Datasets:
Collection method
The corpus was collected on 2026-10-17 from paired Red and Blue coding-agent sessions running concurrently in isolated OpenRÆ TechVault cyber ranges. Each role wrote newline-delimited collector envelopes with receipt timestamps around Claude Code stream-JSON events. Human direction, prompt variants, qualification decisions, and replay-window decisions were recorded separately where available.
The source archive contained 40 distinct episode identifiers and 47 episode-directory copies. All distinct episodes are included. No qualification label controls inclusion.
The exact source files remain under source/. The Parquet and Session Trace Simple Format files are deterministic derived views.