A research workbench that refuses to let the model make things up

What it does

DeepScient is a scientific-research workbench that runs on the DeepSeek Harness (dsh) runtime. It breaks the research lifecycle — topic selection, literature search, experiments, analysis, figures, writing, submission, reviewer responses — into 12 stages with 55 checkable gates, backed by 36 tools, 98 skills, and discipline-specific data for 55 fields. Each gate is judged against real state in the project library, not against the model’s vibes about whether it’s done.

The interesting bit

The design principle is that every conclusion must trace back to data, and every dataset must trace back to a rerunnable experiment — the science_* tools reportedly offer no escape hatch for inventing a number. Where the system genuinely can’t judge (construct validity, novelty, ethics trade-offs), it hands you a checklist instead of a verdict. That’s a refreshingly honest division of labor.

Key highlights

• 55 gates evaluated against actual project state, not model self-assessment
• Reproducibility is on-disk: experiment logs, environment, versions, and random seeds are all persisted
• 8 built-in literature sources (Crossref, OpenAlex, arXiv, PubMed, Semantic Scholar, Europe PMC, etc.) plus an optional MCP layer with 34 more tools
• Journal intelligence covering 5000+ journals for venue selection and formatting
• Self-verification culture: 294 live-instance assertions, end-to-end runs, and a portability check across 78 files — the repo’s stated rule is every claim maps to a runnable command

Caveats

• Windows-first: the desktop path is documented for Windows 10/11 and Apple Silicon macOS; the macOS build is unsigned, so Gatekeeper will intercept first launch
• The desktop bundle is hefty — roughly 635 MB on Windows, 479 MB on macOS
• It’s a shell around DeepSeek Harness plus a science backend; you need dsh installed (or the bundled desktop version) and a real API key for the end-to-end path

Verdict

Worth a look if you want an agent-assisted research workflow with actual provenance and reproducibility discipline rather than a paper generator. Skip it if you just want a chat wrapper that writes you a literature review by dinnertime.