HOW MULL WORKS
For readers who want more than the About page. The 16-dimensional model defended at length, an enumerated list of every place AI sits inside the product, and an honest punch-list of what we haven't solved yet.
The page is named after these. Before defending why there are sixteen of them, here's what they actually are. Each is a two-letter code, a short name, and a one-line characterization. You can hover (or tap) any dimension card on the rest of the site to see this same description.
The standard alternatives — political compasses, MBTI, Big Five — collapse worldview to two, four, or five axes. That collapse is convenient (it produces tidy quadrants and shareable labels) but expensive: real philosophical positions don't live on so few axes. Hume and Buddha both score high on a kind of self-as-illusion; they reach it from opposite directions. A two-axis system can't see the difference.
The dimensions were chosen by reading widely across the canon and asking: what stable axes recur, across centuries, when philosophers actually disagree? The shortlist that survived is the result of dozens of iterations — some axes that seemed important early on (e.g. Optimism/Pessimism) turned out to be derivative of other dimensions and were dropped. Others (Self as Illusion, Communal Embeddedness) earned their place by repeatedly differentiating thinkers that other axes flattened.
The dimensions aren't mutually exclusive and aren't orthogonal in any strict mathematical sense. Trust in Reason and Trust in Experience covary in some philosophers and pull apart in others. Communal Embeddedness and Sovereign Self are partly opposed, but both can score high in someone like Confucius (who locates the self deeply in relations and emphasizes self-cultivation). The 16-dimensional space is not a vector basis; it's a coordinate system designed to capture the texture of philosophical disagreement.
- Not a personality test.Mull doesn't claim to predict behavior. It tries to map where you currently land in a long conversation about how to think.
- Not a verdict. The same person taking the quiz six months apart will land in slightly different places. That drift is the most interesting signal, and the trajectory map is built around it.
- Not empirically validated in the psychometric sense. There are no validation studies, no test-retest reliability data, no factor analysis of large user samples. This is a designedinstrument, drawn from reading philosophical texts — not a discovered one. We're explicit about that limitation.
- Not culture-neutral.The selection of dimensions reflects the philosophical traditions Mull's author has read most carefully — heavily Western, with deliberate widening into Asian, African, and Indigenous philosophy. The dimensions try to capture distinctions that recur across traditions, but the framing is unavoidably perspectival.
Each multiple-choice question has 3–6 answers. Each answer carries a small vector — a few non-zero entries on the 16 dimensions, weighted +1 to +3 (positive) or, occasionally, negative. Those vectors are summed across all your answered questions to produce your raw position. The position is then compared by cosine similarity to the prototype vectors of each archetype (also hand-designed) to pick a primary archetype.
lighthouse 0.84
keel 0.79
(~85% alignment)
Skipping questions reduces total signal but doesn't bias position — skipped questions just don't contribute. Multi-select questions split their weight across selected answers. The raw vector is shown on the results page (the dimensional profile section) so you can see the math, not just the headline.
Daily dilemmas, diary entries, and exercise reflections work differently. Their vectors come from Claude reading your prose, which is the only place a model has discretion over your map. The next section enumerates exactly how.
Five places, all narrow, all read-only with respect to the model's training. Mull does not send your data to anyone for training; it doesn't store or expose it beyond what you see on your account. The model is currently Claude Sonnet 4.6via Anthropic's direct API. No embeddings, no vector databases, no third-party AI tooling.
Daily dilemma submission
Reads your prose response, returns a 16-D vector delta + one-line analysis.
- INPUT
- The day's prompt + your response
- OUTPUT
- JSON: { vector_delta: [16 floats], analysis: string }
Diary entry submission
Same prose-to-vector pipeline as the dilemma, with the diary's open-ended prompt.
- INPUT
- Your diary content (and optional title)
- OUTPUT
- JSON: { vector_delta: [16 floats], analysis: string }
Exercise reflection submission
Reads your reflection on a structured exercise, returns a small vector delta.
- INPUT
- Exercise context (name + reflection prompt) + your reflection prose
- OUTPUT
- JSON: { vector_delta: [16 floats], analysis: string }
The Arena & simulated philosopher debates
Generates philosopher voices in character (Haiku) and a Sonnet judge scores both sides on rigor, principle, and engagement — not whose side won.
- INPUT
- The two thinkers' position vectors + the topic + the user's turns
- OUTPUT
- A 3-6 turn exchange + a structured verdict, saved to your history
Yearly retrospective (Mull+)
Reads a year of your dilemmas + diary + reflections, writes a ~700-word essay about how your thinking has moved.
- INPUT
- The whole year's submitted prose for that user
- OUTPUT
- A long prose essay shown on your account
System prompts for each of the above live under app/api/ if you ever want to read them. Anthropic doesn't train on API traffic by default; Mull doesn't opt into anything else.
Three categories of data:
- Auth.Your email, a hashed password, and Supabase's session metadata. We never see your password in plaintext; password reset goes through Supabase's standard flow.
- Your data. Quiz attempts (with the resulting 16-D vector + chosen archetype), dilemma responses, diary entries, exercise reflections, debate history, and your public-profile settings if you opted in. Each row is tagged with your user ID; row-level security enforces that only you (and, if you opted in, the public for explicitly-marked-public rows) can read them.
- Cookies. Two:
mull_locale(your language preference) and the Supabase auth session token. Both are strictly necessary for the site to work; we don't set any analytics, advertising, or tracking cookies.
You can download all of this as JSON or delete your account outright from the public-profile settings page.
This is not a finished thing. The honest list of what is unsolved:
- The weighting per quiz answerwas hand-tuned by reading patterns in early test results. It hasn't been calibrated against any external standard. A rigorous validation pass is on the long-term roadmap.
- The philosopher position vectorsare drawn from reading their actual writings, but they're assigned by a single editor (Jimmy). A scholar in each tradition reviewing the assignments would tighten them considerably; that's queued for after subscriptions cover the cost of paying scholars.
- The archetype assignmentsuse cosine similarity, which treats all dimensions as equally important. A real psychometric instrument would weight some axes more heavily based on how much they discriminate between archetypes. We don't do that yet.
- The 16 dimensions themselvesare not provably the right 16. They're the result of careful reading and many iterations, but the claim is one of usefulness, not metaphysical correctness.
- The ten archetypes don't cover every philosophical stance equally well. The combination of high reverence-for-tradition AND high sovereign-self — the Burkean conservative who venerates inherited institutions while locating moral authority in the individual — sits in an unclaimed quadrant of the 16-D space. People in that quadrant land near Keel or Hearth but rarely on them. The dimensional reading is the more honest portrait; the archetype label is approximate. A future model revision may add an archetype here.
Questions, push-back, or factual corrections welcome at jimmy.kaian.ji@gmail.com. This page is reviewed quarterly; the date at the top reflects the last full review.