StartupEnv corpus: 2,000 simulated SaaS startup episodes with delayed outcome labels

StartupEnv: A Reproducible Simulation Environment for Longitudinal Synthetic Startup Data This release accompanies the research project StartupEnv: A Reproducible Simulation Environment for Longitudinal Synthetic Startup Data. It provides synthetic longitudinal trajectories of software-as-a-service (SaaS) startups generated using StartupEnv, a Gymnasium-compatible simulation environment designed to study startup development, decision-making, external shocks, and delayed outcomes over time. The primary corpus contains 2,000 simulated startup episodes generated under two decision procedures: a deterministic threshold-based heuristic (P1) and an independent bounded-random decision procedure (P2). Each episode simulates up to 120 monthly steps. The release includes raw simulation records, curated snapshots, episode-level outcome labels, summary statistics, and analysis outputs. The package also contains additional experiments examining the effects of selected startup decisions and external shocks, a randomized-shock-timing corpus, calibration and validation evidence, data-generation and analysis scripts, a data dictionary, software dependencies, and checksum information to support reproducibility and verification. Dataset contents Raw longitudinal simulation records and curated snapshots. Episode-level data and delayed outcome labels. Summary statistics and analysis results. Additional experiments and randomized-shock-timing results. Calibration and validation evidence. Python scripts for corpus generation, analysis, experiments, and calibration metrics. Documentation describing the data fields, package structure, and reproduction workflow. Intended use The release is intended for research on synthetic longitudinal data, agent-based and computational simulation, startup trajectory analysis, decision procedures, external shocks, and simulation-based experimentation. Researchers can use the data to examine patterns produced by the simulation, reproduce reported analyses, and investigate alternative analytical approaches within the limits of the environment. Limitations The corpus is generated by a simulator and is not calibrated to real startup populations. Its distributions and outcomes reflect the model's assumptions, implemented decision procedures, and experimental configurations. It should therefore be interpreted as a controlled simulation dataset, not as empirical evidence about real companies or as a validated predictive benchmark for real-world startup survival or performance. Reproducibility and licensing The release includes documentation, scripts, dependency information, and SHA-256 checksums to facilitate inspection and reproducibility. The dataset and release documentation are distributed under the licence specified in the package's root LICENSE file. The scripts in scripts/ have a separate MIT licence in scripts/LICENSE. These licences apply only to the materials for which the authors have the relevant rights and do not automatically relicense third-party or upstream components. This dataset is associated with the research paper StartupEnv: A Reproducible Simulation Environment for Longitudinal Synthetic Startup Data (CICT 2026, paper ID 747).