Tree B: periscopes. Summary in plain language
The question
Levin looks for hidden intelligence in a simple way. He picks a system, decides what counts as its goal, puts obstacles in the way, and watches how cleverly it gets around them. He uses this on biological constructs that have never existed before, so nothing could have been selected for the cleverness. We use the same method on software that has never been trained: random networks, cellular automata, artificial-life soups and untrained language-model skeletons.
What an experiment looks like
We build one test bench that any of these systems can be plugged into. For each system we write down in advance its goal, the obstacles we will add, and the result that would surprise us. Someone who has not run the system writes this, so we cannot invent the surprise after seeing the data. Each system then runs many times with different random seeds, next to a shuffled copy of itself as a control. We score every run twice. One score uses the goal we chose. The other uses whatever the system's own rules are actually pushing it toward.
What the outcomes could be
- Nothing survives the controls. Then the "unexpected competencies" Levin reports in minimal systems come from how they were scored or scheduled. That gives one paper, plus a test bench others can use.
- One or more effects survive. Some system, under obstacles, does something its rules do not obviously predict, on any measure and in any turn order, and its shuffled control does not. That would be the programme's first real finding. It goes to critics for replication before we say anything about it.
- A mixed result. Effects survive in one class of system (say, continuous artificial life) and not in the others. We would then have a map, and spine/01's per-checkpoint accounting (measure M5, the part of Tree A's ledger that survived) could use it to decide where to look.
What kind of research this is
Artificial-life and dynamical-systems experiments, run with the pre-registration habits psychology adopted. The runs are cheap, deterministic and repeatable. The hard part is writing the surprise criteria honestly.
Timeline
Tree A built the sorting bench and the barrier harness on 2026-09-30, so work can start as soon as this line is scheduled. Working through the zoo of systems takes another quarter.