Reproducibility artifacts for the meta-research study "Available but not runnable out of the box: execution prerequisites and label reporting in shared dysphagia and swallowing AI code." The protocol keeps the working title it was written under. We audited whether published swallowing/dysphagia-AI code is not only available but re-runnable, and we release the full measurement pipeline so the audit is itself reproducible ("we practice what we audit"). Contents: multi-source open-API search & enrichment scripts (01–06), repository discovery / intake / code-link mining / census synthesis (07–10), a blind screening-reliability re-coding (11), post-hoc robustness checks (12–18: search-truncation measurement, backward-citation coverage, reference verification, and full author-list verification against Crossref), and the measurement scripts behind the environment pinning/portability audit, the run-instruction audit and every reported proportion (19–21), whose machine-readable outputs are archived under results/; the transparency rubric, codebook, and RS1–RS6 clinical reference-standard taxonomy; the containerized re-execution harness (Dockerfile templates) with per-repository build/run logs and verdicts; the dated study protocol, a self-audit checklist applying this study's own minimum-reporting recommendation set to itself, the PRISMA-ScR checklist (transparently declined, with its rationale), and that recommendation set; record-level bibliographic corpus metadata; and the candidate-repository inventory at three stages (raw discovery, code-link-mining additions, and the vetted inventory carrying the inclusion decision and reason for every candidate) together with the objective per-repository intake table behind each verdict. Non-distributive: only artifacts produced by the authors are included. Audited third-party repositories—their code, trained weights, and any human-derived (e.g. VFSS/FEES) data—are not re-hosted; each is referenced to its original repository/DOI. Dual license: code (scripts/, harness) under MIT (LICENSE-CODE); data & text under CC-BY-4.0 (LICENSE-DATA). Changes in v1.1.8. Records the re-judged target-validity code for the row whose premise an earlier correction had changed. The target is the instrumental standard itself, because the raters read the videofluoroscopy, and the label is also marked as judged from the same signal the model takes as input, because the input parameters are measured from the recordings those raters scored. That mark would fit almost every imaging row if it were read literally, since a segmentation mask is drawn on the image being segmented and a manometric annotation lies on the trace the model consumes. Those are measurement tasks and the mark would tell a reader nothing their own descriptions do not, so it is applied only where the endpoint is a diagnostic claim, where the accuracy reported is agreement with the raters who read the same recording rather than accuracy against an independently obtained standard. The taxonomy document states that criterion, so the rows without the mark are distinguished by a stated rule rather than by silence. One row carries it. No reported number changes. Changes in v1.1.7. Corrects one label granularity field and reopens one target-validity judgment. The August correction rewrote the scale field of one row but not the binarization field beside it, which still described the segmentation step alone, so the count of releases that reduce the clinical label to two classes moves from four of seven to five of seven. The same correction had flagged the target-validity code of that row for re-judgment, because its premise had changed, and the closure that followed recorded acceptance of the corrected reading rather than a new code, so the flag was cleared while the judgment it named was never made. The row now states that the item is open and why, the released coding table carries the flag beside the code, and the half of the code that the source contradicts, that the segmentation mask is the label, is removed. Which category follows is a judgment and is left to the clinician author. Two records that disagreed with the coding are brought into line: the taxonomy document listed one study among those predicting a proxy although its own code reads target valid, and section D of the codebook described a schema that was never released, which is kept with a note saying what it is rather than rewritten. The row coded from an accepted manuscript names the handle it was read from. The coding generator now rejects a misaligned file, after a comma written into an unquoted field split a row and shifted every column after it without anything noticing. No other number changes. Changes in v1.1.6. Counts the release whose codes come from an accepted manuscript. One audited release has no open published version, and its codes were assigned from the version its authors deposited in an institutional repository, which the deposit records as the version accepted for publication. The row was held outside the clinical denominator while it already carried a confirmed absence for label reliability, which the study's own rule does not allow. It is now counted, its spectrum cell is filled from the same source, and the denominator moves from six to seven with the reliability count from 0 of 5 to 0 of 6, since one release has no rater. Script 21 now reads that denominator from the coding file instead of printing it as a constant. The recommendation set still carried the sample data count from before the rule was corrected in August, and the July intake log presented its figures as current and now says what superseded them. No other number changes. Changes in v1.1.5. A corrective release. Script 31 derived the point shifts of the screening sensitivity from figures that had already been rounded, which moved two of them by 0.1 point, and reported the largest increase only over the signals with a nonzero count, leaving out the larger shift a zero row can take. Both bounds are now