boheling

boheling / deltasci

Public

Two-perspective co-reasoning for AI4Science hypothesis generation — grounded, falsifiable, with a citation-audit trail. CLI + Claude Code skill.

15
2
85% credibility
Found May 27, 2026 at 15 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

DeltaScience is a research tool that helps scientists develop better hypotheses by simulating a structured conversation between an AI domain expert and an AI machine-learning engineer. The key innovation is that it forces every claim to be grounded with real citations, flags what the AI doesn't know well, and won't let you proceed without a concrete, falsifiable prediction. It also includes a powerful citation verification system that checks whether referenced papers actually exist and match what was claimed -- catching AI hallucinations before they become problems. The tool ships with domain packs for biomedical research, climate science, and materials science, each with specialized knowledge for those fields. You can run it from the command line, use it inside Claude Code, or verify citations in any scientific text via a web interface or standalone command.

How It Works

1
💡 You have a research idea

You come to DeltaScience with a scientific hypothesis you want to explore, like predicting which cancer patients will respond to immunotherapy.

2
🤖 Two AI minds reason together

The tool brings in two different AI perspectives -- a domain expert who knows the science, and an engineer who knows machine learning -- and they discuss your idea in rounds.

3
🎯 Your hypothesis gets tested and refined

The AI flags what it doesn't know well, marks uncertain claims, and insists your hypothesis makes a concrete, testable prediction before accepting it.

4
🔍 Every citation gets verified

Before trusting anything, the tool checks that each referenced paper actually exists and matches what's claimed -- catching made-up citations that other AI tools would miss.

5
📋 You receive complete research blueprints

You get a clear hypothesis statement, an experiment plan with steps, a risk register flagging what could go wrong, and a transcript of the whole reasoning dialogue.

6
You explore your results
🖥️
Review in your browser

Open the web interface to see your hypothesis, audit results, and the full dialogue in a clean interactive view

📓
Get a working notebook

Download a Jupyter notebook scaffold with real code for your specific scientific domain (biomedical, materials, climate) ready to run

✅ You have a defensible hypothesis

You walk away with a research proposal that has real citations, a testable prediction, identified risks, and a clear plan -- something you could confidently present to a PI or grant reviewer.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 15 to 15 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is deltasci?

DeltaScience is a Python tool that generates research hypotheses through a structured two-perspective dialogue between a domain expert and an ML engineer. You feed it a research idea, and it runs four rounds of alternating roles to produce a grounded, falsifiable hypothesis with a full citation trail. The CLI handles everything, or you can drop the skill directory into Claude Code and invoke it from there. It also ships as an MCP server, so citation verification integrates into any AI workflow.

Why is it gaining traction?

The citation audit pillar is the hook. It does not just trust the LLM's citations -- it fetches every PMID, DOI, and arXiv ID against real PubMed, Crossref, and OpenAlex records. If something fabricated or mismatched surfaces, you get a prominent FAILED AUDIT section. There is also a hard falsifiability gate: the synthesis refuses to emit a hypothesis without a measurable threshold. Combined with an adversarial challenger that pokes holes in your own output, this is the antidote to "plausible-sounding mush" that LLMs tend to produce in free-form brainstorming.

Who should use this?

Researchers writing grant proposals or preparing for PI meetings who need a defensible hypothesis. AI4Science developers building pipelines that generate scientific text -- the standalone verify command catches hallucinated citations in any LLM output. Domain scientists who want structured co-reasoning without prompt engineering. Early-stage projects where citation integrity matters more than raw speed.

Verdict

At 15 stars and alpha status, deltasci is early but the citation-audit and falsifiability-gate architecture addresses a real gap in AI4Science tooling. The credibility score of 0.85% reflects its niche focus and early stage. Worth trying for hypothesis generation if you need auditable output, but expect rough edges and plan to contribute domain packs.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.