dsh-research-lab

When comparing agent reasoning strategies, the biggest pain point is rarely the model itself — it is the experimental environment: baselines drift, ablation matrices lack unified records, and manifests are impossible to reproduce. dsh-research-lab, a native DeepSeek Harness research plugin, pulls these scattered terminal and notebook steps into DSH Web, returning experiment management to a deterministic, auditable control plane.

Core capabilities revolve around three pillars. First, a Runtime Catalog ships with three verified paths — ReAct Default, Plan Guided, and Code Orchestration — while reserving design-stage slots for Plan-and-Execute, Tree of Thoughts, and Reflexion. Second, a Baseline protocol locks the DSH version, Profile, Preset, model, dataset, budget, metrics, and random seed in one shot, eliminating the awkwardness of reconciling parameters after the fact. Third, an ablation matrix allows free Runtime x repetitions combinations up to 64 planned runs, emitting a normalized JSON manifest with a SHA-256 hash for determinism. Storage is private, atomic, and idempotent; the API never exposes absolute paths or credentials.

It is built for developers running Agent Evaluations, Ablation Studies, or cross-strategy reasoning benchmarks. One caveat: the M0 boundary defines this as a Manifest-first control plane — it does not invoke models, does not incur API costs, and does not claim ToT or Reflexion are production-ready. It solves the problem of experiment management and result reproducibility, not the replacement of reasoning engines themselves.