Scrub Data: A Framework for Reproducible Data Curation with AI Coding Agents

Key results

1. 01

   ShownThe framework took a large, mixed pool of wildlife GPS tracking studies all the way to MoveBench, a benchmark for probabilistic animal-movement forecasting. A small team did the work in a short time.

   Input: 89M raw GPS points from 234 studies (215 Creative Commons Movebank studies, fourteen from a literature review, five suggested by experts). Output: 2.6M points from 807 individuals, 110 species and 147 studies. Two curators took three days, using Codex with GPT-5.4. · 3 Case study: Creating MoveBench; 3.1 Raw data acquisition and ingestion; 3.5 Final results

2. 02

   ClaimedEach dataset version is a node: a file manifest that reuses unchanged files by reference. Each edge stores an executable script with metadata on the request, the user and the affected files. In principle, any version can be rebuilt by replaying the steps from raw data.

   The graph is append-only. A script that fails creates no edge or node. Moving the current-dataset pointer gives undo and branching without deleting history. Default REST routes include steps, analyses, head and undo. The paper reports no test that rebuilds a dataset by replaying the graph. · DAG nodes; API specification; DAG nodes; DAG edges

3. 03

   ShownIn a first cleaning pass run from the terminal, the agent wrote graph steps that thinned fixes to one per hour (GPS rows only), capped data per individual and individuals per study, and assigned a preliminary temporal train/test split.

   Thinning to at most one fix per hour, with ±30 minutes of tolerance, cut the pool to 14M points. Then each individual was capped at two full years of hourly data, fifteen individuals were sampled per study (six were later kept after manual review for anomalies), and each individual's timeline was split 50/50 into train and test. This left 5.9M points. · 3.3 Initial terminal-based data cleaning; 3.4 Iterative visualization and cleaning; C.1 Dataset thinning

4. 04

   SuggestedAn interactive map and timeline viewer, built by the agent, exposed curation issues that shaped the benchmark design. These included uneven sampling rates, deployment lengths and study sizes, plus spatial overlap within subpopulations, which led to a shared train/test cutoff per subpopulation to prevent leakage.

   The viewer used deck.gl, Maplibre GL JS and CARTO basemaps. Some studies recorded at 1hz, deployments lasted from a few days to several years, and studies held anywhere from a single animal to hundreds. After consulting experts, the team set the shortest forecast horizon to one hour. · 3.2 Initial visualization development

5. 05

   SuggestedWith the prompt scaffold in its context, the agent wrote each transformation as a complete Python file, passed as a string to the API. It also built UI buttons that generate step scripts and record a username and a reason for each change.

   This comes from the MoveBench case only (Codex, GPT-5.4). The thinning logic was bundled into one script string and posted as a graph step. UI operations covered keeping or deleting individuals, aligning train/test cutoffs, assigning held-out individuals, removing studies, undo and marking a study complete. · Prompt scaffold; 3.3 Initial terminal-based data cleaning; 3.4 Iterative visualization and cleaning