The science

We map how a person's mind tilts across 29 trait dimensions, named for the constructs behind ten of the most-studied questionnaires in psychology — without ever asking you to fill out a questionnaire. (Calibration measures these as 27 families: two bipolar value dimensions each fold a pair of scales; the projection is spelled out on the methodology page.) Two things must be said before anything else. First: every validation result we hold comes from synthetic personas driven by a language model; no human being has been measured, and whether the results transfer to people is unestablished. Second: the estimator that serves your trait scores is the same one the calibration programme measures, in its deliberate uncalibrated configuration — no fitted parameters are mounted. On the one registered run of the current engine (September 2026; 40 synthetic personas, ~46 completed sessions each), that served estimator recovers planted traits at a median correlation of 0.23, reads ten of 27 families well enough to confirm each on its own, and describes the same person consistently across two halves of their play on only a handful of them (median 0.15). We state both boundaries plainly rather than bury them.

What changed in October 2026. The session itself was redesigned to collect more from each play. You now finish each act with how you do it, cross out the option you would never take, and close the story by choosing where your character goes next. Answers typed in your own words are read as options of their own. In a simulation with ideal players — planted personalities that choose exactly as the scoring model assumes — the new session roughly doubles the information each session carries (×2.12), and 26 of 27 families reach the 0.70 consistency bar within fifty sessions, against 14 before. That is a ceiling, not a result about people. The one published measurement we found for story-choice games puts real players at about half that strength, which leaves a quarter of the information; the one language-model persona test of the new session was inconclusive; and the reading of typed answers missed its pre-registered bar and is scored at reduced weight by an explicit decision. Methodology describes the new session; validation carries the numbers and their limits.

This page is the hub. The deeper material lives in seven sub-pages. If you have ten minutes, read the data layers page; if you have an hour, read all seven in order. If you want the measurement standards we hold ourselves to — and an honest account of which of them we currently meet — methodology is the page you want.


The big picture, in three cards

CardWhat it coversRead
The four data layersSignals → Rules → Constructs → Traits — the four-layer measurement architecture SoulMap was designed around, read with two governing corrections: the rule layer does not execute in production, and a served score today traces to your recorded choices and their table-stamped construct weights.9 min
Synthetic personas + scenariosWhy we model how traits cluster realistically — not as independent dice rolls — using a Gaussian copula whose dependence structure is assembled from open empirical datasets, peer-reviewed meta-analytic estimates, statistical bridges, and explicit conservative uncertainty where neither bridge nor data exists.11 min
The calibration programmeThe run record of the current engine: how the confirm run was registered and executed, what it cost, what it lost and why, and what each banked artifact may and may not support.10 min

And three more:

CardWhat it coversRead
MethodologyHow measurement works end to end — the server-dealt evidence shelf, the session as it is played since October, how typed answers are read, the choice model the product actually serves, the seven broad areas, and the testing discipline: pre-registration, a shuffled null, bootstrap intervals, split-half consistency.16 min
ValidationWhat the current engine has demonstrated, at what strength, on what population — one pre-registered run, the October session measured as a ceiling, typed answers below their bar, every number tied to its banked artifact, the misses stated as plainly as the passes.14 min
RoadmapWhere the programme stands after the confirm run: what is being re-verified, what the next run is for, and what real-user validation would take.6 min

Plus an alphabetical citations roster that backs every footnote on every page in this tier.


Why this exists

Standardised questionnaires are the gold-standard measurement tool in personality psychology. They also have well-known limitations: they ask people to evaluate themselves on abstract claims ("I am the life of the party"), they suffer from acquiescence bias and social desirability bias, and they ask people to remember and average behaviour over weeks or months. Behavioural evidence — what someone actually chooses to do in a situation — is richer than questionnaire evidence, but it is also less standardised: different situations elicit different signals; different people reveal different facets of themselves.

Our research programme applies the inferential machinery that makes questionnaires defensible — explicit measurement models, calibrated uncertainty, pre-registered tests — to short interactive sessions instead. That machinery is validated so far only against synthetic personas; the product serves the measured estimator in its uncalibrated configuration, and we say so rather than blur the two.

Behavioural evidence is richer than questionnaire evidence — but it is also less standardised. The science here is in bridging that gap.

What we are NOT trying to do

The boundaries matter as much as the claims:

  • We do not diagnose. We are not a clinical assessment tool. The trait estimates we produce describe normal-range variation in psychological tendencies, always with their uncertainty attached; they are not psychiatric diagnoses, they are not screening instruments, and they should not be used as either.
  • We do not replace clinical assessment. If you are wondering whether a person is at clinical risk for any condition, the right tool is a licensed clinician, not us. Our calibration data is from non-clinical synthetic populations; we have no claim on clinical sensitivity or specificity.
  • Demographic validity is established against real users, not assumed from synthetic data. Every calibration artifact we hold today is synthetic-population-derived. Real-user recalibration with explicit measurement-invariance testing remains future work, and we will not claim demographic validity before it runs.
  • We do not promise individual prediction at clinical-grade reliability. Every reading is served with an explicit uncertainty band and a plain confidence word, and a number appears only once that trait's own measured consistency supports one; today every reading is also stamped provisional, because on our own measurements only one trait family meets the 0.70 bar for a decision-grade individual reading. Four families whose readings ran backwards on the registered run are withheld from the product until their scoring has been re-verified. We treat these as findings to publish, not caveats to bury.

This boundary is documented as the demographic-blind scoring policy: demographic information shapes the synthetic ground truth and the persona's behaviour in scenarios, but it is never an input feature to the scoring path that produces a served trait score. Scoring is global; measurement-invariance testing across demographic subgroups is the discipline that would prove the global model fair, and it has not yet been run — we state that as a gap, not a formality. The personas page carries the full rationale; this remains the strongest guarantee we make to the people who read their own results, and it avoids the high-risk-profiling flags of the EU AI Act.


A short reading order

For a journalist or a public visitor with ten minutes: just the data layers page, plus the "What we are NOT trying to do" box above.

For a technically trained reader who wants to audit the full chain behind their own readings: data layers → personas → methodology → validation. About 40 minutes.

For a journalist or analyst writing about the methodology: data layers → calibration → validation. About 25 minutes. The citations page is the bibliography.

Every artifact these pages cite is in the repository under calibration/phaseA/ — the confirm run under wp26/, the October work under wp28/ to wp34/ — each with the inputs it was computed from.


Where the data flows

The reflective sessions themselves stay private, and nothing in them trains another user's model. Your session records and your profile are stored in the EU; sign-in details are held by Firebase Authentication in the United States. The scenes are written, and typed answers read, by Anthropic's Claude models, which may process requests outside the EU; Google's models, called through a global endpoint that may also process a request outside the EU, make the opening image and the soundtrack, read some typed answers (to compare them with the options and to write the reflection on your profile), write the summaries of your profile, and can be switched in to do Claude's work. These transfers are covered as the privacy policy describes. Your trait estimates are visible in your account; you can export everything as a single JSON file, you can withdraw from the analytics export at any time, and you can delete your account, which cascades deletion across our stores (copies the model providers keep are deleted on their own schedule; an independently verified end-to-end erasure test is on the roadmap before the analytics export resumes). The product-level data controls live in the in-app Settings → Privacy & Data page.

This is the science page. The infrastructure that makes the calibration runs work on Google Cloud — the architecture, the operations, the cost engineering, the receipts — sits at /about/cloud-infrastructure.