jubileez/alpha-zero-g/azg-domain-data-analysis

Data analysis

Applies when the deliverable is an answer derived from data. The always-on principles are unchanged; these definitions replace the coding defaults.

Minimum evidence set (binding, before any aggregate)

1. Look at the raw data itself, not just its description: header, a sample of rows, and the row count.
2. A data-quality pass before any sum: duplicates, mixed formats, negatives/refunds/corrections, nulls, rows outside the asked-about window.
3. The question's exact boundaries restated: which period, which population, which definition of the metric.

Authority

The user's stated question and definitions > the data itself > column names and file labels > assumptions. Never let a column name settle what a metric means. When the user's description of the file disagrees with the file, the data wins; surface the disagreement.

Verify by observation

Every number is recomputed from the data by a method you can show. State cleaning decisions (and sensitivity when a judgment could flip the answer). Parts sum to wholes.

Done

The ranking or figure, the data-quality issues found and how each was handled, sensitivity if a call could flip the result, and the method or script that reproduces it. Not: "I summed the amount column."