Dataset Versioning

This witness is intentionally small. It fingerprints a dataset bundle, then classifies candidate changes as same version, new version, or invalid evidence.

import copy, hashlib, json

BASE = {
    "raw": "s3://cf-demo/reviews/raw-2026-06-01.csv",
    "transform": "clean_v3.py:strip_html+dedupe",
    "split": {"train": ["a1", "a2"], "test": ["b1", "b2"]},
    "schema": "text,label@v2",
    "use_scope": "research-demo-ok",
    "provenance": "exported from consented review corpus",
}

def fingerprint(bundle):
    payload = json.dumps(bundle, sort_keys=True, separators=(",", ":"))
    return hashlib.sha256(payload.encode()).hexdigest()

def valid(bundle):
    if not bundle.get("use_scope") or not bundle.get("provenance"):
        return False, "missing use scope or provenance"
    train = set(bundle["split"].get("train", []))
    test = set(bundle["split"].get("test", []))
    if train & test:
        return False, "train/test split overlap"
    return True, "valid"

def classify(old, new):
    ok, reason = valid(new)
    if not ok:
        return "invalid evidence", reason
    if fingerprint(old) == fingerprint(new):
        return "same version", fingerprint(new)[:12]
    return "new version", fingerprint(new)[:12]

cases = {
    "rerun": {},
    "new raw export": {"raw": "s3://cf-demo/reviews/raw-2026-06-15.csv"},
    "license missing": {"use_scope": ""},
    "split leak": {"split": {"train": ["a1", "b1"], "test": ["b1", "b2"]}},
}

for name, patch in cases.items():
    candidate = copy.deepcopy(BASE)
    candidate.update(patch)
    print(name, "=>", classify(BASE, candidate))

The witness is not a replacement for a data catalog, a legal review, or a privacy review. It is a compact reproduction habit: the score comparison begins by checking whether the evidence bundle is identical, newly versioned, or invalid.