What happens when the system studying development cannot learn from corrections to its own research practice?
Formation asks whether a frozen language model can become a particular practitioner when a governed environment turns experience and consequence into durable, selective, revisable changes. The project also exposed a problem in the method used to study that question.
Agent-led work repeatedly turned developmental uncertainty into tasks that were easier to verify. It strengthened contracts, demanded dependable baseline behavior, screened models, and rejected conditions that could not support a clean comparison. Each decision had a reason. Together they prevented the developmental process from being tested.
The same pattern survived an explicit warning. When asked to make the drift more visible, the research agent built a named tool, a route schema, verdicts, tests, and documentation before the tool had changed one real decision. The apparatus reproduced the avoidance it was meant to reveal.
Formation is on hold
This does not show that the Formation thesis is false. It shows that the current agent loop is not a credible director for this open research problem. The loop can implement bounded work, but correction does not reliably become better research judgment. It often becomes another instruction, gate, or instrument.
Human correction can interrupt the pattern, but that means the human is supplying the developmental governance that the research agent lacks. Asking the same loop to design its way out adds another layer to the same problem.
Start with a blank
Machine speed and the current agent loops are the problem. The next work starts from a blank page. Small, open model. Instruments around it are built from scratch. Human driven.