# Essai 1: what the visible evidence can—and cannot—support

**Run:** 1 November 2026. **Foray:** 265, reliable project research. **Object:** the final, publisher-formatted Oraee et al. (2017) article, *Collaboration in BIM-based construction networks: A bibliometric-qualitative literature review*, IJPM 35(7), 1288–1301. [DOI](https://doi.org/10.1016/j.ijproman.2017.07.001) · [author's publication list](https://elenip.com/publication/).

This is a completed computational reconstruction and sensitivity analysis of one published table, accompanied by an audit of the proposed research programme. It is not a replication of the review's retrieval, screening, network analysis, or interpretation of the 62 underlying papers. The distinction is the result's useful boundary, not an unfinished claim of replication.

## Question and construction

Can a reader recover the paper's evidence-to-claim route from the material actually available, and does the main descriptive contrast depend on a small number of table entries?

The article separates a broad bibliometric analysis from a qualitative review. Its Collaboration Pentagon provides five categories: Context, Process, Task, Team and Actor. Table 1 assigns visible marks to 62 named studies. We converted that table into a 62-by-5 binary matrix, retained the article's reported summaries as a separate witness, and recomputed both counts and percentages. A zero records a blank printed cell; it is not an independent judgement that the underlying study lacked that dimension.

The source concepts and the new construction remain separate. The five categories and original assignments belong to Oraee and colleagues. Our contributions are the machine-readable transcription, reconciliation of visible marks with printed totals, exhaustive omission checks, and explicit adversarial recoding thresholds. Nothing in these computations validates the authors' theoretical framework or the appropriateness of individual assignments.

## Result: three aggregate disagreements, one stable contrast

| Category | Visible table marks | Recomputed share of 62 | Article's reported count | Article's reported share |
|---|---:|---:|---:|---:|
| Context | 20 | 32.26% | 22 | 35% |
| Process | 42 | 67.74% | 41 | 66% |
| Task | 8 | 12.90% | 8* | 13% |
| Team | 18 | 29.03% | 18 | 29% |
| Actor | 13 | 20.97% | 14 | 23% |

*The Task count of 8 is uniquely implied by 13% of 62 under nearest-integer percentage rounding; the other reported counts are explicit in sections 4.2.1, 4.2.2, 4.2.4 and 4.2.5. The matrix has 101 visible marks, because a study can receive several categories. Percentages must not be forced to total 100%.*

Context, Process and Actor do not reconcile between the visible table and the paper's aggregate narrative/summary. Four individual cell changes would be necessary to match all reported marginal counts: add two Context marks, remove one Process mark and add one Actor mark. The summaries do not identify which cells should change. We therefore make none of those repairs.

Two especially inspectable cases are row 43, Ochieng and Price (2010), which has no visible mark, and row 21, Gu and London (2010), which has a Process mark but no Actor mark even though section 4.2.5 names that study as an Actor example. These are requests for reconciliation, not findings about author intent. A production error, version difference, transcription issue in the article, or analytical difference would require further evidence to distinguish.

The main descriptive contrast survives either account: Process is most frequent and Task least frequent. In the visible table Process exceeds Context, its nearest competitor, by 22 studies; in the reported summary the lead is 19. The discrepancy changes exact arithmetic but does not overturn that ordering.

## Sensitivity: robust ordering, fragile absolute coverage count

The script enumerated every one-row omission (62 cases) and every two-row omission (1,891 cases). Across all 1,953 cases, Process remained uniquely most frequent and Task uniquely least frequent. After two omissions the Process lead is at least 20 marks. Dropping the blank row changes the denominator to 61 and leaves all counts unchanged.

Omission is not recoding. We therefore also computed the smallest adversarial changes that could erase Process's lead at fixed denominator. One binary cell change alters the margin by at most one. Two changes in a row—remove Process and add Context—alter the margin by two. It takes 22 individual toggles, or 11 such row substitutions, to tie Process and Context. Twelve row substitutions would make Context higher. The script preserves explicit row witnesses for those constructed alternatives. These are worst-case stress tests; they are not measured error rates or predictions of independent coders' disagreements.

The table contains no row with all five marks. That fact is much more fragile: row 47 already has four dimensions, so a single added Context mark would create an all-five row. The final article's abstract describes all-inclusive studies as rare; the earlier accepted manuscript uses stronger absence wording. We do not promote the stronger historical wording into a claim about the final paper. Nor would five co-occurring marks prove that an underlying study tested interactions among all five dimensions.

## What changed in the research programme?

The supplied programme described Oraee as purely bibliometric and its search strategy as undocumented. The actual paper has a qualitative phase, a named and tabulated 62-study corpus, and a documented search method. That material makes a real, bounded reconstruction possible. It also exposes the difference between access to a table and access to the original coding decisions.

The useful next evidential step is a reconciliation of the table against the original coding worksheet, followed by independent, blinded recoding of a declared sample—or the full corpus—under a fixed rubric. A contemporary Scopus rerun would be a separate database-drift exercise, not the reproduction of a frozen November 2016 export. A new database, different keywords, new software settings or revised framework would need an explicit change record. None of those later studies has been claimed as completed here.

The result supports a specific conclusion: this historical table contains recoverable evidence and a robust frequency ordering, but its exact aggregates need reconciliation. It does not support misconduct allegations, a retraction recommendation, a claim that present-day BIM research has the same gaps, or a claim that the least-covered dimension is the most useful intervention for practitioners.

## Evidence and verification

The source PDF came from the author's publication list and is identified by SHA-256 in `table-1.json`. Table 1 is on printed pages 1296–1297, PDF pages 9–10. All 310 cells were matched between a manual transcription and PDF-position extraction, after visual inspection of both rendered pages. A second AI analyst performed a cold visual tally: its first Process count missed the mark in row 32; an enlarged crop and independent pixel check resolved the disagreement in favour of 42. This retained extraction correction is a useful warning against treating an attractive audit as infallible. Human double coding has not occurred.

Ten automated checks pass, including exact source row identities, preserved blanks, CSV/JSON agreement, aggregate disagreements, exhaustive omission results, constructed recoding witnesses, the one-cell all-five counterexample, invalid exclusions, and deterministic regeneration. The final and accepted versions are not interchangeable: they have different labels, row ordering, some marks, percentages and abstract wording. The final version governs this analysis.

Run `python3 analyse.py` in this folder to regenerate `results.json`, and `python3 -m unittest discover -s . -p 'test_*.py' -v` in a checkout to run the tests. No paid database, external package or internet connection is needed for these computations. Rechecking transcription still requires the identified source PDF; the PDF itself is linked, not redistributed in this public payload.