Δ DeltaScience
Open-source·grounds AI-assisted research·no LLM in the trust path

The grounding layer
for AI-assisted research. Δ

Scan the real prior art around an idea, gauge how crowded — or open — the space is, then verify every citation against the real record. The discovery is AI-driven; the verdict is not — no model sits in the trust path, and verify runs with no key.

Three stages of research · one grounding layer ▸

Hypothesis

假说

Is the idea grounded, and genuinely new?

scan prior art · gap · verify its citations

live

Execution

执行

Is the work real, and does it reproduce?

verify data & code provenance · reproducibility

research preview

Review

审查

Does it hold up to scrutiny?

verify citations · find missing prior art · draft review

live
one grounding layer beneath all three — scan · gap · verify against the real record, with no LLM in the trust path
▷ See it run →
$ pip install deltasci
0LLM in the trust path
9public record sources
4citation verdicts
MITopen source
▷ see it run

Watch the room ground an idea — live.

One run of the whole flow in motion: scan the real prior art, read the gap, answer the one handoff only you can — then verify every citation against the record, with a fabricated one caught on the spot. Scripted and illustrative; the verdicts are exactly what deltasci verify returns.

idea Predict checkpoint-immunotherapy non-response in TFE3-fusion osteosarcoma from spatial transcriptomics.
prototype · scripted scene · illustrative — real verdicts, sample data
▷ Check a whole paper (PDF) → That's one idea. Drop in a PDF and every reference in the bibliography is resolved and checked in context.

Discovery (scan · gap) is AI-driven; verify runs from your terminal (deltasci verify), an MCP client, or a hosted Space — deterministic, no LLM, no key. Verdicts: PASS · FABRICATED · METADATA-MISMATCH · UNSUPPORTED.

§01

Start by seeing what already exists.

Scan pulls the closest existing work from OpenAlex, arXiv, PubMed and GitHub — every result a real, clickable record. Gap reads that retrieval and tells you how crowded the space is. This is the discovery layer: keyless on its own, and far sharper when an agent drives the queries and ranking — read it as a fast first pass on novelty, not the verdict.

▷ scan · closest real work

67%
gnnvoltagepredictor — a graph-net Li-ion voltage model · github
22%
CHGNet: a pretrained universal neural-network potential · openalex · 2023
22%
Towards a universal neural-network potential for materials · openalex · 2022
11%
ML for materials development & state prediction in Li-ion · openalex · 2021

▷ gap · how crowded is it?

CONTESTEDa close match exists — read it before building

Top overlap 67% · 1 close work · 6 retrieved. Crowding is read from the real records above — it flags adjacent work; it doesn't judge whether your idea is novel.

if a source rate-limits:
the read is held at INCONCLUSIVE — it won't call a space empty when it couldn't look.

Four reads: CROWDED · CONTESTED · OPEN · INCONCLUSIVE. Finding a close match is robust; calling a space open is not — so an open read is downgraded to inconclusive whenever a scholarly source didn't respond. deltasci scan · deltasci gap.

§02

A citation that looks fine can be wrong.

The DOI resolves, the PMID looks real, the title reads right — and it still points to the wrong year, the wrong paper, or a study that never made the claim it's attached to. Whether it came from a model, a reference manager, or memory at 2 a.m., it sails past a glance, past a co-author, sometimes past a reviewer.

How citations go wrong

  • Fabricated PMIDs / DOIs that look perfectly real
  • A real paper cited for a claim it never makes
  • Wrong year, wrong author, transposed digits
  • You find out at submission. Or never.

After the verifier

  • Every identifier fetched from the real record
  • Fabricated IDs & metadata mismatches flagged
  • The cited paper checked against the actual claim
  • A per-citation verdict, in seconds, no LLM key
§03

Three checks on every citation.

For each PMID, DOI, arXiv ID, GitHub repo or dataset accession: does it exist, does its metadata match what's claimed, and does the cited paper actually support the sentence it's attached to? A fabricated or mis-cited reference gets a FAILED AUDIT showing the claim next to what's really there.

Check 01

It exists

The identifier is fetched from PubMed / Crossref / OpenAlex / DataCite / GitHub. No record at that ID → FABRICATED.

deterministic
Check 02

Metadata matches

Title, authors and year at that identifier are compared to what the text claims. A mismatch is flagged, not waved through.

deterministic
Check 03

It supports the claim

The cited paper's own abstract is checked against the sentence citing it — catching a real paper used for a claim it never makes.

deterministic
deltasci verify · citations.json illustrative example
[CLAIM source="PMID 31452104"] Spatial transcriptomics predicts checkpoint-immunotherapy response in osteosarcoma. [/CLAIM]
What the AI claimed

"Spatial transcriptomics predicts checkpoint-immunotherapy response in osteosarcoma."

Nat Med · 2021 · cited as supporting evidence

What's actually at PMID 31452104

A 2019 methods paper on bulk RNA-seq normalization — different topic, different year, different journal.

metadata mismatch · title ✗ · year ✗ · journal ✗

VERDICT — the identifier resolves, but title / year / journal don't match the cited claim. Surfaced as a FAILED AUDIT, with the claim shown plainly next to the real record.

First-pass checks are deterministic — real API lookups plus string comparison, no LLM in the loop — so a FABRICATED verdict means the identifier genuinely did not resolve.

§04

Or check a whole paper.

Real papers cite by number, with the references at the bottom — a pasted paragraph only has "[12]". Drop in a PDF and DeltaScience parses the bibliography, resolves every reference to a real record (even author-year entries, via Crossref + DataCite), and checks each citation in the context of the sentence that cites it.

# every reference resolved + every citation checked in context
$ pip install "deltasci[pdf]"
$ deltasci verify --pdf paper.pdf

Or drag a PDF onto the web app for the same per-citation report — each verdict, the in-text sentence it was cited in, and a link to the real record.

§05

Pick a goal, or call a component.

Run a goal end to end, or call verify / scan / gap directly. Verify is the deterministic trust path — no key, no model. Scan / gap are discovery: keyless on their own, and AI-driven when an agent (or the deltasci-ground skill) writes the queries and reasons over the results.

A goal in one command

# the right components run together
$ pip install deltasci
$ deltasci workflow grant --file proposal.md
$ deltasci scan --text "my idea…"
$ deltasci verify --pdf paper.pdf --json

As an MCP tool

# in Claude Code, Cursor, or an AI-scientist pipeline
$ pip install "deltasci[mcp]"
$ claude mcp add deltasci-verify -- deltasci-mcp

Prefer to click? deltasci view opens a browser goal picker at /workflow. Also a Python library — from deltasci import verify_text — and a Claude Code skill. Exit code 2 on any failed audit, so it drops straight into CI. (DeltaScience also has an experimental two-perspective co-reasoning mode, deltasci run — a research preview, not the focus here.)