We asked AI models to break it.
Before opening the framework up, we put the design through several rounds of adversarial review with capable AI models, each asked to find the holes. Here is what they found, and how we answered.
Three independent AI reviews, one risk list.
We ran the design past three capable AI models — across several rounds, each prompted to try to break it rather than approve it. The useful surprise was that all three converged on the same short list of risks — and that list turned out to be the set of hypotheses that feed into our first pre-registration. The framework knew where its own weak points were.
After we folded in their sharpest suggestions, all three reviewed the revised plan and reached the same verdict: a sound, honest, registrable research programme — pre-data, but ready to register and open up. One of them put it well — the plan had become a project “willing to discover it might be wrong.”
The holes they found, and our answers.
“The structure is a beautifully-reasoned map, not yet a discovered one. Several of the distinctions may be real to experience but not separable in a real population.”
Our answerAgreed — and that was already the central thing the plan sets out to test. The weights are a hypothesis, not a position to defend. We committed in writing to merging, relocating, and dropping archetypes substantially if the data says so.
“If a factor analysis recovers fewer than eleven groups, you'll just call that 'expected' — which makes the model impossible to falsify.”
Our answerFair. So we stated exactly what would falsify it: not merely 'fewer groups', but a specific predicted arrangement — which archetypes sit next to which, where the dividing lines fall. If the empirical map doesn't match the one we drew in advance, the theory is wrong, and we report it as wrong.
“Recognition — 'people see themselves in it' — is the easiest thing in the world to fake. Everyone nods at a flattering description.”
Our answerThis is the sharpest risk, so it gets the most validation effort. Recognition is measured against a flattering generic control: a real description has to be recognised better than the generic one, and strangers have to sort a person to the right archetype better than chance. If a pair fails that, it merges — no exceptions.
“Name the finding that makes an archetype actually disappear. If you can't, this is taxonomy-preservation, not research.”
Our answerWe named three: relocate (it belongs to another pillar), merge (two roles, one underlying thing), and drop (it fails on its own terms and adds nothing beyond the others). Each has a condition set in advance.
“Does an archetype survive once you control for the Big Five personality traits?”
Our answerA friendly correction to the question itself: the Big Five is one yardstick, not the substrate of reality. We test against the whole measurable picture — traits, sensitivity, attachment, life outcomes. A distinction that looks like a personality twin on the Big Five but differs on sensitivity or outcomes has found its footing somewhere the Big Five can't see, and we keep it.
“The quick three-item version is thin. A few items can't carry a confident result.”
Our answerTrue, and it's meant to be a fast sketch, not the last word — every result carries a note that it sharpens as you go deeper. We removed the items that muddied the picture, which the AI reviews credited as the right structural call.
What this kind of review is, and isn't.
These were AI reviews, not peer review. They are a fast, demanding, adversarial first pass — capable AI models trying to break the reasoning — and they are not a substitute for expert human review or, above all, for data. They tell us the design holds together and is honest about its gaps. They cannot tell us the framework is true. Only the staged validation can start to do that, and it has barely begun.
We’re publishing the reviews because the transparency is the point: the design and the adversarial assessment of it, both in the open, holes included.