ValidReuse Method Explorer

Streamlit app for comparing q10(LRV) evaluation methods on the same editable integer input table:

• Empirical bootstrap
• Bayesian approximation
• Bayesian (PyMC)

Use the local Conda environment

conda activate validreuse

Install dependencies

pip install -r requirements.txt

Recommended:

• Use Python 3.13 for the full app, including PyMC.
• The older validreuse/ Python 3.14 environment is still present, but PyMC is not reliable there.

Run the app

streamlit run app.py

Deploy With Docker

Build the image:

docker build -t validreuse-app .

Run it locally:

docker run --rm -p 8501:8501 validreuse-app

Then open http://localhost:8501.

Deploy On A Server

1. Copy the project to the server.
2. Install Docker.
3. Build the image:

docker build -t validreuse-app .

4. Run the container in the background:

docker run -d --name validreuse -p 8501:8501 --restart unless-stopped validreuse-app

5. Put Nginx or Caddy in front of port 8501 if you want HTTPS and a public domain.

Streamlit Cloud

This app can also be deployed from a GitHub repo on Streamlit Community Cloud with app.py as the entrypoint.

Note:

• PyMC makes deployment heavier than a simple Streamlit app.
• If you want the lightest deployment, we can make the PyMC method optional or disable it in cloud builds.

Quality-managed assessments

The QM-Bewertungen workspace adds persistent assessments, original-report retention, authenticated author/reviewer roles, documented suitability checks, versioned paired-LRV decision protocols, controlled exclusions, immutable revision history, independent release and linked amendments. Exports include a printable HTML report, originals and the audit/calculation archive.

See QM_WORKFLOW.md for personal account setup, the first milestone's scope, backup instructions and validation requirements before routine official use. Local accounts must be provisioned before signing into this workspace; no default passwords are provided.

Documentation workspace

The sidebar workspace Dokumentation und Formeln documents every quantity the app reports, with its formula: the three volume methods and the constraints on them, exposure and unit conversion, Eq. 5-1/5-2 and the paired guideline LRV, the model likelihood and priors, the closed-form and numerical population quantiles, the diagnostic thresholds and the status logic, plus the legacy methods and a symbol table. Each section names the function that implements it, and the volume section includes a live calculator that calls sample_data.effective_volume directly, so the worked example cannot drift from the code.

Counts and individual sample volumes

The default workspace now accepts assay-level effluent counts and sample-specific volumes. Select an indicator, enter rows or import the provided Excel template, and inspect the effective-volume preview before fitting. Data for each indicator are kept separately in the session. Download CSV/JSON to retain them outside the session. The old calculator is available from the sidebar as Bisheriger Methodenvergleich.

assets/ValidReuse_Datenvorlage.xlsx contains an Anleitung sheet with the full workflow, a column glossary and all three volume methods. Its documented Zulauf and Ablauf sheets can be imported together. Keep their technical column names unchanged so that the app can validate the workbook safely.

For effluent, enter raw integer CFU/PFU counts (including zero), a sample ID and a replicate ID. For influent, enter positive concentrations in /100 mL, /mL or /L; decimal values are accepted. Result type missing is retained in export and excluded from fitting. Censored, MPN and normalised effluent results can be recorded but block the count fit until a suitable raw-count dataset is supplied.

Volume modes, all in mL:

• concentrated: filtered × assayed / dilution factor / total eluate × recovery.
• direct: assayed / dilution factor × recovery.
• effective: already corrected original-water volume, used once. Other volume fields are ignored in this mode; document the source in recovery_source.

Recovery is a fraction in (0,1], not percent. Dilution factor 10 denotes 1:10. Technical replicates must use nonoverlapping aliquots and the same sample ID; each has its own exposure but shares a latent concentration. Environmental samples must have distinct IDs. A sample location and time should be recorded for traceability.

Both PyMC candidates model count ~ Poisson(effective_mL/100 * concentration). Poisson-lognormal uses lognormal sample concentrations; the negative-binomial comparator uses gamma sample concentrations. The output is the population LRV quantile for independent latent influent/effluent concentrations, with a one-sided posterior credible lower bound. This is distinct from the legacy count-predictive LRV. Exact zero counts need no pseudocount in the likelihood. The preview separately shows the guideline's count-scale zero convention. When matching pair IDs are supplied, the app also reports the paired, measurement-limited LRV from Eq. 5-2. Technical effluent replicates are pooled by adding their counts and disjoint effective volumes. Its displayed q10 is descriptive and has no confidence interval; it is kept separate from the model-based latent population q10. The paired table can be downloaded as CSV and is included in the result JSON when a model run is exported.

Unrounded convergence diagnostics, lower-bound Monte Carlo error, marginal posterior predictive checks at the observed volumes, and candidate spread are reported. JSON export includes inputs, exposures, settings, posterior quantiles and package versions. Input or setting changes hide stale results. Fixed priors are documented in the interface and count_models.py. Results are experimental: paired dependence, uncertain recovery, censored likelihoods and full calibration are not yet implemented. See ROBUSTNESS_REVIEW_NOTES.md for the updated review.

Run focused tests in the Python 3.13 app environment:

python -m unittest discover -s tests -v