SciWay — Provenance-grounded, refusal-first memory for scientific QA

The problem · why now

Teams lose their own memory — as the EU starts mandating they keep it.

What a lab tried, why it decided, and what failed is scattered across ELNs, email, file shares, results and decision notes. Generic memory/RAG tools answer confidently even when the evidence isn't there, and rarely cite or refuse. Meanwhile Horizon Europe, DFG/NWO and NFDI now tie funding to FAIR data-management — turning lab memory into a budgeted requirement.

How it works · the core differentiator

Three honest states: supported, conflicting, refused.

Ask a question and SciWay retrieves the evidence, classifies it, and shows its work: a cited answer when the evidence supports it, both sides when sources disagree, and an honest refusal when the project simply doesn't contain the answer.

Supportedgrounded + cited

SciWay AI agent: a supported answer backed by cited source quotes with provenance

Evidence supports it — answer with exact source quotes you can trace.

Conflictingboth sides shown

SciWay AI agent: conflicting evidence surfaced, showing two sources that disagree on a value

Sources disagree on a value — both are flagged, not silently merged.

Refusedno evidence → no guess

SciWay AI agent: a calibrated refusal showing what was considered but found insufficient

Evidence is missing — it refuses, and shows what it considered.

See it yourself

Paste text. Ask. Watch it ground — or refuse.

This runs SciWay's deterministic engine right here — no sign-up, no API key, nothing stored. The answer is a verbatim sentence from your text; ask something it doesn't cover and it refuses instead of guessing.

Try: "what temperature was the stability study?" (answered) vs "what protein concentration was used?" (refused). Synthetic / non-PII content only.

Your grounded answer — with the cited source sentence — will appear here.

Features

Built as a trust layer, not another chatbot.

A grounded-QA workspace for your team's own corpus — every capability keeps the same provenance + refusal contract.

🔎

Grounded QA & provenance

Every claim carries its source system, title, page / row / line, exact quote and URI. Inspect the evidence behind any sentence.

🛑

Calibrated refusal

When the evidence isn't there, SciWay abstains instead of bluffing — and shows what it considered but found insufficient. On by default.

⚖️

Conflict surfacing

When sources disagree on a value, both are flagged with provenance — so decisions rest on the full picture, not the first hit.

📥

Multi-format ingestion

Drag-and-drop multiple files (PDF / DOCX / XLSX / CSV / MD / TXT / JSON) or paste text into an isolated project — with page / row / line citation provenance.

🕸️

Knowledge graph

An interactive concept map of your project's corpus — see how documents, entities and findings connect at a glance.

🧠

Project memory

Saved Q&A history per project, so the team's questions and grounded answers accumulate into a reusable record.

📤

Citable evidence export

Export the project's documents, saved Q&A and provenance as a citable evidence log (Markdown / JSON), plus FAIR metadata.

🔒

Answers without an LLM

The deterministic engine returns the verbatim answer sentence + provenance with no API key — nothing leaves the machine by default.

🇪🇺

EU governance & FAIR

FAIR-aligned export (DataCite / Dublin Core / RO-Crate), EU / self-hosted inference with PII minimization — stated as design intent for pilots.

Roadmap

Where SciWay is going.

Honest status — shipped, in build, and exploring. The trust contract (provenance + refusal) holds at every stage.

Shipped

Grounded QA with citations, conflict & calibrated refusal
Bring-your-own multi-file upload + paste, per-project isolation
Project memory (saved Q&A history) & knowledge-graph concept map
Citable evidence export (Markdown / JSON) + FAIR (DataCite / Dublin Core / RO-Crate)
Read-only Zotero integration (connect a library, sync read-side)
Reproducible open-dataset eval (QASPER / SciFact / SQuAD 2.0)

In build

Semantic embeddings on by default (portable cosine floor)
EU / self-hosted LLM-written answers (opt-in)
Evidence-based conflict detection (beyond keywords)
Read-side connectors designed: Google Drive, Benchling / LabArchives ELN, Zenodo

Exploring

Multi-hop reasoning across silos
Institutional / RDM-office deployments
Audit-grade evidence logs & provenance graphs
Design-partner pilots (German/Dutch life-science labs)

Trust & EU

Honest by design.

SciWay is a venture-track prototype. Compliance features (GDPR-aware data handling, EU hosting, FAIR-aligned export) are stated as design intent for pilots — never "certified". Inference is local-by-default; external inference is opt-in, key-gated, and PII-redacted. Bundled lab data is fully synthetic; the demo accepts synthetic / non-PII content only.

View the open-source repo →

Help shape SciWay — and try it in 2 minutes.

Are you a PI, researcher, or RDM/data steward who loses track of what your team already tried? Tell us where it hurts — your answers directly steer the roadmap.

Open the MVP on Render → Take the 2-min survey

Get in touch

Pilots, partnerships, questions

We're talking to academic and translational life-science labs and their institutions about design-partner pilots.

Email · hello@sciway.software

Code · github.com/Rebell-Leader/sciway-eu

See it

One app, on Render

Open the MVP: a built-in synthetic demo lab to explore with zero sign-up, plus bring-your-own projects — create an account, upload your docs, and ask.

Open the MVP → View the code

The memory layer for a research team's own work.