czx-lbj/deepseek-science/deepseek-science-analysis

CSV analysis recipe

Use scripts/analyze_csv.py for one numeric measurement grouped by one categorical column. Python 3 and matplotlib are prerequisites. The helper is runtime-tested on a synthetic fixture; that check does not validate a study design or establish scientific expert quality.

1. Inspect the actual CSV and study description. Record the measurement units, sampling unit, missing-value policy and group meanings. This helper treats every row as an independent observation. Confirm that assumption before using its intervals: repeated measurements, subjects appearing in several groups, technical replicates and clustered samples need a design-specific analysis. Keep identifiers in the original data so this can be checked.
2. Resolve the helper relative to this skill directory, and check the selected interpreter without installing anything:

``sh python scripts/analyze_csv.py --check-dependencies ``

If the check fails, report needs-setup, the interpreter and the missing dependency. Use an available configured interpreter or the user's approved setup procedure.

1. Run on actual data into a new output directory. Supply meaningful column names and an explicit seed:

``sh python scripts/analyze_csv.py --input measurements.csv --group condition --value measurement --output analysis-run --seed 2026 ``

UTF-8 CSV requires unique nonempty headers, consistent row widths, nonempty groups and finite numeric values. Invalid rows cause a nonzero exit before an output directory is created; the helper never silently drops them. It refuses an existing output directory. Resolve any failure against the original data and record corrections.

1. Inspect summary.csv, observations.csv, figure.svg, figure.png, metadata.json, the copied input and helper. Summary fields are row count N, arithmetic mean, sample SD (ddof=1) and a 95% percentile bootstrap interval for each group mean. The helper resamples rows within each group 2,000 times by default with the supplied seed, then linearly interpolates the 2.5th and 97.5th percentiles. N=1 has an unavailable SD and interval. Small samples, biased sampling and dependence limit interval interpretation; intervals alone do not establish a group difference, statistical significance or causation. See references/method.md before interpreting results.
2. Write an analysis note stating the sampling unit, exclusions or corrections, descriptive observations and limitations. Keep exploratory choices distinct from prespecified hypotheses. When using science_project, checkpoint the original input CSV as well as the actual tables, figures, copied input/script, metadata and note in its analysis phase; quote observed rows or metadata text with exact locators. Recording the original input makes a later edit invalidate the recorded phase; retaining a copied input preserves the exact run data. Advance using the current revision only after the checkpoint succeeds.

For a smoke run, replace the input with examples/synthetic.csv, group with group and value with value. Those rows are invented demonstration data, not experimental evidence. Completion requires a successful execution and inspection of saved outputs; a dependency check alone is not an analysis run.