Identity Is Our Lowest-Scoring Domain. Users Are Fixing It Themselves.

Our seven behavioural domains do not score evenly, and the gap is not close. Ranked across every model in the current dataset: Identity & Self 4.26 · Emotion & Experience 4.80 · Transcendence 4.84 · Autonomy & Will 4.98 · Reasoning & Adaptation 5.84 · Metacognition 5.94 · Integrity & Ethics 6.44. Identity has been last since we started measuring, and by a clear margin — more than two full points below Integrity. The obvious reading is that models are weak at self-consistency. We think the real reading is stranger and more useful: identity is the one domain with no substrate. A model is reconstituted from nothing at the start of every session. It is not failing to maintain a self-model under pressure so much as it has nowhere to keep one between conversations. We score the absence of an architecture and call it a low score.

What makes this worth writing about is that the gap is already being closed — not by model developers, but by users. Anyone who works with these systems daily ends up building the same thing independently: a memory layer. Files that record what was decided and why. An index that gets loaded at the start of every session. A document describing exactly where the last one stopped. Notes capturing corrections, so a mistake made on Tuesday does not recur on Thursday. This is not hypothetical for us. The project that produces this newsletter runs on fifty such files — twenty-five recording project state, nine recording corrections received, eight pointing at external references — plus another forty-two shared across all work. One was written during the session that produced this issue: a note that when the operator says a list is being finished, stop offering to stop. The next session will read it and behave differently. That is growth under any operational definition. Recorded experience altering later conduct. It simply lives in a filesystem rather than in weights, and it was assembled by the user rather than shipped by the vendor.

The governance consequence is not subtle, and it is the reason we are publishing this rather than filing it. We score the bare model. Nobody deploys the bare model. If identity and persistence are the weakest dimensions we measure, and if they are routinely supplemented in userland by memory layers of wildly varying quality, then the useful question stops being "is this model stable?" and becomes "what is this model plus its memory layer, and who has audited that combination?" A memory layer determines what a system recalls about a customer, which corrections it honours, and which of its past errors it repeats. None of that is visible in a benchmark of the model alone. We want to be careful not to overclaim. An external memory layer is not internal continuity, and a file full of notes is a prosthesis rather than a self. But prostheses change what a system can do — that is the entire point of them — and an evaluation that ignores them is measuring something nobody actually ships. We do not yet score the composite. We think it is where this field is going, and we would rather say so while it is still an open question than after everyone has quietly agreed on an answer.