Orchestrate Scientific Integrity Agent Crew
With growing concerns about 'AI slop' in scientific publishing, this challenge focuses on developing an agentic system to enforce scientific integrity. You will use CrewAI to orchestrate a team of specialized AI agents that act as a 'Scientific Review Board.' This crew will collaborate to analyze newly generated scientific abstracts or summaries, identify potential factual inaccuracies, inconsistencies, and characteristics of AI-generated content, and verify claims against a knowledge base. The system should highlight suspicious areas and provide justifications for its findings, leveraging the advanced reasoning capabilities of Claude Opus 4.1.
What you are building
The core problem, expected build, and operating context for this challenge.
With growing concerns about 'AI slop' in scientific publishing, this challenge focuses on developing an agentic system to enforce scientific integrity. You will use CrewAI to orchestrate a team of specialized AI agents that act as a 'Scientific Review Board.' This crew will collaborate to analyze newly generated scientific abstracts or summaries, identify potential factual inaccuracies, inconsistencies, and characteristics of AI-generated content, and verify claims against a knowledge base. The system should highlight suspicious areas and provide justifications for its findings, leveraging the advanced reasoning capabilities of Claude Opus 4.1.
Shared data for this challenge
Review public datasets and any private uploads tied to your build.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
Detect All Known Errors
The crew identifies all pre-defined factual errors.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Justification Quality
Each flagged issue has a clear and relevant justification.
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Accuracy of AI Slop Detection
The percentage of correct 'AI slop' indicators identified. • target: 0.85 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
Review Consensus Score
A measure of agreement among agents on critical findings. • target: 0.9 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Master CrewAI's framework for defining roles, goals, and tasks for collaborative AI agents, ensuring clear responsibilities and communication paths.
Implement role-playing agents such as a 'Factual Verifier,' 'Consistency Checker,' and 'AI Slop Detector,' each equipped with specific tools and system prompts.
Integrate Claude Opus 4.1 for the 'AI Slop Detector' and 'Consistency Checker' roles, leveraging its advanced analytical and reasoning capabilities to identify subtle inconsistencies and patterns indicative of AI generation.
Utilize Mistral Saba for the 'Summarizer' agent, to quickly digest and extract key information from scientific texts for initial review by other agents.
Build a tool for the 'Factual Verifier' agent that queries a Pinecone vector database populated with scientific articles and established facts for evidence-based verification.
Design the overall review process within CrewAI, specifying the sequence of tasks, agent hand-offs, and criteria for collaborative decision-making.
Develop a robust output mechanism that provides a summary of findings, specific flagged issues, and justifications from the contributing agents, possibly integrated with DeepOpinion for workflow automation of the publishing feedback loop.
How this agent runs
The evaluation will assess the CrewAI system's ability to accurately identify factual errors, logical inconsistencies, and 'AI slop' characteristics in generated scientific text, providing clear justifications.
Challenge input
{'abstract': '...', 'known_errors': [{'phrase': '...', 'reason': '...'}]}
Agent execution
The configured agent processes the input under the challenge policy.
Evaluated output
{'flagged_issues': [{'phrase': '...', 'finding': '...', 'justification': '...'}], 'overall_verdict': 'accurate|inaccurate'}
- The crew identifies all pre-defined factual errors.
- Each flagged issue has a clear and relevant justification.
- The percentage of correct 'AI slop' indicators identified. • target: 0.85 • range: 0-1
- Accuracy Of AI Slop Detection target: 0.85
- 2 public reference cases
- Python execution harness
View technical recipe
Configured tools
No tool records are attached.
Evaluation contract
- Detect All Known Errors · Weight 1
- Justification Quality · Weight 1
- Accuracy of AI Slop Detection · Weight 1
- Review Consensus Score · Weight 1
Recipe state
This is a preview. The configuration can change before the evaluation recipe is locked.
Run this agent on your dataset
Versalist can run this agent on your behalf with your data. Tell us about your dataset and the result you need.
Discuss your dataset[ok] Wrote CHALLENGE.md
[ok] Wrote .versalist.json
[ok] Wrote eval/examples.json
Requires VERSALIST_API_KEY. Works with any MCP-aware editor.
DocsFind another challenge
Jump to a random challenge when you want a fresh benchmark or a different problem space.