lzszq/dsh-scholar/domain-data-science

Domain Pack: Data Science

Applies on top of the research-core skill for data-science projects.

Data conventions

1. Every dataset must be versioned with a content hash (preprocessing hash in

the ExperimentContract). A change in preprocessing invalidates prior runs.

1. Record the data license and provenance in the corpus/snapshot metadata;

private data is flagged and never uploaded to external models (external_model_upload policy).

1. Splits: record the exact split rule (stratified, temporal, grouped by

entity). Leakage checks are mandatory before formal runs.

Statistical conventions

1. Report effect sizes with bootstrap CIs (the analysis pipeline does this);

never report a bare p-value without an interval.

1. Multiple-testing correction (Holm default) when more than one metric or

subgroup is compared.

1. Class-imbalanced targets: macro-averaged metrics + class counts.
2. Uncertainty sections must state: seeds, n, CI method, hardware.

Baseline & comparison

1. Reproduce the baseline first (baseline_prepare + baseline_verify); only

within-tolerance reproductions may be compared against.

1. Record deviations verbatim in the RunManifest; never adjust official

baseline numbers.

Pipeline reproducibility

1. Every transform step must be deterministic (fixed seeds, no

time-dependent sorting) or have its hash recorded.

1. The release bundle must include: data manifest + hashes, preprocessing

script, environment lockfile, and the exact analysis commands.