Start with a working database.

Passenger records from the 1912 RMS Titanic voyage - the canonical beginner classification dataset. This bundle carries the **complete** Kaggle "Titanic: Machine Learning from Disaster" distribution: all 1309 passengers, the official 891/418 train/test split, and the `gender_submission` baseline, restructured for Strata so the split and the curation steps are branches rather than a folder of CSVs.

A synthetic, deterministic A/B test - one experiment, three arms - laid out so each arm lives on its own Strata branch (`control`, `variant-a`, `variant-b`). It exists to demonstrate the cross-branch comparison workflow: the same query, run per branch, is the entire analysis loop. ## What's in the data

Population, GDP, and inflation for every country and region the World Bank tracks, 2000-2025 - the "how big / how rich / how fast are prices moving" table agents reach for in analysis tasks, one document per country instead of six thousand API rows.

The three IANA HTTP registries - status codes, methods, and header fields - as one-read lookups. The reference an agent writing or debugging HTTP calls needs constantly, without bundling a copy or guessing from memory.

Fisher's 1936 iris measurements - the smallest, most familiar dataset in machine learning, restructured for Strata as a tutorial-grade bundle whose version history walks through the dataset's real, documented curation corrections.

108 version strings with their validity verdict under Semantic Versioning 2.0.0, 45 ordered precedence pairs, and a full precedence ranking of every valid string - the edge cases dependency logic gets wrong, as one-read lookups. Plus a second branch that reads the same corpus through the tolerant dialect real tooling actually uses, so `v1.2.3` is one `--branch` flag away from being valid.

GitHub Events-shaped developer activity - pushes, pull requests, issues, stars, forks - restructured for Strata as an append-only event log grouped into streams by event type, the natural shape for "what happened, in what order" questions over developer activity.

A fully synthetic, deterministic snapshot of a working AI agent's memory - episodic memories, an append-only event trace, and live working state - pre-forked into three experiment branches that each apply a different memory-retention policy, so you can compare retention strategies against identical starting conditions with a single clone. ## What's in the data

The SPDX License List - 727 licenses and 84 exceptions, release 3.28.0 - as one-read lookups. What an agent reviewing dependencies, generating a NOTICE file, or gating redistribution needs from the list, without bundling a copy of it or recalling it from memory.

Country, currency, and language codes from three ISO standards, compiled directly into Strata as a lookup table - the small, dependency-free reference data an agent developer reaches for constantly and shouldn't have to bundle by hand.

The GroupLens MovieLens 100K catalog restructured across three Strata primitives - movie metadata as JSON documents, catalog aggregates as KV, and per-movie genre vectors in a similarity-searchable vector collection - so a single clone gives you a recommendation sandbox with no preprocessing step.

Thirty real benchmark runs from Strata's own performance suite, replayed as an event log - a first-party history of how the engine's latency, scaling, and workload numbers moved across roughly two months of development, and a worked example of storing run history as events rather than a flat CSV.

Wikipedia article metadata restructured for Strata as one JSON document per article, with a KV title-to-slug index over the top - the shape a future retrieval-augmented-generation corpus would build on once article text and embeddings are added.

Stack Overflow questions and answers restructured for Strata as JSON documents with their comment threads embedded inline, so a single `json get` returns a full discussion - no join against a separate comments table.