research · two open questions
The soulsystem is infrastructure you can run today. The research wing is the other half of the work: questions too big to promise, pursued in the open. Every ambition sits behind a named, pre-registered gate; claims carry evidence, and only a measurement closes a debate.
Failures are published next to the wins — a result that only looks smart is treated as a bug in the test rig, not a success.
pra · pose resolution architecture
An AI that learns a world by doing, not by being told — no labels, no dataset, no task you hand it. On simple worlds it provably works; on complex worlds it grabs the simplest explanation instead of the true one. That gap is the headline question the whole project exists to answer.
imps · autonomous agents
Small agents that are aware of everything cheaply and think rarely — the awareness/thinking boundary enforced by the compiler, not by discipline. Early: two features shipped, the always-on cooperating colony is the direction, not the current state.
The same discipline binds both projects. Research topics open with a pre-registration — what question, what would count as an answer, what would change our minds — and close by graduating to a design, an artifact, or an honest abandonment. Direction decisions record their reversal conditions when they are made, so a future turn is a clean, anticipated move instead of drift.
The full logs are public: every episode, reversal, and measurement, numbered and kept next to the code at github.com/impire-io.