[ autoresearch ]

The loop

A loop for learning how an AI system actually behaves.

Recovering one function by hand is interpretability. Doing it in a loop, on its own, is autoresearch.

The loop has four parts. System identification is the engine: probe a target cheaply (read the weights, isolate a part, ablate in context) and propose candidate edges. Adversarial validation is the feedback: try to refute each edge before trusting it. The causal graph is the memory: keep only the surviving, boundary-scoped edges. A controller picks the cheapest next experiment to shrink what is still unexplained.

the ringengine · feedback · memory · controller

system identification  ->  adversarial validation  ->  causal graph
        ^                                                   |
        +------------------  experiment selection  <--------+

In control terms it is a self-tuning estimator: a plant (the target), an estimator (system identification), a residual test (adversarial validation), a state estimate (the causal graph), and an experiment-selecting controller. The difference is that the state it estimates is understanding of the target, and the goal is to reduce uncertainty rather than hit a setpoint.

Two choices make it research and not a chatbot loop: the feedback is adversarial (refute, not reward), and the memory is a scoped causal graph (not a flat log).