Docs

Catalyst Docs

Documentation on iterative AI, continual learning agents, memory, retrieval, judgment profiles, feedback loops and evaluations.

Catalyst asks one question:

Can an AI system improve because you used it yesterday, without hiding what changed?

Remembering an old message is not enough. We need to know what evidence the system saw, what rule it proposed, where the rule applies, whether the next result improved and how to undo it.

These notes explain what we're trying, what we think, and what is still unproven.

Start here#

• Purpose
• What we believe about AI
• What we are building
• Iterative AI
• Terms
• Projections

The working loop#

correction or outcome
-> evidence with a source
-> narrow proposed rule
-> review
-> guidance for a matching task
-> comparison
-> keep, limit, or roll back

A correction is not automatically learning. It only counts when the system can show the evidence, the changed behavior, the evaluation and the rollback path.

Current work#

Meraki is the product people see.

Meraki Core is the engine trying to make the loop real. The repository contains the public technical docs and the current proof audit.

This is still early. The local engine has automated proof. The hosted invited-user loop and independent human evaluation are unfinished.