EARLIER HANDS-ON CHECK RECORD, OCTOBER 1, 2026. Practical Picks adopted a research-only method on October 3, 2026. This check is kept as a dated historical observation, not current testing; the comparison was re-evaluated from public sources on October 19, 2026. Versions named below were current on October 1, 2026. marimo fixes security issues only in its latest release; use current versions for any replay. PP-260 — local Python notebooks on a changed Parquet file Private release-candidate method, checked 2026-10-01 UTC. Article publication requires owner approval. Scope Linux x86-64, Python 3.12. JupyterLab 4.6.4 / Jupyter Server 2.21.1 and marimo 0.25.0 were installed in separate task-scoped venvs. JupyterLab 4.6.3 was the initial test and was superseded because 4.6.4, available September 21, patched four advisories affecting 4.6.3. Both also used pandas 2.3.3, PyArrow 25.0.1 and Matplotlib 3.10.7. No account or subscription. The browser UI used Playwright Chromium. Servers bound explicitly to 127.0.0.1, were checked with ss -ltnp and stopped. Jupyter authentication was disabled only during synthetic loopback automation; do not copy that setting into normal use. Bundle: /assets/pp260/replay-bundle.zip Files: fixture.csv, fixture-updated.csv, make-fixture.py, fixture.parquet, fixture-updated.parquet, marimo-analysis.py, jupyter-analysis.ipynb, marimo-observed-updated.html, jupyter-observed-updated-4.6.4.ipynb, and initial historical jupyter-observed-updated.ipynb. The source notebooks expect active.parquet in their own directory; copy one frozen fixture to that filename before a run. Run from a trusted local directory with a current Python environment. Frozen contract Parquet schema: order_id, region, channel as strings; amount as nullable int64. Eight fictional rows; one null amount. Exclude null amounts, group sum by region, sort region. Original: East 24, North 7, West 35, grand total 66; seven nonmissing rows. The C-order West amount changes 20 to 25 in the updated fixture. Updated: East 24, North 7, West 40, grand total 71. The marimo slider defaults to zero for this core result. Raising the in-notebook minimum to 10 on the updated data yields three rows, East 12, West 35, grand total 47. Relevant SHA-256 values fixture.parquet: 8c00f917733a1672ed2510a0111d1401a1a70f032c88af72c76f895642404ab7 fixture-updated.parquet: 5e77400e407a3c00d4df208e5c520430aadbd1bdaa52c2c896417bf497f77161 marimo-analysis.py: 92b1d3e5376ec5b632298e6e0538482ccd592634176a1bb56ba48fa5f9020b7f marimo-observed-updated.html: 07a5ef848f1d4b217bc4d38bbca5f6e804a4164cb2198152e4040405012d8eb0 jupyter-observed-updated-4.6.4.ipynb: 5f5e9c9e103787a1150579ec92e9d35203d347e646459cb7a3247bbfcf4029ee jupyter-observed-updated.ipynb (initial 4.6.3): fc0b6ab900786d1085a23d0837a892fb84a92ea1f5d0dde1ad6f127c3844eff1 Source jupyter-analysis.ipynb gained stable cell IDs after the observed browser run; its code cells did not change. Check its current bundle hash with sha256sum. Replay in a fresh task-specific directory python3.12 -m venv marimo-venv marimo-venv/bin/python -m pip install --no-cache-dir --only-binary=:all: marimo==0.25.0 pandas==2.3.3 pyarrow==25.0.1 matplotlib==3.10.7 python3.12 -m venv jupyter-venv jupyter-venv/bin/python -m pip install --no-cache-dir --only-binary=:all: jupyterlab==4.6.4 pandas==2.3.3 pyarrow==25.0.1 matplotlib==3.10.7 cp fixture.parquet active.parquet marimo-venv/bin/marimo edit marimo-analysis.py --host 127.0.0.1 --port 4493 In a browser, Run All and inspect rows, missing amount, group sums and chart. Stop the server. Replace active.parquet with fixture-updated.parquet. Reopen and Run All. Move the slider from 0 to 10 to check dependency recomputation. Stop the server. cp fixture-updated.parquet active.parquet marimo-venv/bin/marimo export html marimo-analysis.py -o replay-marimo.html For Jupyter, copy the original fixture to active.parquet, then launch: jupyter-venv/bin/jupyter lab --ip=127.0.0.1 --port=4494 --no-browser Open jupyter-analysis.ipynb in the authenticated local browser, Run All and save. Stop Jupyter, replace active.parquet with fixture-updated.parquet, reopen. Notice saved old output before execution; Run All (prefer Restart Kernel and Run All), inspect 71/40 and chart, then Save. Stop Jupyter. Do not disable the default Jupyter token for routine replay. Observed checks and limits Both browser Run All attempts on the original and updated fixtures completed and displayed expected totals and a bar chart. JupyterLab 4.6.4 Ctrl+S persisted .ipynb outputs; its reopened old 66 total was visible before execution, disappeared after Run All, and updated 71/40 was saved. The 4.6.4 saved output has execution counts 1/2/3 and output counts 1/1/2. The older 4.6.3 artifact is retained only to document the superseded initial run. marimo Save wrote a separate edited copy; the canonical .py holds no stale output. Clean-process marimo export exited 0 with 71/40 and embedded one chart image. Changing the marimo slider recalculated dependent totals to 47 without Run All. Neither product automatically detected the external file replacement in the tested setup. We did not test external watchers, scale, crashes, other operating systems, privacy/network telemetry, remote hosting or third-party extensions. Outputs depend on the separately installed Python packages and source file. This is not a performance or security ranking.