dataset-report

Dataset Report

Deliver data.csv, plot.png, dataset-report.md and clio-dataset-report.json in a new output directory. This bounded workflow supports a numeric table with a selected value column: source and CSV at most 64 MiB, CSV at most 100,000 rows. Stop and explain when those limits do not fit; do not silently sample and report whole-data statistics.

Discover the HDF5, Pandas and Plot MCP tools. Inspect HDF5 metadata, dataset shape and supplied column names/units before reading. Ask for the dataset path or column meaning when ambiguous; do not infer physical units from values.

Before data operations, run the helper beside this skill (resolve its absolute path from this installed skill directory):

python3 /absolute/path/to/this/skill/scripts/verify_report.py prepare --source /absolute/input.h5 --output /absolute/new-report --column signal_mV

It records the source hash and planned output paths. Preserve this baseline on retries. Hashing reads the source file; it is intentionally limited to 64 MiB.

1. Open the HDF5 source read-only. Inspect the chosen dataset and export its numeric values with export_dataset as JSON. Close the file afterward.
2. Map the exported matrix to the metadata's named columns. Use Pandas save_data with data={"time_s": [...], "signal_mV": [...]} and index=False to write data.csv. A columns/data envelope loses names. Compare saved rows and column names against the exported values; this is the provenance check the helper cannot perform.
3. Call Pandas statistical_summary on the selected CSV column. Record unrounded count, mean, median (the returned 50%), min and max in the manifest's statistics object, preserving its other fields.
4. Use Plot line_plot for an actual ordered axis or histogram_plot for a distribution; follow the live tool schema. Save plot.png and inspect it.
5. Write dataset-report.md: source/dataset, columns and units, row count, the five statistics, figure, data handoff and limitations. Counts have no units; mean/median/min/max retain the measurement units. State observed patterns separately from models. CV/skewness alone cannot establish growth; no fit, mechanism or extrapolation is established by this workflow. Derive any reported interval or rate from the actual ordered time/value pairs and show the calculation; otherwise omit it.
6. Run the helper's check /absolute/new-report/clio-dataset-report.json. Fix concrete failures without changing the original baseline. The native package's client hook runs the same read-only check after native write/edit of the manifest or report. Shell/MCP writes do not trigger this hook, so the explicit final check remains required. A hook message is feedback, not a prevention mechanism or a scientific certification.
7. Give the raw data, report and helper output to the installed scientific-evidence-reviewer. Claude's native marketplace names it clio-agents:scientific-evidence-reviewer; project installations use its unqualified name. In OpenCode, request that named subagent. In Codex, use the named agent when discovered, or run clio-kit plugin run-agent scientific-evidence-reviewer --client codex --project /absolute/project --prompt "Review these observed results: ...". Supply actual evidence in the prompt; do not invent a review response. If discovery, sandbox access or the provider blocks review, report independent review as unavailable. Treat reviewer corrections as claims requiring evidence, not authority.

Keep the final response to verified project-relative output paths, the five checked statistics, check/review status and limitations. Do not introduce numerical claims absent from the reviewed report. A PASS establishes CSV numeric checks, source integrity and output presence; it does not establish HDF5→CSV fidelity, plot semantics, the truth of prose or a fitted scientific model.