computed from the exact fractions and rounded away from zero, and the increase over every signal is recorded beside the earlier figure. In the adapted reporting checklist, the item on unadjusted estimates pointed only at the footnote that explains the row given without an interval, not at the table body or the results subsection that carry the proportions, and the funding item named only the title page heading while the statement also stands under Declarations in the manuscript. The archive documents name the article's sections as the article now titles them. No other number changes. Changes in v1.1.4. Adds what a second pre-submission review of the manuscript asked for and corrects what it found. Five scripts and their outputs: the journal-policy mapping (30) now reads every availability section of each article, separates data from code statements and records where a repository is named, after its first version read only the first section, stopped at the first nested subsection and skipped one open full text, which misstated four of seven rows; strata, joint signal patterns and corrected denominators (31: imaging and non-imaging, publication-linked, the 0/6 conditional environment denominator, the 1/18 relaxed composite, channel recall of 0 of 2 in-scope comparator records, and screening disagreements); a re-issue of the code-link mining extraction without its de-duplication step (32), because the original script dropped links already in the inventory before recording them, so the reported absence of overlap between channels could not be read from the record; an as-declared install of the four environment files, in three releases, that were never built, followed by pip check and an import of every declared distribution with the network off (33: 3 of 4 repositories installed, each with a clean pip check and loadable imports, and Video-SwinUNet failed to install); and a comparison of the predictions of the one shipped model that loaded in July across the library versions recorded in its file and those its unpinned install resolved to (34: identical predictions on all 2,002 inputs under the versions recorded in the model file and under those the install resolved to, while scikit-learn moved forward without CatBoost cannot predict at all). Both need Docker, record base images by digest and are not run by the reproduce driver; neither reassigns a verdict. The measurement definitions now state that the harness entry condition was written after the three builds, and correct a count made before the sample-data correction; the reference-standard taxonomy carries a revision note; one build log records that its access date and commit cannot be fixed to one state; the author-list check treats a group author as closing the byline; and the reproduce driver runs the new offline script. Wording that implied a rule stated in advance is corrected in the protocol, the measurement definitions, the self-audit checklist and script comments, and earlier descriptions of the pre-submission review as external review are corrected. The README lists scripts 31 to 34 and the Docker checks and gains version-history entries for v1.1.3 and v1.1.4. Every number reported before this release is unchanged except the policy mapping, which is corrected. Changes in v1.1.3. Adds the provenance the earlier releases lacked, in response to a pre-submission review. Six scripts and their outputs: the commit each audited repository pointed at on its access date, recovered for 22 of 23 rows, the exception being a repository that is empty and therefore has no commit to record; the depositor's published checksum for the external weight archive, which fixes which file was downloaded without establishing bit-identity with the copy we ran and deleted; a re-reading of every intake signal at those commits, which reproduced all six signals in all 20 repositories the intake script had coded; the screening sensitivity in the direction the disagreements p
Sefa Tunçer, Nazife KAPAN· Zenodo (CERN European Organi...· 0 citations
Reproducibility artifacts for the meta-research study "Available but not runnable out of the box: an audit of execution prerequisites and label reporting in shared code for dysphagia and swallowing artificial intelligence." The protocol keeps the working title it was written under. We audited whether published swallowing/dysphagia-AI code is not only available but re-runnable, and we release the full measurement pipeline so the audit is itself reproducible ("we practice what we audit"). Contents: multi-source open-API search & enrichment scripts (01–06), repository discovery / intake / code-link mining / census synthesis (07–10), a blind screening-reliability re-coding (11), post-hoc robustness checks (12–18: search-truncation measurement, backward-citation coverage, reference verification, and full author-list verification against Crossref), and the measurement scripts behind the environment pinning/portability audit, the run-instruction audit and every reported proportion (19–21), whose machine-readable outputs are archived under results/; the transparency rubric, codebook, and RS1–RS6 clinical reference-standard taxonomy; the containerized re-execution harness (Dockerfile templates) with per-repository build/run logs and verdicts; the dated study protocol, a self-audit checklist applying this study's own minimum-reporting recommendation set to itself, the PRISMA-ScR checklist (transparently declined, with its rationale), and that recommendation set; record-level bibliographic corpus metadata; and the candidate-repository inventory at three stages (raw discovery, code-link-mining additions, and the vetted inventory carrying the inclusion decision and reason for every candidate) together with the objective per-repository intake table behind each verdict. Non-distributive: only artifacts produced by the authors are included. Audited third-party repositories—their code, trained weights, and any human-derived (e.g. VFSS/FEES) data—are not re-hosted; each