White paper · Draft v0.1 · October 2026

Strata

A university-hosted research workspace that turns a doctoral student's papers into a map of claims and evidence.

Design stage · nothing described here is built or measured yet

01 · Abstract

Automate the early stages. Keep the judgment.

Collecting sources, extracting content, organizing it and finding how papers relate is repetitive work. Strata does it; the researcher keeps every scientific decision and the standard methodology.

Not a new LLMA knowledge layer on top of existing models, built on published RAG 8 and GraphRAG 10 methods.
Graph, not chatPapers are split into meaningful units and linked in a multi-layer semantic graph.
Small contextA hybrid index passes only relevant passages to the model, so library size stops mattering.
On campusThe model is replaceable and can run on university servers or a private cloud the university controls.

Team name and host institution.

02 · The problem

The literature doubles every 17 years.

4.10% a year

growth in scientific publications 1

67.3weeks

mean time for a five-person team to conduct and publish one medical systematic review 2

Read the full section

Much of that time goes to repetitive work: collecting, first-pass screening, data extraction and comparing findings 26. A doctoral student usually does it across disconnected tools: a browser and publisher sites, a reference manager, a note-taking app and, lately, a chatbot. None of them keeps the relationships between the findings of different papers.

Local evidence: interviews with a few PhD students or a short survey on time spent on literature review.

Annual output relative to today, modeled from the 4.10% growth rate 1. Each marker is one doubling.

02.2 · Why chatbots are not enough

Fluent, confident, and often unsourced.

Chatbots are built for conversation, not for working on a specific library under a scientific procedure. The weakness is not the model. It is the missing data, retrieval and procedure layers around it.

All six weaknesses, with evidence

Fabricated references. 55% of GPT-3.5 and 18% of GPT-4 citations were fabricated in one test 3; hallucination is a known property of language models 4. Strata answers only from registered sources and links every claim to a passage.
Limited context. Of models claiming 32K tokens or more, only about half hold up at that length 6. Strata keeps an external index and sends only relevant passages.
Lost in the middle. Accuracy drops when key information sits mid-input 5. Strata keeps the context short and focused.
No structured memory. Relations between concepts are not stored 9, so questions like "where is the gap?" get weak answers 10. Strata keeps a multi-layer graph.
No procedure. No inclusion criteria, decision log or PRISMA flow 7. Strata builds them in.
Data leaves the university. Raw data and drafts cannot be used. Strata can run inside the university boundary. Reference to a university data policy or research-ethics guideline.

GPT-3.5 citations

55%

fabricated 3

GPT-4 citations

18%

fabricated 3

Each dot is one citation in a hundred. Walters & Wilder 3. And of models that claim a 32K-token context, only about half perform acceptably at 32K 6.

Same question, two answers

A general chatbotno library

Stratayour library · 30 papers

Both answers were written for this page. The left one reproduces failure patterns documented in 3, 5 and 6, and its citation is deliberately invented. The right one shows the answer format Strata is designed to produce; its citations point to this paper's real reference list. Tap a citation chip to see the source.

03 · Design principle

The machine suggests. The researcher decides.

Decisions stay humanInclusion, exclusion, interpretation and conclusions are recorded only with the researcher's approval, as in accepted active-learning screening tools 26,27.

Every output is traceableEach sentence links to a paper, page and passage. Unsourced claims are not displayed.

Procedure stays putProtocol, criteria and reporting follow standards such as PRISMA 2020 7, only faster and better documented.

04 · Architecture

Six layers. The model is one of them.

05 · Semantic graph

This paper's own library, as a graph.

The 30 references below, split into findings, concepts and claims, the way Strata would. Hand-built for this page; findings are paraphrased. Drag to rotate, pinch or scroll to zoom, tap a node.

Click the graph to explore · scroll zooms while active

Ask the graph

05.1 · How a paper is split

Four cuts turn a PDF into evidence.

The approach is close to GraphRAG 10 and LightRAG 24, with a schema designed for the scientific literature. Every unit stays linked to its source sentence.

Rhetorical roleEach sentence is labeled background, aim, method, finding or limitation, after Argumentative Zoning 14.

Entities & claimsConcepts, methods, data and claims are extracted with a fixed schema.

Citation intentUses a method, agrees or disagrees, read from the text around the citation 15,23.

Claim vs evidenceDoes a passage support or refute a claim? A known task with a standard dataset 16.

Passage level

"What was the sample size?"

Precise questions

Claim level

"Who disagrees with this?"

