Glossary

What is Data Traceability?

When a number in a report or a model’s output is questioned, traceability is what lets a team answer three things: where the value started, what was done to it, and which version of it was used. Without it, the investigation becomes a search through pipelines, notebooks, and shared drives.

A traceable record links:

  • the original source, kept unchanged
  • each derived version and the rules or steps that produced it
  • where each version was delivered, and whether it was applied there
  • the runs or reports that used it

The last two are easy to confuse. A version can be delivered and never applied, or applied and never used in a run, so the preparation, delivery, and run histories need to be kept as separate records.

How is data traceability different from data lineage?

  • Data lineage maps how data moves and is transformed between tables, columns, and jobs; tools build it from pipeline code or query logs.
  • Traceability follows a specific data version to where it was delivered, applied, and used in a run, including steps a lineage graph may not capture.
  • Column-level lineage can show what feeds a column; traceability also shows which version of it a given result used.

Where does data traceability break in AI work?

A credit model’s approval rate changes after retraining. The data team knows the training table was refreshed, but not whether the refresh included last month’s corrected income fields, which preparation rules ran, or which version the scoring service now reads. Each of those could explain the change, alongside the retraining itself, and none of the data questions can be checked.

The check is whether each result can be traced back to a fixed data version and the transformations behind it. In Syntitan the original Dataset is kept unchanged and preparation results are recorded as project Versions. Traceability is one of the six axes Ready for Raw Data diagnoses, alongside reproducibility, which asks whether the same evaluation can be repeated. Execution Trace, which will tie each run to the Version it used, is in development.

Frequently asked questions

Why is data traceability important for AI?

When an AI result changes or is challenged, traceability shows which data version and transformations were behind it, so the team can narrow down the cause instead of rebuilding the history.

What is the difference between traceability and reproducibility?

Traceability lets you verify where data came from and what was done to it. Reproducibility lets you identify the data and conditions behind a past result again and repeat the same evaluation.

Does traceability require keeping every copy of the data?

It requires a record of the original and of each derived version and the steps that produced it, enough to verify origin and changes. Whether the data itself is kept depends on the retention and deletion rules that apply; where data was deleted, the record should show what can still be verified.