is referenced to its original repository/DOI. Dual license: code (scripts/, harness) under MIT (LICENSE-CODE); data & text under CC-BY-4.0 (LICENSE-DATA). Changes in v1.1.6. Counts the release whose codes come from an accepted manuscript. One audited release has no open published version, and its codes were assigned from the version its authors deposited in an institutional repository, which the deposit records as the version accepted for publication. The row was held outside the clinical denominator while it already carried a confirmed absence for label reliability, which the study's own rule does not allow. It is now counted, its spectrum cell is filled from the same source, and the denominator moves from six to seven with the reliability count from 0 of 5 to 0 of 6, since one release has no rater. Script 21 now reads that denominator from the coding file instead of printing it as a constant. The recommendation set still carried the sample data count from before the rule was corrected in August, and the July intake log presented its figures as current and now says what superseded them. No other number changes. Changes in v1.1.5. A corrective release. Script 31 derived the point shifts of the screening sensitivity from figures that had already been rounded, which moved two of them by 0.1 point, and reported the largest increase only over the signals with a nonzero count, leaving out the larger shift a zero row can take. Both bounds are now computed from the exact fractions and rounded away from zero, and the increase over every signal is recorded beside the earlier figure. In the adapted reporting checklist, the item on unadjusted estimates pointed only at the footnote that explains the row given without an interval, not at the table body or the results subsection that carry the proportions, and the funding item named only the title page heading while the statement also stands under Declarations in the manuscript. The archive documents name the article's sections as the article now titles them. No other number changes. Changes in v1.1.4. Adds what a second pre-submission review of the manuscript asked for and corrects what it found. Five scripts and their outputs: the journal-policy mapping (30) now reads every availability section of each article, separates data from code statements and records where a repository is named, after its first version read only the first section, stopped at the first nested subsection and skipped one open full text, which misstated four of seven rows; strata, joint signal patterns and corrected denominators (31: imaging and non-imaging, publication-linked, the 0/6 conditional environment denominator, the 1/18 relaxed composite, channel recall of 0 of 2 in-scope comparator records, and screening disagreements); a re-issue of the code-link mining extraction without its de-duplication step (32), because the original script dropped links already in the inventory before recording them, so the reported absence of overlap between channels could not be read from the record; an as-declared install of the four environment files, in three releases, that were never built, followed by pip check and an import of every declared distribution with the network off (33: 3 of 4 repositories installed, each with a clean pip check and loadable imports, and Video-SwinUNet failed to install); and a comparison of the predictions of the one shipped model that loaded in July across the library versions recorded in its file and those its unpinned install resolved to (34: identical predictions on all 2,002 inputs under the versions recorded in the model file and under those the install resolved to, while scikit-learn moved forward without CatBoost cannot predict at all). Both need Docker, record base images by digest and are not run by the reproduce driver; neither reassigns a verdict. The measurement definitions now state that the harness entry condition was written after the three builds, and correct a count made before the sample-data correction; the reference-standard taxonomy carries a revision note; one build log records that its access date and commit cannot be fixed to one state; the author-list check treats a group author as closing the byline; and the reproduce driver runs the new offline script. Wording that implied a rule stated in advance is corrected in the protocol, the measurement definitions, the self-audit checklist and script comments, and earlier descriptions of the pre-submission review as external review are corrected. The README lists scripts 31 to 34 and the Docker checks and gains version-history entries for v1.1.3 and v1.1.4. Every number reported before this release is unchanged except the policy mapping, which is corrected. Changes in v1.1.3. Adds the provenance the earlier releases lacked, in response to a pre-submission review. Six scripts and their outputs: the commit each audited repository pointed at on its access date, recovered for 22 of 23 rows, the exception being a repository that is empty and therefore has no commit to record; the depositor's published checksum for the external weight archive, which fixes which file was downloaded without establishing bit-identity with the copy we ran and deleted; a re-reading of every intake signal at those commits, which reproduced all six signals in all 20 repositories the intake script had coded; the screening sensitivity in the direction the disagreements point rather than away from it; the publication-confirmed subset with the two signals previously omitted from it; and a mapping of journal sharing policy to printed availability statement to observed artifact for the seven studies where a policy applies. The self-audit checklist records the immutable-state item as met retrospectively and adds an item for the measurement error of the audit itself. Two inputs resolved to the authors' working copies rather than the English sources this archive publishes; re-running the census after the fix left every reported number unchanged. Changes in