EMA paper and reference research

This manuscript revision uses the completed r07 study. It adds a critical literature map, bounded reference acquisition and a compiled paper with portable sources. The original science, attempts, bundles, report and audits remain unchanged. No new training or native model session was run.

Read the paper, download its source archive, or start with the reference index. The paper’s author label allagma was explicitly supplied by the owner. Other researchers configure their own attribution; it is not a framework default.

What the reference research changed

The eight-entry map covers seven primary papers and the exact upstream code snapshot. Each entry separates verified bibliographic metadata from located source support, limitations and consequences for the study. The search record and critical map identify competing findings and untested baselines, with individual notes available on demand.

The search found Ajroldi et al. (2025), which directly studies averaging and annealing. The manuscript therefore makes a bounded case-study claim rather than presenting their qualitative relationship as new. Switch EMA supplies actual counterexamples to universal ordinary-EMA benefit. Its intervention and Morales-Brotons et al.’s bootstrapping experiment are kept distinct. The search is a targeted narrative map, not an exhaustive systematic review, and was conducted after the frozen experiment.

The asset manifest pins seven PDFs and their available TeX sources, three relevant upstream code files at commit 1de1dbc1f4ee2c5f61e9c94348d55eb51d7fa2eb, one real saved model artifact and one held-out/generation split from r07. The retrieval manifest records 19 available assets, their hashes, sizes, relative cache paths and licenses. Total transfers were 139,596,456 bytes; expanded source files account for another 86,420,758 bytes. These raw files remain outside Git, releases and Pages. Paper PDFs retained under arXiv’s distribution license are not relicensed for redistribution by this project.

Recompute and rebuild without the cache

From the repository root, using Python 3.11+:

python3 studies/ema-schedule/publications/reference-r2/derive.py
python3 -m allagma reference check \
  --directory studies/ema-schedule/publications/reference-r2/references --require-review
python3 -m allagma paper check --study studies/ema-schedule/publications/reference-r2
python3 -m allagma paper build --study studies/ema-schedule/publications/reference-r2 \
  --destination work/my-paper-rebuild
python3 tools/verify_paper_output.py --directory work/my-paper-rebuild --author allagma

The paper build additionally needs an existing TeX distribution and OS isolation (macOS Seatbelt or Linux bubblewrap). The final verification command uses Poppler’s pdfinfo and pdftotext. These dependencies belong to the optional paper integration; the default conformance kit and toy study remain offline and standard-library-only. No reference download, model account, scientific training environment or cache is needed for these commands.

The source archive also works independently of the repository:

mkdir paper-source
tar -xzf paper-source.tar.gz -C paper-source
python3 paper-source/anc/build.py --output clean-paper-build

Use new output directories. derive.py recalculates all 114 retained summaries from the selected seed-level measurements. It checks means, uncertainty, sign-flip tests and leave-one-seed-out means, without repeating training or extracting metrics from all saved samples. Source selection binds each copied scientific file to the original package index. The complete study’s separate restoration instructions remain the path to its full scientific package.

Inspect or reacquire the references

For this machine’s existing cache, a new CLI process can verify and reuse every asset without network access:

python3 -m allagma reference verify-cache \
  --directory studies/ema-schedule/publications/reference-r2/references
python3 -m allagma reference acquire \
  --directory studies/ema-schedule/publications/reference-r2/references

For a new cache, --online explicitly permits retrieval of the public paper and code assets. The model and dataset use acquisition: provided: restore the r07 package if needed, then provide a local JSON mapping their asset IDs to the exact package_path files in that package. Keep this local binding outside Git. The provenance URL points to the package index, not a direct download of those individual arrays. A missing supplied asset produces an explicit unavailable state and a nonzero exit, while successful downloads remain reusable.

Cache defaults are 256 MiB per asset, 512 MiB expanded, 1 GiB per study, 4 GiB per cache and 20 GiB minimum free disk. Transfers, retries, failed extraction, quarantine and protected input copies are accounted for. The initial TLS failures from the Mac’s missing Python CA link remain in the acquisition ledger; the adapter now uses the existing system trust bundle without disabling verification. The explicit gates for terms and adopted-limit expansion were exercised by offline conformance cases; no gated terms or larger limits were accepted here.

Retained validation

Integration evidence records offline reuse with an unchanged transfer count, immutable preparation and the selected assets. The actual broker probe read 36 model arrays and two evaluation arrays, with network, protected-input writes and shared-cache reads denied. An earlier preparation is retained; a new preparation revision clarifies the evaluation-file description without changing its bytes.

Scientific review and humanizer review are bound to the manuscript, section roles, citations, attribution and selected evidence. They are same-assistant reviews, not independent peer review. The delivery receipt binds the PDF and source archive to two clean builds, including invocation of the unpacked archive’s own builder. The PDF inspection records the page review. Local TeX Live 2024 results are identified separately from the supported-release CI check. No arXiv processing or acceptance is claimed, and no submission was made.

The PR review repairs changed packaging and review identity. The manuscript and PDF bytes remain unchanged. The original scientific review, prose review and delivery remain as historical evidence. The current archive isolates reading notes from generated ancillary files. Repair evidence records all four findings and their regressions.

---

Canonical source · Generated from the maintained repository document at this build’s commit.