Core implementation is currently private while patent and licensing options are evaluated. Clair is a solo-built correction-first cognitive agent prototype using three core loops:
- Reasoning: solves tasks and forms candidate answers
- Calibration: checks truth, confidence, and evidence before output or storage
- Maintenance: audits memory, reduces drift, and preserves system health
The goal is to build an inspectable agent architecture that can improve under benchmark pressure without poisoning its own memory.
Experimental prototype. Not production-ready.
- Behavior: 50/50
- Answer quality: 14/50
- Average score: 0.28
- Generic responses: 0
- Unrelated memory errors: 0
- Needs tool/document support: 33
Most agent systems focus on acting and remembering. Clair focuses on correction before memory.
The system is designed around:
- strict module boundaries
- verification-based storage
- correction dominance
- memory governance
- failure-class testing
- benchmark-driven improvement
[Insert diagram here]
- Intake
- Perception
- Reasoning
- Calibration
- Planning
- Action Selection
- Execution
- Evaluation
- Reflection
- Memory
- Maintenance
- Add real PDF extraction
- Add real XLSX extraction
- Improve document support solver
- Reduce tool/document failures
- Push GAIA-style answer quality from 28% toward 55%+