An unfamiliar machine can still be understood one transition at a time. A state
is not a story about the past; it is the part of the past that can change what
happens next. When a model copies a sequence, each emitted symbol is both an
answer and a new input. The resulting trajectory can remain stable, correct a
small error, or fall into a repeating orbit.

This reference text is deliberately ordinary rather than randomly sampled. The
training prompts contain token identifiers drawn uniformly with replacement, so
natural prose provides a useful test of whether copying extends to structured
token sequences. Replace this file with a larger corpus, such as held-out
FineWeb text, when reproducing the original validation regime.

Three memories divide the work. The prompt memory stops changing after the
first boundary. The reasoning memory records the generated trace. The final
memory begins only after the second boundary. Every layer may read the memories
that already exist, while only the bank belonging to the current phase is
updated. One global rotary clock advances through all three regions.

