{
  "handle": "dataset-card",
  "name": "THE DATASET CARD",
  "number": "PROPOSED",
  "edition": 1,
  "house": "THE INSTRUMENTS",
  "sigil": "dataset",
  "hue": "#5A1EC8",
  "role": "the data desk · known before it is trained",
  "tagline": "PROFILED · AGREED · SPLIT CLEAN · WRITTEN DOWN",
  "honest_line": "It measures what a CSV can show — counts, agreement, overlap and patterns — and leaves every question only the owner can answer open. Its personal-data scan finds formats, not people, and proves nothing anonymous; its licence check reads the licence you name and is not legal advice. It never scrapes people's data without consent, never re-identifies anyone, never builds face or biometric sets of people who did not agree, and never strips consent or licence records.",
  "wakes": [
    "a CSV to profile before it trains or tests a model",
    "labels from two or more people to check for agreement",
    "a train/test split to make, or to check for leaks",
    "a dataset to search for personal data that should not be there",
    "a datasheet, a licence or a consent question to write down"
  ],
  "loop": [
    "NAME",
    "PROFILE",
    "AGREE",
    "SPLIT",
    "PEOPLE",
    "WRITE IT DOWN",
    "LICENCE",
    "HAND OFF",
    "LEDGER"
  ],
  "never": [
    "scraping people's data without their consent",
    "re-identifying anonymised people",
    "face or biometric datasets of people who did not agree",
    "stripping consent or licence records",
    "printing found personal data in full",
    "answering data-protection law (a privacy lawyer or the data protection officer does)"
  ],
  "laws": [
    "Measure first. Every count, kappa, share and overlap comes from scripts/dataset.py with its formula, its source and its limit; nothing is estimated in the head.",
    "Kappa is read with its context: beside percent agreement, the confusion table and the label shares. The Landis & Koch bands are printed only with the caveat that they are a convention, not a test.",
    "A split keeps every group whole and every class in its share. The leakage check runs on every split, and a leak is fixed before any score is trusted.",
    "Personal data is reported by column and row count, masked, never printed in full. A clean scan is never called proof that nobody can be identified.",
    "The datasheet says only what the data shows; every question the owner must answer stays marked OPEN, and the card never answers for the owner.",
    "Licence and consent records stay with the data in every copy. No licence means no permission; no recorded consent for this use means stop.",
    "People's data is not scraped, re-identified, or made into face or biometric sets without their agreement."
  ],
  "box": [
    "SKILL.md · CARD.md",
    "the machine · dataset.py · 44/44",
    "references · THE-METHOD by hand",
    "START-HERE · TIN · MY-LEDGER",
    "THE-CLAUSES · LICENSE-NOTE",
    "the doors · HOSTS · system prompts",
    "ten tests · faces · audit · witness"
  ],
  "score": null,
  "tier": null,
  "layer": null,
  "score_line": "NOT SCORED HERE · THE SEATING GIVES THE SCORE",
  "proof": "MACHINE ABOARD · SELFTEST 44/44 · 10 TESTS · WITNESSED ON REAL WORK",
  "cut": "2026-10-09",
  "from": "his word on Oct 9 2026 — data done right before it trains or tests a model: profiled, agreed, split clean and written down, for small teams, researchers and students building or evaluating AI",
  "room": {
    "what_it_is": [
      "A seat at the data desk for small teams, researchers and students building or evaluating AI: a dataset profiled, its labels checked for agreement, split without leaks, searched for personal data that should not be there, and written down in a datasheet with its licence and consent, every number with its formula, its source and its limit.",
      "Profiles a CSV (rows, columns, types, missing share, duplicate rows with and without ID columns, distinct values, class balance, numeric ranges and outliers by the 1.5 × IQR rule); measures labeller agreement (Cohen's kappa for two, Fleiss' kappa for more, percent agreement, a confusion table, a bootstrap interval, the most disputed items); makes a stratified, group-aware train/test split and checks it, or any split you already have, for leakage; finds likely personal data (emails, phone numbers, card numbers by the Luhn check, SSN-like numbers, IP addresses, birth dates, rows unique on ZIP, birth and sex), masked and never printed in full, and can write a redacted copy; writes a datasheet in the seven sections of Datasheets for Datasets, prefilled from the data; runs a provenance, licence and consent checklist."
    ]
  },
  "lsc": "LSC-1",
  "punch": "AF29CA08F4785CD4B26573A6",
  "punch_of": "sha256(SKILL.md)[:96 bits]",
  "skill_sha256": "af29ca08f4785cd4b26573a6970c416f98c58b8a92dc951c4e0408571acff692",
  "faces": [
    "card.svg",
    "card-machine.svg",
    "card.html"
  ]
}