v1.1.2. A corrective release; the analysis, the data and every reported number are unchanged. The protocol's remaining-steps line still listed notifying the authors of the audited repositories as a step, which contradicted the ethics paragraph of the same document, where it is stated that they are not notified. No notification round was carried out, and the manuscript and the self-audit checklist both record that item as not met. The line is corrected and the earlier wording is quoted in place rather than deleted. The citation file, which still declared v1.1.0 and had never recorded the version DOIs minted for v1.1.0 and v1.1.1, is brought up to date. Changes in v1.1.1. A corrective release; the analysis, the data and every reported number are unchanged. Four files described a personal relationship between the two authors, which is not an interest in the audited subject matter and is no longer stated; the methodological facts that were doing the work remain in place, that the primary measurements are script-derived and that the exploratory clinical axis was coded by one clinician with no second coder. The paper's title is corrected in every file that repeats it. The README version history records the version DOIs. Changes in v1.1.0. This version adds the two artifacts the manuscript cites as archived but which v1.0.0 omitted — the dated protocol document and the self-audit checklist; replaces the census-synthesis and screening-reliability scripts with the versions that produce the figures reported in the manuscript; adds the post-hoc robustness checks (12–18); adds three measurement scripts (19–21) that put a released command behind the pinned-and-portable environment row, the run-instruction signal and every reported proportion; adds a bootstrap interval and the prevalence and bias indices to the screening-reliability output; and archives the machine-readable script outputs under results/. Every document in the archive is in English. v1.0.0 remains a
Sefa Tunçer, Nazife KAPAN· Zenodo (CERN European Organi...· 0 citations
Reproducibility artifacts for the meta-research study "Available but not runnable out of the box: an audit of execution prerequisites and label reporting in shared code for dysphagia and swallowing artificial intelligence." The protocol keeps the working title it was written under. We audited whether published swallowing/dysphagia-AI code is not only available but re-runnable, and we release the full measurement pipeline so the audit is itself reproducible ("we practice what we audit"). Contents: multi-source open-API search & enrichment scripts (01–06), repository discovery / intake / code-link mining / census synthesis (07–10), a blind screening-reliability re-coding (11), post-hoc robustness checks (12–18: search-truncation measurement, backward-citation coverage, reference verification, and full author-list verification against Crossref), and the measurement scripts behind the environment pinning/portability audit, the run-instruction audit and every reported proportion (19–21), whose machine-readable outputs are archived under results/; the transparency rubric, codebook, and RS1–RS6 clinical reference-standard taxonomy; the containerized re-execution harness (Dockerfile templates) with per-repository build/run logs and verdicts; the dated study protocol, a self-audit checklist applying this study's own minimum-reporting recommendation set to itself, the PRISMA-ScR checklist (transparently declined, with its rationale), and that recommendation set; record-level bibliographic corpus metadata; and the candidate-repository inventory at three stages (raw discovery, code-link-mining additions, and the vetted inventory carrying the inclusion decision and reason for every candidate) together with the objective per-repository intake table behind each verdict. Non-distributive: only artifacts produced by the authors are included. Audited third-party repositories—their code, trained weights, and any human-derived (e.g. VFSS/FEES) data—are not re-hosted; each is referenced to its original repository/DOI. Dual license: code (scripts/, harness) under MIT (LICENSE-CODE); data & text under CC-BY-4.0 (LICENSE-DATA). Changes in v1.1.6. Counts the release whose codes come from an accepted manuscript. One audited release has no open published version, and its codes were assigned from the version its authors deposited in an institutional repository, which the deposit records as the version accepted for publication. The row was held outside the clinical denominator while it already carried a confirmed absence for label reliability, which the study's own rule does not allow. It is now counted, its spectrum cell is filled from the same source, and the denominator moves from six to seven with the reliability count from 0 of 5 to 0 of 6, since one release has no rater. Script 21 now reads that denominator from the coding file instead of printing it as a constant. The recommendation set still carried the sample data count from before the rule was corrected in August, and the July intake log presented its figures as current and now says what superseded them. No other number changes. Changes in v1.1.5. A corrective release. Script 31 derived the point shifts of the screening sensitivity from figures that had already been rounded, which moved two of them by 0.1 point, and reported the largest increase only over the signals with a nonzero count, leaving out the larger shift a zero row can take. Both bounds are now computed from the exact fractions and rounded away from zero, and the increase over every signal is recorded beside the earlier figure. In the adapted reporting checklist, the item on unadjusted estimates pointed only at the footnote that explains the row given without an interval, not at the table body or the results subsection that carry the proportions, and the funding item named only the title page heading while the statement also stands under Declarations in the manuscript. The archive documents name the