Archival status: the Zenodo DOI below is reserved, but the archival record is not yet published because checkpoint-volume transfers remain incomplete. The source and manuscript are publicly available in GitHub release v1.0.1; Zenodo checkpoint downloads are not yet available. The release checksum files describe the immutable v1.0.1 tag and its source ZIP, not this subsequent status note on the default branch.

Study repository: PeterPonyu/mixture-regularisation-gating. Versioned archive: 10.5281/zenodo.22914503. Use CITATION.cff for release 1.0.1; release-manifest.json records distributed-file checksums. The archive and source repository describe this study only.

Author-owned software is MIT licensed. The author's manuscript, figures, generated results and model weights are CC BY 4.0. Source-study data, labels, annotations and other third-party materials retain their original terms; they are not relicensed. Read LICENSE and NOTICE.md before reusing mixed-content files.

This research package studies how a capped mixture loss and its annealing schedule control realised gradient exposure in an attention autoencoder. The saved experiment compares a zero-weight reference with cap-5 and uncapped losses under 20- and 100-epoch ramps. Two fixed data backgrounds, Setty and Endo, and three seeds produce six shared warmups and 30 matched endpoints. Saved continuation histories cover epochs 200–219 and the epoch-220 endpoint. Historical capped and uncapped sweeps on four backgrounds supply additional context.

The reproducible product is analysis of those saved computations: per-batch exposure checks, paired endpoint contrasts, robustness analyses, figures and the article build. All commands below run on a CPU with no training, GPU, raw dataset or model checkpoint. Large .pt files are distributed separately in this article's Zenodo checkpoint archive; the default GitHub analyses and figures do not read them.

For operations that consume model states, download this article's checkpoint archive from its Zenodo DOI record and restore the package-relative paths listed under optional_checkpoints in publication-manifest.json. A GitHub source checkout alone does not supply those model states, and these checkpoints do not constitute a complete training-replay environment.

Use Python 3.11 or newer with NumPy and SciPy from requirements-analysis.txt. The recorded validation used Python 3.13.5, NumPy 2.2.6 and SciPy 1.16.3. From the package root:

python3 -m venv .venv
. .venv/bin/activate
python -m pip install -r requirements-analysis.txt
export OMP_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2 MKL_NUM_THREADS=2
python scripts/exposure_diagnostic.py
python scripts/robustness_audit.py
python scripts/verify_publication.py

exposure_diagnostic.py uses only the Python standard library. It matches all 30 original and continuation endpoints, verifies all 8,700 saved per-batch records, and recomputes summed weighted latent-gradient norms. It writes evidence/exposure_diagnostic.json, including within-warmup contrasts, descriptive correlations and analytic sensitivity of the clipping gates to a simultaneous rescaling of latent coordinates and Gaussian parameters.

robustness_audit.py also reads the 30 continuation .npz archives. It checks finite latents, responsibility normalization, shared batch-history identifiers, zero exposure in capped arms and equality of capped latents with the zero-weight reference. It reports per-background and per-schedule associations, normalization by update count, removal of individual seeds or warmups, and the original mixture-fit convergence flags. It creates validation/ when needed and writes validation/robustness-audit.json.

verify_publication.py uses only the Python standard library and checks publication-manifest.json. The manifest distinguishes original source digests from current distributed-file digests, identifies source snapshots, and records the semantic relabeling of machine-specific paths. Checkpoints are optional for this check; scientific arrays and required analysis inputs are verified even when every .pt is absent.

Figure regeneration requires R with ggplot2, gridExtra, jsonlite, the standard grid package, and Cairo graphics support. Tested versions are R 4.3.3, ggplot2 4.0.3, gridExtra 2.3 and jsonlite 2.0.0. Install missing R packages in an existing R environment with:

Rscript -e 'install.packages(c("ggplot2", "gridExtra", "jsonlite"), repos="https://cloud.r-project.org")'

Arial and the system librsvg, Cairo and GObject shared libraries must be installed. The vector renderer uses Python's standard library; no Python PyGObject or Pycairo package is needed. python3 must be on PATH; PYTHON can select another executable. A TeX installation with XeLaTeX and BibTeX, the packages named in standalone-assets/preamble.tex, and the TeX Gyre Termes Math font is required for the article.

Rscript scripts/build_figures.R
bash build.sh

The first command rebuilds seven data-figure PDF/PNG pairs and the vector architecture PDF, and writes figures/plot_manifest.json with panel counts. All numerical inputs are local JSON and CSV files listed in figure-build.json. Non-finite values in historical score JSON are interpreted as missing values during plotting without editing the archived scores. The six UMAP panels reuse the saved 450-point coordinates and labels; they do not fit UMAP.

build.sh regenerates the figures, runs XeLaTeX, BibTeX and two further XeLaTeX passes, and copies build_independent/paper.pdf to paper.pdf. It overwrites generated figures and the article PDF. Pre-rendered figures are also included. Neither figure generation nor compilation changes model parameters or produces new training outcomes.

• evidence/factorial_results.json contains the 30 epoch-200 endpoints, shared-warmup checks, original mixture parameters and convergence flags.
• evidence/run-archive/ contains the 30 continuation endpoint records and arrays, six permutation schedules, gradient histories and mixture refits. evidence/new_results.json is the same aggregate used by the figures.
• evidence/new_protocol.json records the continuation design, and evidence/EXPOSURE_METHODS.md explains the exposure estimand and limits of the analytic scaling calculation.
• evidence/source/dpmm_transformer.py is an unmodified model snapshot whose SHA-256 matches the original experiment. It is provided for inspection of the loss and gating implementation; it is not a standalone importable training library. The nonportable historical training drivers are excluded. Their original hashes remain in the result records and publication manifest.

The available arrays support the stated saved-result audits. They do not provide a self-contained raw-data preprocessing or training pipeline, a reconstruction of all epoch-200 checkpoint states, or regeneration of the archived graph scores. Checkpoint files alone do not remove these limits.

The exposure estimand is the sum of effective weight times the norm of the clipped mixture-loss gradient with respect to the latent representation. It is not parameter displacement. Per-batch gradient magnitudes were not archived for epochs 180–199 and cannot be recovered from active-batch counts. Continuation refit convergence flags were not recorded; finite values do not prove convergence. The original fit records include 46 non-converged refits and are retained.

Six warmups and two fixed backgrounds are the computational units of the paired analysis; arms, batches and cells are not independent donors. Exposure is a post-treatment quantity and depends on latent scale and fitting. Its association with an endpoint is descriptive, not an identified causal dose–response. The analytic rescaling calculation shifts the clipping gates without retraining and makes no prediction of a new trained endpoint.

Release 1.0.1 changes only checkpoint distribution and documentation; scientific results, model-state bytes and the manuscript are unchanged from 1.0.0. To make large downloads recoverable, the study-specific Zenodo record stores the checkpoint ZIP as ordered 20 MiB volumes. checkpoint-downloads.json lists each volume URL, byte count and SHA-256, plus the reconstructed ZIP identity. These volumes belong only to this study and must not be mixed with other releases.

From the repository root, use the Python standard-library helper:

python3 scripts/restore_checkpoints.py --download-dir ../checkpoint-downloads --download
unzip -n reproduced/checkpoints.zip -d .

If the volumes were downloaded separately, omit --download. The helper validates every volume and the reconstructed ZIP and refuses to overwrite different existing files. It does not deserialize models or execute training. After extraction, the original package-relative .pt paths and hashes are unchanged. Default CPU saved-result analyses do not need these downloads.