Archival status (14 September): GitHub v1.2.0 is published and its asset checksum is verified. Zenodo DOI 10.5281/zenodo.22743916 is reserved, but upload to that draft is incomplete; the new Zenodo record is not yet published. Until it is published, obtain and cite the versioned GitHub release.
The current closed-loop audit study is available in studies/closed_loop_audit, with its own reproduction entry points, frozen informative-observation experiment and independent Julia/R checks. The repository-root workflow and the historical version 1.0.0 citation below describe the earlier release. They remain preserved for reproducibility. Version 1.2.0 identifies the current September study; its corresponding Zenodo DOI is https://doi.org/10.5281/zenodo.22743916. Manuscript and submission documents remain outside this repository.
The September 8 extension adds 2,200 baseline/reliability fits, complete numerical certification of the original regression and dynamic conditions, and conditional-mixture diagnostics. Results include a diminishing baseline bias and nonmonotone parameter-specific sensitivity; the smaller independent batch did not reproduce the original previous-failure contrast sign. Complete results and portable checks are linked in the study README.
Muxue Zhou, Zhen Xiao, and Xiaokai Xia, School of Educational Science, Hengyang Normal University. Xiaokai Xia is the corresponding author. See CONTRIBUTORS.md for the confirmed contributions.
Accompanying manuscript: Interpreting adaptive learning logs: An audit of inferential targets and observation mechanisms. No journal publication DOI is claimed. Manuscript authorship is separate from the historical records below.
Zhou, M., Xiao, Z., & Xia, X. (2026). Closed-loop adaptive-learning audit:
Julia and R reproducibility materials (Version 1.2.0) [Computer software].
https://doi.org/10.5281/zenodo.22743916
GitHub's citation panel now describes v1.2.0. Both CITATION.cff and
.zenodo.json have matching creators, title and version. See
RELEASE_NOTES.md for the archived-version boundaries.
The numerical evidence remains frozen at computational commit
52eddb7a7e72c665f0353e278bdb5e5cd96b651b; the metadata update does not
represent a new simulation run.
Public R and Julia reproducibility materials for:
Auditing psychological inferences about memory from adaptive learning logs: Cognitive-model recovery, policy logging, and randomized validation
Authors: Muxue Zhou and Xiaokai Xia, School of Educational Science, Hengyang Normal University.
This repository supports reader inspection and computational reproduction of the article's simulations, empirical models, sensitivity analyses, figures, tables, and validation checks. The manuscript, supplementary document, submission correspondence, reviewer response, and document templates are not distributed here.
The project evaluates when delay effects can be interpreted in adaptive, closed-loop learning systems. Julia implements the known-truth simulations and model-recovery experiments. R analyzes the de-identified empirical transition table, summarizes simulations, builds figures and tables, and runs the machine-readable validation suite.
Included materials:
- R and Julia analysis source code;
- one minimized de-identified empirical analysis table;
- formal simulation and empirical result files;
- generated figures and tables;
- validation records and runtime information.
Excluded materials:
- the article and all submission documents;
- raw Anki exports and event-level logs;
- names, emails, telephone numbers, device identifiers, free text, source filenames, original event identifiers, and linkage keys;
- absolute dates, absolute timestamps, and exact within-day clock times;
- n-back, Stroop, FSRS-state, and participant-summary data that are outside the article's analyses.
The public empirical workflow begins at
data/analysis_transitions_deidentified.csv.
It contains 2,045 cross-day transitions, 10 released fields, 12 pseudonymous
participant codes, and 100 pseudonymous card codes. The codes were created
during private de-identification; their crosswalk is absent from the repository
and release package. No card content is included.
The private event-to-transition preparation was audited before release, but raw
and event-level records are withheld to reduce longitudinal re-identification
risk. Consequently, readers can reproduce the statistical analyses that begin
with the released transition table, while the private de-identification and
event reconstruction step is outside the public computational boundary. See
data/DATA_DICTIONARY.md and
data/PUBLIC_DATA_AUDIT.md.
The formal results were produced with:
- Julia 1.11.7;
Project.tomlsupports Julia 1.10 or later and uses no external Julia packages; - R 4.5.3;
- R package
lme42.0.6 for two crossed participant/card random-intercept logistic models; base and recommended R packages are used elsewhere.
The tested dependency versions are also recorded in
R_DEPENDENCIES.csv.
