The question is not simply whether an AI system can learn again. It is whether it can preserve provenance, recognise pressure, adapt structure, and remain accountable while doing so.
Active research programme · Evidence evolving
Continual learning.Without starting over.
CLEO explores how intelligent systems can retain useful context, adapt across time, and remain governed as their capabilities evolve.
Research architecture →01
Research direction
A system that changes without losing itscoordinates.
02
Evidence policy
Claims needcoordinates.
Every material result should lead back to a method, source, status and accountable owner. Concepts, prototypes, benchmarks and deployed capabilities must remain visibly distinct.
Research records
Architecture becomes
accountable evidence.
- TRSarchitecture · needs confirmation
Temporal Regulation System
Tracks temporal signal frequency across tasks and informs the system’s internal timing without dependance upon external tool calls or timers; one of the core components for eliminating catastrophic forgetting.
- DRSarchitecture · needs confirmation
Dynamic Routing System
Explores adaptive pathfinding and bounded learning pressure in multi-dimensional modality.
- NASHarchitecture · needs confirmation
Equilibrium Output Gateway
An optional internal ability to explore output consensus through equilibrium rather than simple majority vote via internal deliberation where game theory applies; used where human and self-safety and wellbeing are paramount, such as conflict resolution scenarios.
- 04paper · verified
Principle of Persistent Structurization/Scale-Timescale Optimization Corollary
The crown jewel of CLEO: the physical axiom that is a pre-condition for Constructal law flows to appear. The original white paper earned us an invitation to orally present the PPS in Paris at the 16th Constructal Law Conference.
Open source record ↗ - 05benchmark · needs confirmation
CLEO Benchmark Cards
Our internal benchmarks to build CLEO to ensure the physics works. Three benchmarks, back to back without internal resets between them, over a ~5.6 hour period on Kaggle standard data sets with a T4 x2 accelerator. NO EWC. NO replay buffers. NO task ID cheats. Parameter and expert count is dynamic and not hardcoded; it will grow and shrink depending on its current state.
Open source record ↗
The connected story continues
Continue to Michael & Dan→