Generating ideas is easy. Determining if they are new and valuable is the real challenge. The system includes an automated audit module that acts as a preliminary "Peer Reviewer".
Generative AI systems can produce hypotheses that sound plausible but have actually already been explored in the literature. Without external validation, you risk:
Every idea is validated against global scientific literature in real-time
AI2 (Allen Institute)
OurResearch (ex Microsoft Academic)
Three steps for each generated idea
The system builds semantic queries from the proposed idea and actively searches Semantic Scholar and OpenAlex for papers that have already explored that specific combination of concepts.
Example query:
"transfer learning computer vision natural language processing cross-domain application"
A specialized evaluator model (LLM with review prompt) analyzes:
Based on the recovered evidence, the idea is classified into one of three states:
Automatic classification of novelty level based on scientific evidence
Genuinely New
No significant evidence was recovered connecting the key concepts of the proposed idea. It is a genuine research gap with high probability of being unpublished.
Metrics:
novelty_score > 0.8🚀 Recommended action:
✓ High priority for research
✓ Proceed with formal hypothesis development
✓ Design preliminary experiments
✓ Secure resources and funding
Recent Trend
Scattered or very recent evidence found. The topic is beginning to explore, there are recent papers (last 1-2 years) but no consensus or standard solution yet.
Metrics:
novelty_score 0.4 - 0.7⚠️ Recommended action:
✓ Review the found papers in depth
✓ Differentiate the proposal clearly from them
✓ Good for incremental publication (State of the Art + Delta)
Consolidated
The connection is well established in prior literature.
The system provides existing references (prior_art_refs) for consultation.
Metrics:
knownness_score > 0.75🔄 Recommended action:
× Discard or pivot the original idea
✓ Read existing papers (learning)
✓ Look for a completely unexplored angle
✓ Consider novel extensions or variations
Ambiguous
Insufficient or contradictory evidence. The LLM evaluator could not confidently determine if the idea is new or existing.
Possible causes:
🔍 Recommended action:
⚠ Requires expert human eye
⚠ Review recovered evidence manually
⚠ Consult with domain expert
⚠ Refine the idea formulation and re-validate
Specialized evaluator model (e.g., GPT-4o mini) that analyzes abstracts
Judge Verdict
The specific combination of concepts does not appear in the recovered evidence. The analyzed abstracts show no direct interaction between the proposed ideas.
Judge Verdict
The concepts appear directly interacting in the evidence. Multiple papers demonstrate that the connection has already been explored or implemented.
Judge Verdict
The interaction appears only in recent literature. The topic is young and is in active exploration phase by the scientific community.
Range: 0.0 - 1.0
Estimated probability that the idea is unpublished. Calculated from:
Typical threshold: novelty_score > 0.8 → NOVEL_BRIDGE classification
Range: 0.0 - 1.0
Degree of certainty that the idea already exists in the literature. Conceptual inverse of novelty_score.
Typical threshold: knownness_score > 0.75 → KNOWN_LINK classification
Each classification must be backed by evidence to ensure system reliability:
Minimum references
min_evidence_refs = 2 (default)
At least 2 recovered documents are required to make a reliable classification.
Truth sources
• OpenAlex (250M+ works)
• Semantic Scholar (200M+ papers)
• Real-time query
The real impact of automated validation
Of AI-generated ideas without validation are redundant or existing
Faster than exhaustive manual literature review
Savings in research resources avoiding duplication of efforts
How it works under the hood
From the proposed idea, the system generates multiple optimized queries:
Retrieved papers are ranked by relevance and top-K is analyzed:
Model specialized in critical idea review:
Each result includes actionable metadata:
Don't let months of research end in a "this was already done in 2019". Validate automatically before you start.