(6), 1071–1083. DOI 10.1016/j.ijproman.2016.11.006.
The abstract gives five journals, pre-2016 research and 68 selected articles. The seed’s approximate 300-paper corpus would test a different design.
Publisher abstract and institutional metadata checked; complete coding not reconstructed.
Wiewiora et al. (2014)
Interactions between organizational culture, trustworthiness, and mechanisms for inter-project knowledge sharing · PMJ 45(2), 48–65. DOI 10.1002/pmj.21407.
The publisher identifies a multiple-case study. Qualitative credibility and the chain from interviews to interpretation need their own test, not a generic missing-code score.
Identity and study type checked against PMI; interview data not independently analysed.
Locatelli et al. (2023)
A Manifesto for project management research · European Management Review 20(1), 3–17. DOI 10.1111/emre.12568.
A conceptual agenda with no new empirical dataset. Evaluate its argument and programme rather than treating missing raw data as a defect.
Publisher source and data statement checked; conceptual merits not adjudicated here.
Editorial thresholds, access and practical programme design
Missing data, a failed replication and evidence of misconduct are different states. COPE’s August 2025 retraction guidance concerns seriously unreliable findings and assigns the editorial decision to editors. It does not supply an automatic 10% rule or make inaccessible information sufficient grounds for retraction.
The supplied calendar and budgets are proposals, not evidence that preregistration, staffing or ethics approval has happened. A real survey replication needs a model-specific power or precision justification. Citation ranking needs a dated source and sensitivity analysis. Those programme components remain unexecuted.
Repeat and inspect the attempt
The data, source/version metadata and calculation are deliberately small. Download the two inputs (CSV and dataset JSON) and script into one folder, then run python3 analyse.py. The script uses Python’s standard library and reproduces results.json. It does not fetch sources or require a paid database.
Attribution: original study labels and visible codes from Oraee et al.; transcription, sensitivity construction and this explanation produced with Codex on 1 November 2026. Full papers are linked, not rehosted. The original programme remains a private input. Public downloads are research artefacts, not user-created records.
What can be relied on now?
The exact visible-table frequencies are reproducible from the retained data. The mismatch with three summary categories is independently checked within this agent-assisted workflow. Process’s leading frequency survives every tested one- and two-row omission and the specified small recoding budgets. These are bounded, useful positive results.
We have not independently recovered the historical Scopus export, reproduced bibliometric networks, recoded the 62 full texts, verified every inclusion decision, tested the framework’s completeness or evaluated its recommendations on projects. No disciplinary misconduct or retraction conclusion follows.
The next discriminating essai would freeze a coding rubric and independently recode the original studies, initially examining the rows where narrative examples and visible marks diverge. Its purpose would be to distinguish a presentation inconsistency from a substantive coding disagreement, then test the whole 62-study matrix if feasible. A separate impact-ranking study can choose which other claims most deserve that effort. Neither route replaces the wider programme.