Reading the audit trail

Clean data is only trustworthy if you can see how it got clean. Dotwave's audit trail is the running record of everything that happened to a dataset during cleaning, written in plain English rather than code. It lets you retrace every decision, explain your work to a client, and tell at a glance which changes you made by hand and which the AI proposed. This article explains what the audit trail records, shows sample entries, and covers how to filter it when a dataset has been through many steps.

What the audit trail records

The audit trail captures every cleaning operation, in the order it happened. Each entry is a small, complete account of a single change, and it records:

• What was done, described in plain English rather than as a formula or a snippet of code.
• Which column the operation touched, so you always know where a change landed.
• How many rows were affected, giving you the scale of each edit at a glance.
• Who did it — you, working manually, or the AI acting on a suggestion.
• When it happened, so the whole sequence reads as a timeline.

Because the entries are kept in order, reading the audit trail top to bottom is like watching the dataset go from raw to finished. Nothing is hidden and nothing is summarized away — each step stands on its own.

Sample audit entries

To see how readable this is in practice, here is what a handful of real entries look like:

• Removed 47 duplicate rows
• Filled 23 blank values in age with the median (43.0)
• Converted invoice_date to real dates
• Capped outliers in price_col
• Added calculated field "profit_margin"

Notice that none of these require you to read code to understand them. Anyone — you, a colleague, or a client — can look at "Filled 23 blank values in age with the median (43.0)" and know exactly what happened and why it mattered. That plain-English framing is deliberate: the audit trail is meant to be read by people, not just by machines.

Filtering the audit trail

A thoroughly cleaned dataset can accumulate a long list of operations, so the audit trail lets you narrow it down. You can filter by operation type, by column name, or by date. Filtering by operation type lets you see, say, only the outlier-handling steps or only the deduplication steps. Filtering by column pulls up everything that ever happened to a single field — useful when a client questions one specific number. Filtering by date isolates the work done in a particular session.

One of the most valuable filters is by who performed the step. Reviewing what the AI did versus what you did manually helps you audit the assistant's contributions before you stand behind them, and it lets you demonstrate to a client exactly where automated help ended and your judgment began.