adhd: Independent Reasoning Architecture: Novelty 8 vs 3, Trap 8 vs 2

Existing Chain-of-Thought or Tree-of-Thought approaches are limited in exploration scope because they are anchored to the first answer or share a single shared context. ADHD defines this as an architecture problem rather than a prompting problem, generating N independent reasoning processes under intentionally distorted cognitive frames. In the divergence phase, context sharing is completely blocked to secure ideas from diverse perspectives.

Diverged ideas undergo scoring, clustering, trap removal, and deepening of surviving ideas through a separate critique pass. Unlike judgments made by a single model at once, independent LLM judges evaluate based on Breadth, Novelty, and trap detection capability. Benchmark results show that the ADHD approach outperforms single-shot methods with Breadth 9 vs 6, Novelty 8 vs 3, and trap detection capability 8 vs 2.

It is suitable for prompts in the form of 'suggest multiple methods' for design decisions, ambiguous debugging, naming, API surface design, and strategy formulation. Several open-source projects such as repowire, mstack, and zk-flow-oss have officially adopted or integrated this framework. In particular, repowire has converted ADHD into message orchestrator raw data and included it in the default template, while mstack is distributing it as a 'think' plugin.