Run commands from the repository root. Rscript is required for every mode;
julia is additionally required for full and smoke modes.
Rscript run_all.R --mode=referenceThis mode starts from the included formal replication-level outputs. It regenerates summaries, five main and three supplementary figures, seven main and eight supplementary tables, runtime manifests, and the 38-check validation report. It does not rerun the longest Monte Carlo studies, empirical bootstrap, or mixed-effects models.
Rscript run_all.R --mode=fullThis mode reruns all Julia simulations and both R empirical analyses before rebuilding and validating the generated assets. The empirical sensitivity grid contains:
4 planned-interval thresholds x 2 outcome definitions x 3 penalties
= 24 cells x 1,200 participant-cluster bootstraps
= 28,800 estimates
The full run also fits 24 leave-one-participant-out specifications and two crossed random-intercept benchmarks. Runtime depends on available CPU resources.
Rscript run_all.R --mode=smokeSmoke mode runs one replication per Julia design cell and writes only to
validation/smoke_outputs/. It checks execution and output schemas without
overwriting the formal results.
| Path | Contents |
|---|---|
src/julia/ |
Closed-loop, cognitive-state, calibration, WCLS, and context-proxy simulations |
src/R/ |
Empirical analyses, summaries, builders, public-data checks, and global validation |
data/ |
Minimized de-identified transition data, dictionary, and release audit |
outputs/ |
Formal simulation and empirical outputs plus run manifests |
figures/ |
Generated PNG and SVG figures |
tables/ |
Full-precision CSV and display-rounded Markdown tables |
validation/ |
Public-data, model-extension, empirical, runtime, and 38-check records |
run_all.R |
Main reproduction entry point |
Project.toml |
Julia project metadata and compatibility declaration |
The project-level gate is implemented in
src/R/validate_outputs.R. The released reference
results pass all 38 checks. The checks cover formal row counts, known-truth
conclusions, WCLS/LPM numerical equivalence, context-proxy behavior,
observation-model boundaries, calibration robustness, the minimized empirical
schema, sensitivity results, generated asset inventories, and the R/Julia-only
source boundary.
Scientific reproduction is assessed against the declared designs, row counts, estimands, direction, bias, coverage, recovery criteria, and validation checks. Floating-point and platform differences can prevent byte-identical Monte Carlo files across systems.
| Item | Link |
|---|---|
| Repository | https://github.com/dddd1007/adaptive-memory-audit |
| GitHub release | https://github.com/dddd1007/adaptive-memory-audit/releases/tag/v1.0.0 |
| Version DOI (v1.0.0) | https://doi.org/10.5281/zenodo.21850352 |
| Concept DOI (all versions) | https://doi.org/10.5281/zenodo.21850351 |
| Zenodo record | https://zenodo.org/records/21850352 |
The original citation metadata is preserved in
metadata/history/v1.0.0/CITATION.cff.
Use this historical citation only for the corresponding earlier materials.
Suggested historical software citation:
Zhou, M., & Xia, X. (2026). Adaptive-memory audit: R and Julia reproducibility
materials (Version 1.0.0) [Computer software].
https://doi.org/10.5281/zenodo.21850352
Prefer the version DOI (10.5281/zenodo.21850352) when citing this exact
release. Use the concept DOI (10.5281/zenodo.21850351) only when you want
to cite the software as a whole across versions. The current September study uses the v1.2.0 citation above.
| Material | License |
|---|---|
run_all.R, Project.toml, and src/ |
MIT License |
data/analysis_transitions_deidentified.csv |
Creative Commons Attribution 4.0 International |
outputs/, figures/, tables/, validation records, and Markdown documentation |
Creative Commons Attribution 4.0 International |
See LICENSE, DATA_LICENSE.md, and
LICENSES/CC-BY-4.0.txt. The software and materials
are provided without warranty. Users remain responsible for ethical and lawful
handling of human-participant-derived data.
The independent learning-reconstruction study adds 3,200 frozen-design simulations of update-family identification, state reconstruction and forward prediction under random/adaptive selection. Its validation record and complete replication outputs accompany independent Julia/R checks. Correct family selection can coexist with substantial update-parameter error under response-model misspecification. These conditional simulation findings do not identify causal effects or validate psychological parameters in the empirical case. Prior modules and release tags are preserved.