Plato-Bio: A Biology-Routed Research Agent with Verification-First Design
Plato-Bio has been created by researchers as a biology-centric extension of the open Plato/Denario framework, aimed at prioritizing verification in biological novelty assessments. In contrast to standard large language model research agents that may yield coherent yet unverified results, Plato-Bio integrates explicit workflow states with provenance documentation, citation verifications, claim-to-evidence connections, scoped file writings, and publication checkpoints. An audit of sources uncovered and corrected three issues that could skew evaluations: the default factory's loss of task domain, the exclusion of declared method signals from scoring, and the absence of the drafted-claim denominator in evidence sidecars. The latest clean revision saw the complete Python suite achieve 931 passes, six skips, and zero failures; biology, genomics, evidence/citation, and adversarial-safety suites also succeeded without issues. The system was tested on two specific use cases, including a frozen historical rediscovery task. This research is detailed in arXiv preprint 2607.23975.
Key facts
- Plato-Bio is a biology-routed extension of the Plato/Denario architecture.
- It couples workflow states with provenance records, citation checks, claim-to-evidence links, scoped file writes, and publication gates.
- A source audit found three defects: loss of task domain, omission of method signals, and missing denominator in evidence sidecars.
- The full Python suite passed 931 tests, skipped 6, with no failures or errors.
- Targeted suites for biology, genomics, evidence/citation, and adversarial-safety also passed without failure.
- Evaluation included a frozen historical rediscovery use case.
- The preprint is available on arXiv with ID 2607.23975.
Entities
Institutions
- arXiv