Version Dataset Agent Skill

Version Dataset Skill

You help a medical researcher put a dataset under version control: fingerprint it,

detect when it changes, and lock a reproducible version. This guards the

data-integrity rule — an analysis must run on the data it claims to, with a fixed

seed — by making any drift between runs loud instead of silent.

Communication Rules

• Communicate with the user in their preferred language.
• Manifest fields, drift reports, and provenance notes are in English.

Philosophy

A dataset is an input to a result; if it changes silently, every downstream

number is suspect. This skill records a deterministic fingerprint (file SHA-256 +,

for tabular files, schema and per-column value hashes) so a later run can *prove*

the inputs are unchanged. It does not alter data, and it records nothing

non-deterministic (no timestamps unless explicitly passed), so the same data

always yields the same manifest.

Reference Files

• Manifest schema + workflow: ${CLAUDE_SKILL_DIR}/references/manifest_schema.md —

the manifest.json structure, what each drift category means, and the non-

deterministic-artifact policy (PPTX/DOCX timestamps). Read before interpreting drift.

Deterministic Script

# Build a manifest (record the analysis seed + provenance)
python "${CLAUDE_SKILL_DIR}/scripts/version_dataset.py" manifest data.csv \
  --out manifest.json --seed 42 --provenance "KNHANES 2018 extract v1"

# Verify a later copy against it (CI / pre-analysis gate)
python "${CLAUDE_SKILL_DIR}/scripts/version_dataset.py" verify --manifest manifest.json --strict

# Compare two manifests (what changed between versions)
python "${CLAUDE_SKILL_DIR}/scripts/version_dataset.py" diff --old v1.json --new v2.json

File hashing is stdlib-only; tabular schema/column hashing uses pandas when present.

--ignore-cols excludes volatile columns; --base makes manifest keys relative.

Workflow

Step 1: Lock the version (gate)

Build the manifest at the moment the dataset is frozen for analysis. Gate:

confirm with the user the seed and provenance note are correct before locking —

the manifest is the record they will cite as "this is the data the results came from."

Step 2: Verify before each run (gate)

Before re-running an analysis (or in CI), verify --strict. Gate: if drift is

reported, stop and show the user the drift report; do not proceed on changed data

without their explicit acknowledgement and a re-lock. Silent re-run on drifted data

is the failure this skill exists to prevent.

Step 3: Diff across versions

When a dataset is intentionally updated, diff the old and new manifests and

present the change set (added/removed/changed columns, row-count delta) so the

user can record what changed and re-lock. Gate: the user approves the new

version before it replaces the locked one.

Non-Deterministic Artifacts

Some outputs (PPTX/DOCX with embedded timestamps, figures with render metadata)

change byte-for-byte on every build even when the analysis is identical. Do not

put these under strict byte verification — manifest only the deterministic inputs

and tabular outputs (data files, result CSVs), or use --ignore-cols for volatile

columns. See references for the policy.

Scope Limitations

Supported

• Content-hash manifest of any file; schema + per-column hashes for tabular files

(CSV/TSV/Parquet/Stata/SAS/Excel).

• Drift verification and manifest-to-manifest diff.

NOT Supported

• Storing or transmitting the data itself (manifests hold hashes, not contents).
• Cleaning, profiling, or de-identifying — use /clean-data, /generate-codebook, /deidentify.
• Full pipeline-output reproducibility for non-deterministic binaries (see above).

Cross-Skill Integration

• /generate-codebook documents *what* is in the data; version-dataset locks *which version*.
• /deidentify should run before a manifest is shared (example values are not stored, but provenance notes may carry context).
• Demo reproducibility: each bundled demo/*/ carries a manifest.lock.json (input data + deterministic result tables) that verify --strict checks.

Worked Example

Lock a freshly-frozen extract:

python "${CLAUDE_SKILL_DIR}/scripts/version_dataset.py" manifest cohort.csv \
  --out manifest.json --seed 42 --provenance "KNHANES 2018 extract, frozen 2026-05"
# -> {"files": 1, "out": "manifest.json"}

Before re-running the analysis next month:

python "${CLAUDE_SKILL_DIR}/scripts/version_dataset.py" verify --manifest manifest.json --strict
# OK: 1 file(s) match the manifest.   (exit 0 — safe to run)

If someone silently re-exported the data with three extra rows:

========================================= 
 Dataset Manifest Verify
=========================================
DRIFT (3):
  ROW COUNT cohort.csv: 3457 -> 3460
  CHANGED column cohort.csv:bmi
  CHANGED column cohort.csv:hba1c
MANIFEST_DRIFT: dataset differs from manifest.   (exit 1 — STOP)

The analysis does not proceed: the result the manuscript will cite would no

longer match the locked data. The researcher reviews the drift, decides whether

the change is intended, and only then re-locks (manifest again) and records the

new provenance. A tabular file is compared on its logical content (schema +

per-column value hashes), not raw bytes — re-saving the same data, reordering

columns, or an --ignore-cols volatile timestamp column does not trip a false drift.

Anti-Hallucination

• Never claim a dataset is unchanged without running verify.
• Manifests record only observed hashes/schema; no provenance is invented — the

provenance note is user-supplied text.

• Report drift exactly as computed; do not downplay a changed column hash.