For people who use Quilt-powered tools but don't write code
You're using a tool built on Quilt, a system that uses three kinds of AI together:
Most AI tools you use today are LLM-only. Quilt tools use all three. That means fewer wrong answers, fewer made-up facts, and clearer "I don't know" when the AI is uncertain.
Every Quilt app keeps a witness log — a permanent record of important decisions. We log:
We do not log your private inputs unless they directly affect an answer you asked for.
You can ask at any time: "Show me everything you've remembered about me."
Every AI answer is signed by the witness log. If something breaks, you can file a witness: "This answer was wrong because…" The next version of the model will be trained on that correction.
The witness log is permanent and auditable. You cannot be gaslit about what the AI told you — the log shows exactly what it said, when, and with what confidence.
JEV is a hosted service from typesafe.ai. We send only the question + a short state. We do not send your full history unless you ask us to.
You can delete your cells at any time. "Forget me." The system will remove every cell that originated from you.
One model is brittle. Two is rigid. Three is a team.
JEV alone can't write a poem. LLM alone might hallucinate. JEPA alone can't explain. Together: JEV decides, LLM explains, JEPA confirms. You get the speed of intuition, the rigor of decision, and the clarity of language.
Embeddings are the muscle-memory layer: they remember the shape of past decisions so you don't repeat mistakes. "Was this email a newsletter last time? What did we do with it?" — embeddings find that pattern by trajectory, not keyword.