Comparison

Cluster level

"Where is the gap?"

Leiden communities 25, as in GraphRAG 10

Summary tree

"Summarize the field."

Paragraph to library, like RAPTOR 11

Final choice of graph tooling (for example Graphy, GraphRAG or an in-house build) and a reference to its documentation.

06 · Retrieval

Three searches, fused, then a short context.

Keyword search (BM25) 17 catches exact terms. Dense vectors 18 on an HNSW 19 or FAISS 28 index catch meaning. Graph traversal catches related evidence whose words look nothing like the question 10,24. Reciprocal Rank Fusion merges them 20.

Do long-context models make retrieval unnecessary?

RRF constant k 60

Passages sent to the model 4

Name and documentation of the token-saving index tool chosen (for example TokenSave) and which parts above it covers.

06.2 · Why the window stops mattering

Input size depends on k, not on your library.

k · c + q ≪ ~~N · L + q~~

N papers of average length L tokens, k retrieved passages of c tokens, q tokens of question and instructions. As N grows, full text grows linearly; retrieval stays flat.

Papers N

Paper length L

Passages k

A structural argument, not a measurement. L, k and c (500 tokens) are assumptions you can change. Answer quality depends on retrieval quality and must be measured with standard RAG metrics 29.

07 · Model & deployment

The model is a part, not the product.

We train no new model. Knowledge lives in the graph and index, so it updates without retraining and every answer shows its source 8. Swapping the model never means rebuilding the graph.

Commercial API

Provider's servers

Early stages on published papers. Fastest start; only retrieved passages leave.

Private cloud

University-controlled

Semi-sensitive data and research groups, with open-weight models such as Llama 3 22 or Qwen2.5 30.

University servers

Campus data center

Raw data, interviews, unpublished drafts, served efficiently with engines such as vLLM 21.

Inside the university boundary

Access to papers and copyright

Strata has no direct access to publisher databases and does not harvest publisher content in bulk. The researcher downloads each paper with their own legal access, and the browser extension registers it as a source with its metadata (DOI, title, authors, journal). Each library is limited to its owner or group.

Legal review of major publishers' terms (Elsevier, Springer, Wiley) on AI text processing, and the host university's data policy.

08 · Landscape

Each tool solves a slice.

Strata's difference is bringing the slices together on the researcher's own library, under university-controlled deployment.

General chatbotsChatGPT · Claude · GeminiFluent writing and summariesSource tracing, memory, procedure

AI literature searchElicit · AnswerThis · Consensus · SciSpaceFind and summarize open papersFull text of your own papers, linked

Citation mapsResearchRabbit · Connected Papers · sciteHow papers connect 23Claims and methods, not just papers

Review managersCovidence · Rayyan · ASReview 27Standard screening workflowAutomatic capture, graph synthesis

Reference managersZotero · Mendeley · EndNoteStore and citeUnderstands content, semantic search

Reconcile with the competitive analysis and check each tool's site before publishing; these tools change quickly.

09 · Limits, evaluation, roadmap

Every claim here is a hypothesis.

Extraction makes mistakes. Retrieval reduces hallucination but does not remove it 4,9, which is why approval is part of the design.
PDF conversion (equations, tables, scans) depends on how the paper is built.
Strata only sees papers the researcher adds. Search completeness stays their responsibility.
On-campus models may be weaker than commercial ones. The trade-off must be measured.

Hybrid retrieval finds the right passagesRecall@k · nDCGQuestion set with reference answers from a real library

Answers are faithful to sourcesFaithfulness · context precision 29Automatic evaluation plus human sample review

Zero fabricated references% traceable to a passageMethod of Walters & Wilder 3

Graph edges are correctPrecision · Cohen's κTwo experts judge a sample of edges

Early-stage time goes downHours per small reviewUser study with a control group

PHASE 1

Prototype

Extension, processing, hybrid index, cited answers.

PHASE 2

Graph

Semantic graph and technical evaluation.

PHASE 3

Pilot

A group of PhD students and a user study.

PHASE 4

On campus

University infrastructure, sensitive data.

Pilot library size, participants and discipline; timeline per phase; a specific ask of the reader.

References

Figures from references 1 to 3 were checked against the papers. Page and volume details of the others should be reconciled in a reference manager before publication.

Host the first pilot.

We are looking for doctoral programs and research groups.

Join the waitlist

Strata · White paper draft v0.1 · October 2026 · All examples on this page are illustrative.