article's sections as the article now titles them. No other number changes. Changes in v1.1.4. Adds what a second pre-submission review of the manuscript asked for and corrects what it found. Five scripts and their outputs: the journal-policy mapping (30) now reads every availability section of each article, separates data from code statements and records where a repository is named, after its first version read only the first section, stopped at the first nested subsection and skipped one open full text, which misstated four of seven rows; strata, joint signal patterns and corrected denominators (31: imaging and non-imaging, publication-linked, the 0/6 conditional environment denominator, the 1/18 relaxed composite, channel recall of 0 of 2 in-scope comparator records, and screening disagreements); a re-issue of the code-link mining extraction without its de-duplication step (32), because the original script dropped links already in the inventory before recording them, so the reported absence of overlap between channels could not be read from the record; an as-declared install of the four environment files, in three releases, that were never built, followed by pip check and an import of every declared distribution with the network off (33: 3 of 4 repositories installed, each with a clean pip check and loadable imports, and Video-SwinUNet failed to install); and a comparison of the predictions of the one shipped model that loaded in July across the library versions recorded in its file and those its unpinned install resolved to (34: identical predictions on all 2,002 inputs under the versions recorded in the model file and under those the install resolved to, while scikit-learn moved forward without CatBoost cannot predict at all). Both need Docker, record base images by digest and are not run by the reproduce driver; neither reassigns a verdict. The measurement definitions now state that the harness entry condition was written after the three builds, and correct a count made before the sample-data correction; the reference-standard taxonomy carries a revision note; one build log records that its access date and commit cannot be fixed to one state; the author-list check treats a group author as closing the byline; and the reproduce driver runs the new offline script. Wording that implied a rule stated in advance is corrected in the protocol, the measurement definitions, the self-audit checklist and script comments, and earlier descriptions of the pre-submission review as external review are corrected. The README lists scripts 31 to 34 and the Docker checks and gains version-history entries for v1.1.3 and v1.1.4. Every number reported before this release is unchanged except the policy mapping, which is corrected. Changes in v1.1.3. Adds the provenance the earlier releases lacked, in response to a pre-submission review. Six scripts and their outputs: the commit each audited repository pointed at on its access date, recovered for 22 of 23 rows, the exception being a repository that is empty and therefore has no commit to record; the depositor's published checksum for the external weight archive, which fixes which file was downloaded without establishing bit-identity with the copy we ran and deleted; a re-reading of every intake signal at those commits, which reproduced all six signals in all 20 repositories the intake script had coded; the screening sensitivity in the direction the disagreements point rather than away from it; the publication-confirmed subset with the two signals previously omitted from it; and a mapping of journal sharing policy to printed availability statement to observed artifact for the seven studies where a policy applies. The self-audit checklist records the immutable-state item as met retrospectively and adds an item for the measurement error of the audit itself. Two inputs resolved to the authors' working copies rather than the English sources this archive publishes; re-running the census after the fix left every reported number unchanged. Changes in v1.1.2. A corrective release; the analysis, the data and every reported number are unchanged. The protocol's remaining-steps line still listed notifying the authors of the audited repositories as a step, which contradicted the ethics paragraph of the same document, where it is stated that they are not notified. No notification round was carried out, and the manuscript and the self-audit checklist both record that item as not met. The line is corrected and the earlier wording is quoted in place rather than deleted. The citation file, which still declared v1.1.0 and had never recorded the version DOIs minted for v1.1.0 and v1.1.1, is brought up to date. Changes in v1.1.1. A corrective release; the analysis, the data and every reported number are unchanged. Four files described a personal relationship between the two authors, which is not an interest in the audited subject matter and is no longer stated; the methodological facts that were doing the work remain in place, that the primary measurements are script-derived and that the exploratory clinical axis was coded by one clinician with no second coder. The paper's title is corrected in every file that repeats it. The README version history records the version DOIs. Changes in v1.1.0. This version adds the two artifacts the manuscript cites as archived but which v1.0.0 omitted — the dated protocol document and the self-audit checklist; replaces the census-synthesis and screening-reliability scripts with the versions that produce the figures reported in the manuscript; adds the post-hoc robustness checks (12–18); adds three measurement scripts (19–21) that put a released command behind the pinned-and-portable environment row, the run-instruction signal and every reported proportion; adds a bootstrap interval and the prevalence and bias indices to the screening-reliability output; and archives the machine-readable script outputs under results/. Every document in the archive is in English. v1.0.0 remains a
