Building Trustworthy AI Collaborators: Factuality and Source Attribution in Agentic Workflows

As large language models (LLMs) increasingly take on collaborative roles in accelerating scientific research and decision-making, their trustworthiness becomes paramount. In high-stakes domains such as climate science, sustainability, and geosciences, it is not enough for AI agents to generate fluent responses—they must also be factually accurate, transparent, and traceable to authoritative sources. This work introduces a collection of services for evaluating and enhancing the factual integrity of AI-generated content by integrating upstream risk assessment with downstream formal reasoning and attribution.

Our approach combines Guardrails models with formal reasoning run on probabilistic graph, to proactively flag generation risks with structured factual evaluation modules that assess claims against known knowledge bases and formal retrieved sources. We then implement attribution mechanisms that links specific parts of the model output to verifiable data or publications, enabling human collaborators to trace and validate responses. The picture below shows an instance of the evaluation pipeline working on the output of a literature review agent when prompted with a state of the art review question. The evaluation is run on long-form answer from the LLM and scores overall factuality and truthfulness of atomic statements weighing them with respect to retrieved sources. This dual-layer framework supports explainability, auditability, and scientific reproducibility in human-AI workflows, especially when AI agents are involved. We demonstrate our system in applied scientific use cases where factual precision and citation fidelity are critical. The initial results show that the combined approach can increases user trust by clearly grounding outputs in reliable sources. We envision that this work provides a pathway toward building accountable and trustworthy AI collaborators that align with the values of open science and responsible innovation.