Method
Is it signal, or is it error?
The values we review are held to two questions. AI agents gather the evidence; a person makes the decision.
The two questions
Does this value look wrong?
Error
Corrected, with the reason and evidence recorded.
Real signal
Kept as it is and annotated, so nobody “fixes” it later.
Undecided
Flagged to your team with the evidence. Not guessed.
Is this value missing?
Found
Filled from the source it was found in, and marked as filled.
Derivable
Calculated from reliable related values, with the method shown.
Genuinely missing
Left empty and labelled with the reason, not invented.
Principles
Never invent a value
A value we can’t find or reliably derive stays empty, with the reason recorded. A filled-in guess is worse than an honest gap.
Correct errors, keep events
An outlier is investigated before anything happens to it. Real spikes, drops and one-offs are part of the signal.
Show the uncertainty
When the evidence doesn’t settle a question, you see the question and the evidence. We don’t hide it behind a confident number.
Keep changes reversible
Changes are recorded with what they were, why they were made and who approved them, so they can be undone.
Who does what
Agents for scale. People for judgment.
Language and vision models let us review far more records than a manual sample. They don’t get the final word: the changes they propose are reviewed by a person, individually or as an approved rule, before they reach your data.
| Task | Done by |
|---|---|
| Read records, files and fields at scale | agent |
| Check values against sources and history | agent |
| Search for missing values elsewhere | agent |
| Propose a fix and write the reasoning | agent |
| Decide the rules for your data | human |
| Approve or reject each proposed change | human |
| Settle disputed and ambiguous cases | human |
| Sign off the final dataset | human |
The audit trail
Changes, accounted for.
Illustrative entries.