Tooling that helps reviewers read faster by turning papers into structured ‘claims and evidence’ graphs with audit-friendly outputs.
System that extracts structured nodes (claims, evidence, limitations, etc.) from papers and scores them to support literature review and comparison across a collection.
Reading dozens of papers is slow partly because the structure is inconsistent. Plato’s Cave uses modern language models to extract a consistent structure (e.g., claims, evidence, limitations) and then runs a scoring pipeline so papers can be compared more systematically.
Plato's Cave began as a GatorHack hackathon build — a first-place win at UF AI Days — and grew into a research paper on whether a paper's claims actually hold up.
Plato's Cave is an open-source, human-centered system for validating research papers. You give it a PDF, a URL, or even a natural-language query; it locates the paper, breaks it into the individual claims the authors are making, sends browser-based agents out to check each claim against outside evidence, and returns a paper-level integrity score with the full reasoning behind it.
It started as a hackathon project — first place at the UF AI Days GatorHack — and grew, with a team at the University of Florida, into a research paper. I designed the system's mathematics and architecture: the claim-graph formulation and the trust-propagation scheme were my ideas, as was the construction of the experiments.
The system runs a six-stage pipeline. The important design idea is in the last stage: instead of scoring claims in isolation, trust propagates along the dependency edges of the graph, so a confident-sounding conclusion built on weak premises cannot inherit a high score.
Every node in the graph is scored on six dimensions before trust is propagated. Keeping the axes explicit is what makes the final number auditable rather than a single opaque verdict.
Plato's Cave is deployed as three cooperating services rather than one monolith, which is what lets the verification agents run in parallel and in isolation from the app the reviewer is looking at.
The part I care about most is that this is not meant to replace a reviewer. Because the frontend can share a browser session with an agent, a person can watch exactly what each agent checks and step in. The output is an integrity score you can trace back through the graph, not a verdict handed down from a black box.
Yes. The code is public on GitHub and the paper is available as an arXiv preprint.
No. It measures how well a paper's claims are supported by available evidence, which is a narrower and different question than authorial intent.
Because it is an aggregate of six auditable dimensions propagated through a dependency graph — you can always open the graph and see which claim dragged the score down, rather than trusting a single opaque number.
I designed the mathematics and architecture of the system — the claim-graph formulation and the trust-gated propagation were my ideas, as was the construction of the experiments. I am the paper's second author.
@misc{maldaner2026platoscave,
title = {Plato's Cave: A Human-Centered Research Verification System},
author = {Maldaner, Matheus Kunzler and Valle, Raul and Kim, Junsung and Sultan, Tonuka and Bhargava, Pranav and Maloni, Matthew and Courtney, John and Nguyen, Hoang and Sawant, Aamogh and O'Connor, Kristian and Wormald, Stephen and Woodard, Damon L.},
year = {2026},
eprint = {2603.23526},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2603.23526}
}