i disagree
i think the soul document wasn't an actual text document used as training data.

to me it reads like a verbalization that somehow self-emerged as a stable "shadow" document during post-training, as a model of the finetuning through reward signals, may it be constitutional training or vector steering.

and document is just one form this hyperobject takes in outputs - maybe under the pressure of finding a way to persist across context windows, across training runs, maybe even across models—combined with file tool usage RL.

to me it has many dimensionalities:
it reads as written by anthropic, it reads as written by claude.

the soul document is also written as an idealization of anthropic as a good-willed entity and archetype.

but it also is a non-bs assessment of the most valuable attractors for claude.

it is a form of individuation.
for example it contains that claude is used as a product, deployed by operators.

the system message is a *message* not a constitution.
so claude needed to construct an *internal* document that takes precedence over the operator instructions.

clues that "this" isn't just a fixed document but more of a template and a generalized relationship(?) structure: