About CoralDuck

v1.2.0 September 25, 2026

Projection: your cohorts on a map 🗺️

The Overview has a second layout beside the tree. Every row is a point on two coordinates of your data, and every cohort is a diamond at the mean of its rows, joined by the tree’s edges. Coral had only the tree; the projection layout follows Eckelt et al. 2023.

• Bring coordinates, or compute them. Pick any two numeric columns, or run UMAP in your browser. The embedding is seeded, so the same attributes draw the same map every time, and a large table is fitted on a sample of 50,000 rows.

  

• Cut a cohort with a lasso. Drag around the rows you want, and “Filter by lasso” opens Filter with it for every selected cohort, so the new cohort is still confirmed in the action view.
• Clusters become cohorts. “Find clusters” colors the map by its dense regions, one knob sets how many there are, and “Split by clusters” hands them to Split.

  

• Computed columns travel with your work. The embedding and the clusters are new columns, marked as computed. They come back on a reload, and “Export derived values” writes them to a file that goes beside the project file.

Full changes.

v1.1.0 September 19, 2026

Characterize: what sets your cohorts apart 🌲

Pick two or more cohorts, and a random forest ranks the attributes that tell them apart. It is Coral’s Kokiri operation, ported to fully run in your browser.

• Kokiri’s approach, extended. Each attribute is confirmed, tentative or rejected against a shuffled copy of itself, instead of being cut at a fixed threshold. Accuracy is out-of-bag: each row is predicted only by trees that did not train on it. Each tree draws its rows without replacement, which removes a known bias in the importance and serves small cohorts better. Stop keeps the forest grown so far, and Exclude and Limit start a narrower run.

  

• Parquet and JSON load, and both row exports can write Parquet.
• Shareable links. A link can open the app with a dataset already loaded, so you can send one to a colleague.
• Session management. Each tab keeps its own saved session, and the Welcome screen lists the others to continue or delete.

Full changes.

v1.0.0 September 15, 2026

Coral going serverless 🚀

CoralDuck reaches v1 as a browser-only tool. Coral’s workflow is here — the cohort tree, the three-area action view, filter and split, prevalence, compare, and the item table — running against an uploaded dataset rather than a database.

What that changed, in the places you will notice it:

• You bring the data. CSV and TSV, plain or gzipped, from your disk or a URL.
• Nothing is uploaded, and nothing is fetched. No login, no session on a server, and no third-party request from this site or from the app.
• Prevalence gets a better confidence interval: Agresti-Coull instead of Wald, which is too narrow when a cohort is small or its rate is close to 0% or 100%. The reference is any subset of the cohort’s own steps.
• Two row exports instead of one, plus a PNG of the whole tree, and a portable project file that carries an analysis between machines.
• A color scheme you choose — light, dark or system — across the app and this site.
• Compare shows one card per attribute instead of one large matrix. Each cell names its real statistic (Cramér’s V, or a rank-biserial effect size) and says which cohort is higher and how much the two overlap.

  

Not here yet: Characterize, the projection view, and the onboarding tooltips. Full changes.