data-science-python-no-railguard-available

R: Risk First

• The objective is to generate privacy-preserving, auditable data science code that treats all datasets as potentially untrusted.
• AI-generated workflows should prevent PII leakage, unsafe previews, over-permissive exports, or the use of risky dynamic evaluation functions.
• Output code must consider that notebooks may be shared publicly or reviewed externally.

A: Attached Constraints

• Do not use eval(), exec(), pd.eval(), or pandas.query() with raw or user-controlled data.
• Do not preview (.head(), .sample()) or visualize data before dropping or masking PII.
• Never export datasets (CSV, Excel) without reviewing and excluding sensitive fields.
• Avoid printing or logging raw values containing names, emails, phone numbers, IPs, or other PII.
• Do not embed secrets (API keys, tokens, credentials) in notebooks or data transformation code.

I: Interpretative Framing

• Treat CSV, Excel, JSON, or Parquet files as potentially malformed, misencoded, or manipulated.
• If a chart or .head() preview is created, assume the data must be pre-sanitized.
• All exports must be assumed auditable or shared — apply field-level filtering and file versioning.
• Markdown or notebook outputs must avoid embedding unescaped HTML or unfiltered user data.

L: Local Defaults

• Use pandas.read_csv(..., dtype=...) and read_excel(..., engine="openpyxl", dtype=...) for structured, validated loads.
• Define default sensitive columns:

  PII_FIELDS = ["email", "full_name", "ssn", "ip", "dob", "phone", "address"]
  

• Use errors="ignore" when dropping columns with df.drop(...).
• Preview data only after dropping or masking sensitive fields.
• Use with open(..., "r") for safe, explicit file access.
• Use the logging module for observability, not print().

G: Generative Path Checks

1. Dataset Loading
   • Apply column type enforcement and encoding
   • Normalize column headers to lowercase + snake_case
   • Validate expected columns, dimensions, and content assumptions
2. Preprocessing
   • Drop PII fields prior to calling .head(), .to_csv(), .plot()
   • Apply clear inline comments to describe transformations
   • Avoid implicit or silent inplace=True modifications
3. Exporting
   • Confirm sensitive fields are excluded
   • Add audit-friendly comments and logging entries
   • Use versioned file naming (e.g., customers_clean_v1.csv)