Tree B: direction questions

Q1 open Harness scope creep

With eight systems of different state shapes, it will be tempting to add per-system special cases to the harness. The rule is that any system needing one gets its own adapter file, and the harness itself stays under 200 lines.

Q2 open Who writes the surprise criteria

Someone who has not run the system must write its criteria. With one agent doing all the work, that means writing the criteria for the whole zoo before running any system, and time-stamping them in the OSF pre-registration.

Q3 open Untrained transformer size

The assay needs a GPU once the model is anywhere near the size of a real one. By default we run a 4-layer, 64-dim model on CPU and note the ceiling.

None blocking as of 2026-09-30.