0
ML Weft
Adaptive models within bounds you can check
ML Weft is an open research program from Moissanite Labs. It studies models that keep adapting after deployment, under limits that a program can verify.
Papers
The ML Weft series
One overview, then a paper for each part, each with its evidence labelled.
Every paper names the repository commit it describes and includes the commands that reproduce its results. Each one is available as a web page and as a PDF.
1
Models as Cells
Self-modification bounded by distance from a saved baseline
A model that keeps adapting after deployment needs a limit on how far it can drift, and the limit has to be something a program can check. ML Weft treats each model as a cell.
2
Checkable Computation
Interaction combinators, a binary graph format and committed execution records
A system that changes itself should be able to show what it computed. ML Weft approaches this with three small pieces: a reducer for interaction combinators, a graph-rewriting model in which every step is a local rewrite of one pair of nodes; a canonical binary format for the graphs; and an execution record that commits to the initial graph, every step and the final graph with SHA-256.
3
Eido
Style as a point in space
4
Constitution vs. Contract
Provenance, metering and governance for adaptive models