Allaun Silverfox
|
29752fb145
|
feat(dag): Resumable DAG with manifold coordinate transforms
Complete model for chunked NP-hard solving with Fisher manifold
coordinate transforms between exploration chunks.
Key innovation (not divide-and-conquer, not branch-and-bound):
1. Wind up: start computation chunk
2. Run: evaluate subset S_k until limit
3. Pause: save checkpoint (distribution, Fisher matrix, best energy)
4. Transform: compute eigenstructure of Fisher matrix, rotate coords
5. Resume: restart from uniform in NEW manifold coordinates
6. Repeat: build DAG of checkpoints
SilverSight integration:
- ChunkLib: evaluate, eigenstructure, transform, resume
- MetricLib: Fisher matrix computation from partial results
- DAG state is the resumable checkpoint (serialize → resume anywhere)
- Each chunk produces a Receipt with parent link (DAG edge)
Scaling: n=40, 1K parallel branches → ~100s exact (vs 10^12x brute-force)
Refs: ChentsovFinite.lean (metric uniqueness), Fisher information geometry
|
2026-06-23 00:54:24 -05:00 |
|