Data science projects accumulate a large number of choices — which dataset to use, which algorithm to try, why a certain approach was accepted or rejected. Without a structured record, those decisions become invisible and your work becomes hard to reproduce or audit.
Babel’s traceability panel lets you log three types of records directly within each CRISP-DM phase: datasets, experiments, and decisions. Every record is linked to a phase, so you always know when in the lifecycle something happened.
Accessing the traceability panel
Expand any phase card on the project pipeline page. Below the Kanban board you will find the Traceability panel, with tabs for Datasets, Experiments, and Decisions. Records displayed are automatically filtered to the current phase.
Phase filter
Each record type supports a ?phase= query filter. In the UI, the traceability panel automatically applies the filter for the expanded phase, so you only see records relevant to that part of the lifecycle. You can switch to a different phase using the phase selector at the top of the panel if you need to review records across phases.
Record types
Datasets
Experiments
Decisions
A dataset record documents a data source that was ingested or used during this phase. Log it as soon as you acquire a new dataset so the team knows its origin, size, and acquisition context.To add a dataset: open the traceability panel, select the Datasets tab, and click Add dataset.A short, recognizable name for the dataset. For example: Customer transactions Q1 2024.
Where the dataset came from — a database name, file path, external URL, or API endpoint. This is the primary reference for reproducibility.
A plain-text description of what the dataset contains, its domain, any known quality issues, and how it relates to the project objectives.
The number of rows or records in the dataset at the time of acquisition. Useful for tracking if the dataset changes over time.
The date the dataset was obtained, in ISO-8601 format (YYYY-MM-DD). Used to establish a timeline in the audit trail.
The CRISP-DM phase during which this dataset was acquired. Accepted values: business, data_understanding, data_preparation, modeling, evaluation, deployment.
Record every version of a dataset separately if it changes significantly between phases (for example, the raw extract vs the cleaned version after ETL).
An experiment record captures a modeling trial — the algorithm used, the parameters configured, and the outcome. Logging experiments systematically prevents duplicating work and provides evidence for model selection decisions.To add an experiment: open the traceability panel, select the Experiments tab, and click Add experiment.A unique, descriptive name for this run. For example: RandomForest baseline v1 or XGBoost depth-6 tuned.
The name of the algorithm or model architecture used. For example: Random Forest, Logistic Regression, BERT fine-tuned.
The hyperparameters and configuration used for this run, in any readable format. For example: n_estimators=200, max_depth=8, learning_rate=0.05.
The evaluation metric used to judge this experiment. For example: F1-score, RMSE, AUC-ROC. This field is used by the dashboard to sort and display the top experiments.
The numeric or qualitative result achieved on that metric. For example: 0.87 or RMSE = 142.3.
Your interpretation of the result. Did the experiment succeed? Should this model be promoted? What should be tried next? This is the most valuable field for future team members reading the record.
The CRISP-DM phase this experiment was conducted in. Most experiments will be logged under modeling or evaluation.
The dashboard surfaces the top 3 experiments by metric value in the Experiments widget. Keep the result field consistent (all numeric, all using the same scale) so the ranking is meaningful.
A decision record documents a significant choice made during the project — a methodology choice, a scope change, a data exclusion, a stakeholder agreement. Capturing decisions closes the gap between what you built and why you built it.To add a decision: open the traceability panel, select the Decisions tab, and click Add decision.A clear statement of the decision that was made. Write it as a resolved statement, not a question. For example: We will exclude records with more than 30% missing values from the training set.
The reasoning behind the decision — the evidence, constraints, or stakeholder input that led to this choice. This is essential for future audits and handovers.
Optional. Other options that were considered and why they were not chosen. Recording alternatives demonstrates due diligence and prevents revisiting rejected paths.
The CRISP-DM phase during which this decision was made. Decisions can appear in any phase, from business (scope agreements) through deployment (monitoring thresholds).
Log decisions as they happen, not at the end of the project. Memory fades quickly, and the context that makes a decision meaningful is often lost if you wait.
Deleting a traceability record
Each record card has a Delete button (trash icon). Deletion is immediate and permanent after confirmation.
Deleting a traceability record removes it from the project report as well. If the record has already been included in a shared PDF report, the report is unaffected — but future report exports will not include the deleted record.
Traceability in the project report
When you export a PDF report from the project dashboard, all datasets, experiments, and decisions are included in dedicated sections, organized by phase. This makes the traceability log directly usable as an appendix to a client deliverable or internal review document. 