Methods and limitations
How the character quiz works
An independent entertainment quiz with original behavioural questions and transparent, versioned comparisons.
The same model, different worlds
A shared bank contains 60 independently written behavioural spectra. The compact, main and fuller trial forms contain 24, 35 and 60 questions. Each future series supplies a separately reviewed portrayal-specific cast, biographies, images and frozen editorial profiles. The questions, interaction and matching methods are shared.
Question order is shuffled once per attempt. The descriptions and their 0/100 directions stay fixed. Choosing 50 deliberately counts as an answer; untouched, skipped and unknown responses remain distinct. New attempts need all personal questions answered before matches appear. Longer forms cover more behaviours; that alone does not establish greater accuracy.
Three ways to compare
- Pearson correlation · default
- Compares the relative pattern across traits after subtracting each profile’s mean. The displayed index is 50 × (r + 1). A constant profile has no defined correlation.
- Mean absolute difference
- Compares the average absolute gap on the native scales. The index is 100 minus that average gap. A person choosing the midpoint everywhere can still receive a distance ranking.
- Midpoint-centred cosine · experimental
- Subtracts 50 from both profiles and compares the direction of the resulting vectors. The index is 50 × (cosine + 1). A neutral profile has no defined direction.
Traits have equal weights. Near-constant population variance is treated as undefined at 10⁻¹²; ties use an anchored tolerance of 10⁻⁹. Compatible historical reports can retain missing responses and their original approved ten-character snapshot. Their coverage minimums are 20, 28 or 48 shared traits; missing responses are never filled with 50. New reports use the current cast and complete personal responses.
These are similarity indices, not percentages of accuracy, probabilities of being a character, or population percentiles. Method choice can change rankings. Scatter plots use fixed 0–100 axes and an equal-score diagonal; they do not fit a trend line.
What the profiles can tell you
The current character values are provisional AI-assisted editorial starting estimates, written independently for specific fictional portrayals. They are not measured community consensus or actor personality assessments. Some everyday preferences are weakly evidenced in fiction. Character descriptions and profile estimates should be read with that uncertainty in mind.
Software checks establish reproducible scoring and interface behaviour. They do not establish empirical matching accuracy. We have not yet measured that with fresh participants. Future evidence should examine independent rater agreement, repeat responses, trait interpretation, sample composition, selection bias and uncertainty.
Optional character contributions
The three forms offer up to 3, 4 or 7 extra tasks during the personal quiz. People can decline immediately, choose familiar characters and give separate development permission. With at least two familiar characters, slider and pair tasks are mixed; with one, only sliders are used. Unknown, skipped and stopped responses stay distinct. A pair tie means “about the same”, not a numeric distance.
After personal completion, people can view results immediately or volunteer for one further batch with fresh unchecked permission. Original answers, timings and in-test contributions then remain locked. Additional ratings cannot change that completed result. Private ratings are retained separately from editorial starting profiles. There are no automatic profile updates or published aggregates. Any later aggregation and profile publication requires review.
Private recording and your choices
Starting creates a pseudonymous session. Completed personal answers and results use TypologyOnline’s shared private assessment recorder, including questionnaire/profile versions, frozen scoring inputs, actual question order, client-reported timing and revision metadata, and server receipt times. Timing may help identify quality concerns; it cannot prove attention or honesty.
Character-development permission applies only to optional character ratings. Personal-answer research has its own unchecked opt-in after completion. Neither is required to receive a complete result. We do not inherit consent from other tests. Imported offline files stay local and are not trusted native submissions; old file permissions do not authorise a new server upload.
During this open visit you can withdraw both optional purposes or delete the attempt. Withdrawal removes character contribution content and excludes personal responses from research; ordinary personal results remain privately stored. Deletion removes the stored personal responses, results and contributions, retaining only a minimal session tombstone to prevent replay. Account deletion cascades to linked session/contribution records. Keep your private attempt reference for later requests through Contact; an identifier alone is not proof of ownership.
Records expire after two years under the dedicated retention command; it must be scheduled by the operator before public collection. Restricted exports separate synthetic fixtures and exclude withdrawn records. Application-held research exports are invalidated when permissions change or data are deleted, and expire after seven days. Operators must honour withdrawals in any analysis copies taken outside the application. Raw individual data are never public. Browser access tokens remain in page memory; private downloads do not contain them.