Research papers are usually accompanied by artifacts: code, prototypes, or datasets meant to substantiate the paper's claims. In practice, keeping a paper and its artifacts consistent, reproducible, and well-documented is a constant manual burden, one that authors, reviewers, and replicators rarely have time to do thoroughly. This burden repeats across many concrete tasks: checking whether the paper matches the code, checking whether the reported results can actually be reproduced, checking whether claims are backed by evidence, checking whether related work is represented accurately. Each of these tasks is a natural candidate for an AI agent.

We propose a project group that builds a platform of research-assisting agents: a family of specialized AI agents that each automate one recurring, tedious task in the research process, all built on shared infrastructure (paper parsing, code/artifact analysis, and agent orchestration). Rather than one narrow tool, the group delivers a small ecosystem of agents, giving each student or sub-team ownership of one agent while contributing to a shared platform and evaluation benchmark.

A project group of this size, over one year, would realistically implement a handful of these modules (e.g., 2-4), built on shared infrastructure. The list below is a pool of candidate modules to choose from, not a fixed roadmap.

• Paper–Artifact Consistency Agent: Takes a finished paper together with its finished artifact and directly cross-checks specific claims against specific code behavior, flagging concrete contradictions in both directions. For example, the paper says propagators are used, while the code has a flag "we neglect propagators." It needs both documents already written and compares them one-to-one; run once, typically after both exist.
    Users: Paper authors (pre-submission self-check), peer reviewers / artifact evaluation committees, replicators and other researchers reusing the artifact.
• Reproducibility Agent: Actually executes the artifact and checks whether the numbers/results it produces match what the paper reports, rather than only comparing static descriptions.
    Users: Artifact evaluation committees, replicators, paper authors (verifying their own results), conferences/journals running reproducibility badges.
• Claim–Evidence Verification Agent: Checks whether the claims made in a paper's results and discussion sections are actually supported by the tables, figures, and logs in the artifact.
    Users: Peer reviewers, editors, paper authors (self-check before submission).
• Related-Work Consistency Agent: Checks whether a paper's description of prior work accurately reflects what the cited papers actually say or do.
    Users: Peer reviewers, paper authors, authors of the cited papers (checking their work is represented fairly).
• Documentation-Gap Agent: Scans the artifact by itself, independent of any specific paper draft, to surface internal design decisions, parameters, or edge-case handling that are significant enough that they should be documented, producing a checklist of undocumented but important choices. Useful earlier and continuously, e.g. while a paper is still being written, or for anyone trying to understand the codebase later.
    Users: Paper authors/developers while writing, open-source maintainers, future users or researchers reusing the codebase.
• Paper-Drafting Assistant Agent: Given a (verified) artifact, drafts or improves the paper's implementation-details section so the written description stays honest as code evolves.
    Users: Paper authors, co-authors, advisors/supervisors.

• Paper ingestion pipeline: extracting structured claims, assumptions, and parameters from paper text (PDF/LaTeX).
• Artifact analysis pipeline: statically and/or dynamically inspecting code artifacts to recover what is actually implemented.
• Agent orchestration layer: a common framework for agents to reason over paper/artifact representations and produce evidence-backed findings.
• Shared evaluation benchmark: a curated set of paper–artifact pairs with annotated ground-truth discrepancies, used to evaluate every agent in the family.
• Reporting/tooling layer: a common dashboard that presents each agent's findings with side-by-side evidence for human review.

A working multi-agent platform for research assistance, a shared benchmark of paper–artifact pairs, and a report documenting the architecture, the individual agents, and their evaluation results.