Scientific Novelty & Evidence Agent
An expert scientific evidence agent that discovers literature, extracts study-level evidence from accessible full text, maps claims to sources and preserves uncertainty for human review.
What it is designed to do
- Map evidence relevant to a scientific or technological question
- Structure claim-to-source traceability
- Identify research gaps and unresolved questions
- Support novelty assessment before expert review
- Generate a structured evidence report for human evaluation
What it receives and returns
- Research question, hypothesis or technical claim
- Optional manuscript, abstract or technical description
- Optional domain constraints, dates, keywords and source preferences
- Provisional source-bounded scientific answer
- Evidence map with DOI or stable source identifiers
- Study-level extracted findings, conditions and methods when accessible
- Claim-to-source matrix
- Research-gap and insufficient-evidence flags
- Traceable structured report
Evidence before confidence.
Material claims should be linked to inspectable supporting sources when source access is available. The agent should flag insufficient evidence rather than convert uncertainty into a definitive conclusion.
Human review of material sources remains required for final scientific judgment.
Evaluation is published only when it is reproducible.
20 / 20 technical regression · scientific score pending
Technical regression suite: 20/20 passing for the current release candidate. This is not a scientific certification. The human-reviewed Gold Scientific Benchmark remains pending before a public scientific performance score is published.
Limitations
- Does not establish legal patentability or freedom to operate
- Does not replace peer review or subject-matter expert judgment
- Does not guarantee exhaustive coverage of all literature or databases
- May inherit limitations, access restrictions or errors from external sources and model providers
Human gate
Final novelty, publication, patentability or research-priority conclusions must be made or approved by a qualified human reviewer.
AX-SCI-001 is not yet a public one-click execution.
We are keeping the first release controlled while the public benchmark, execution profile and reproducible evaluation record are completed.