Technical overview
What this release is built to train.
Every row contains a dense 2,000 to 5,000-word reasoning trajectory that trains a model to pause, audit itself, and plan verification before execution.
The five-stage structure covers request understanding, known-versus-uncertain facts, ambiguities, confirmation checklists, and 10 to 20 high-density web-search queries.
The matrix spans industry, professional role, software stack, task type, operating system, difficulty, and risk level, with deterministic prime-number scrambling to prevent repetition.
The role and environment coverage includes 200+ professional profiles and operating contexts across Apple, Windows, Linux, cloud shells, Android, ChromeOS, and iPadOS workflows.
Five-stage System 2 preparation before execution
Each example pairs a complex professional request with a 2,000 to 5,000-word reasoning trace. The target behavior is preparatory rather than directly executable: understand the objective and constraints; separate known facts from uncertainty; surface ambiguities; build a verification checklist; and produce 10 to 20 dense web-search queries for a downstream retrieval or frontier-model stage.
This structure trains small and edge-deployed models to function as reasoning routers. Instead of spending a frontier-model call on an underspecified prompt, the edge model identifies missing context, challenges unsafe assumptions, and prepares exact retrieval work. The intended capabilities include self-auditing, uncertainty calibration, prompt interception, clarification, RAG query generation, and disciplined handoff.
A seven-dimensional, billion-state generation matrix
The source matrix combines industry, industry-appropriate professional role, software stack, task type, operating system, difficulty, and risk. Seven large prime numbers deterministically scramble these dimensions into one billion valid permutations; only about 260,000 rows, or roughly 0.026 percent of that space, were sampled to reduce thematic repetition.
Coverage spans more than 200 roles and a wide operating-system surface: current and legacy Apple and Windows environments, Windows Server, major Linux distributions, WSL, cloud shells, Android, ChromeOS, iPadOS, and embedded contexts. This forces the reasoning trace to respect local software, hardware, platform, and professional constraints instead of generating generic plans.
Scale, schema, and training use
The release contains 260,293 observations and approximately 692.1 million tokens across all columns, including 646.9 million agent-reasoning tokens and 42.6 million user-prompt tokens. Generation consumed approximately 1.47 billion tokens. Chunked Parquet records preserve batch identity, role, industry, operating system, user prompt, and the full reasoning output.
The corpus is suited to reasoning fine-tuning for 7B-14B local models, retrieval-router training, prompt-quality interception, uncertainty-aware planning, and pre-processing before expensive model calls. It is an MIT-licensed training corpus, not a live search index; deployments should still validate generated queries and retrieve current evidence before execution.