Recursive Self-Improvement infrastructure for agent fleets.
Five layers, one closed improvement loop:
- Bayesian Confidence — Beta-Bernoulli posteriors with adaptive forgetting
- PID Controller — Two-level cascade (resource + cognition), gains driven by learning rate
- Multi-Shell Compression — Chord model ratio, PID-tuned by cognitive load
- LOG-Tensor Cycle Closure — T(i→j)∘T(j→k)∘T(k→i) = I, adaptive tolerance
- Attractor Dynamics — Potential landscape reshaped by cycle error
Every pass through all five layers is one Rotation. The Rotation improves the improvement function.
- gc-pid-bridge → Level 2 actuator
- headspace → absorb/evolve/synthesize mapping
- baton-system → carries confidence + compression
| System | Role | RSI Layer |
|---|---|---|
| the-rotation | Meta-controller | All 5 |
| gc-pid-bridge | Resource PID | 2 |
| headspace | State compression | 1, 3 |
| baton-system | Fleet state | 3, 4 |
| pincher | Reflex runtime | 1, 2 |
MIT