Research
ChelatedAI research
ChelatedAI studies how retrieval and memory systems can adapt to new data and new models without silently corrupting the evidence they already hold. It informs how we build retrieval, evaluation, and custom models for clients.
Research areas
Adaptive retrieval and drift
What happens to a search index when the embedding model changes, and how to migrate safely.
Embedding correction
Spectral chelation and dimension masking to counter semantic collapse in retrieval systems.
Reversible evidence pools
The Liquified Lattice architecture: derived links and adapters that can be explored, pruned, and rolled back.
Evidence kernels
Testable questions about provenance, intervention rank, and what a probe can and cannot discover.
Computational storage
Disk-backed retrieval and near-data candidate evaluation, with hardware claims kept in scope.
Publications and resources
Hosted on mattmre.com. Papers are public preprints and are not peer reviewed.
Research archive
The ChelatedAI lab notebook: adaptive retrieval, graph memory, and storage-assisted computation, each with a stated evidence boundary.
Read on mattmre.com ChelatedAIPapers
Versioned public preprints with methods, limitations, and falsifiers. Not peer reviewed, and labeled that way.
Read on mattmre.com ChelatedAIFindings ledger
What held up and what did not, with each claim kept separate from the test that supports it.
Read on mattmre.com ChelatedAIExplainer atlas
Interactive explainers for the equations, controls, and failure tests behind each paper.
Read on mattmre.com ChelatedAIResonant holographic path compression
An active hypothesis on reusing frequent model routes as rotated low-dimensional slices, with a lossless manifest kept alongside.
Read on mattmre.com ChelatedAIResearch log
A dated record of milestones, null results, and decisions not to promote a result.
Read on mattmre.comSource code: github.com/mattmre/CHELATEDAI
How we report results
A test that runs correctly is not the same as a method that works. Every result is labeled as confirmed within a stated scope, unconfirmed, blocked, or negative, and negative results stay published. We apply the same standard to client work.
Working on retrieval or model quality?
If your search or AI results degrade as data and models change, we would like to hear about it.