Soultrace 3.0: The Trait Model 🚀
We shipped a big overhaul to how Soultrace works under the hood. Version 3.0 replaces direct color classification with a two-stage latent trait model. Here's what we changed, and why.
The old way (v2.x)
The previous engine updated color probabilities straight from your answers:
Answer → P(color | answer) → Updated color distribution
It worked. It also had three specific headaches we kept patching around.
Colors were computed from raw answer patterns instead of being anchored in known psychological constructs, which meant we couldn't explain why a given answer nudged you toward blue vs. black beyond "that's what the table says". Extreme responders (people who always pick 7) walked away with different results than moderate responders (people who pick 5) even when their underlying personality was identical. And calibrating those likelihood tables — one per question × color × score combo — was a chore that scaled badly.
The new way (v3.0)
We now infer 8 latent psychological traits first, then transform those into colors:
Answer → Trait Updates (Bayesian) → Weight Matrix × Traits → Softmax → Colors
The 8 traits
| Trait |
What It Measures |
| Conscientiousness |
Organization, dependability, self-discipline |
| Need for Cognition |
Enjoying effortful thinking |
| Analytical Thinking |
Preference for systematic reasoning |
| Agency Motivation |
Drive for achievement and power |
| Promotion Focus |
Orientation toward gains and aspirations |
| Sensation Seeking |
Need for novel, intense experiences |
| Emotional Expressivity |
Comfort with displaying emotions |
| Communion Motivation |
Drive for connection and belonging |
None of these are invented. They come from decades of personality work — Costa & McCrae, Cacioppo & Petty, Higgins, Zuckerman, and the tradition built around them.
ERS conditioning
Biggest win: Extreme Response Style correction.
A chunk of users systematically pick endpoints — 1s and 7s all the way down. A different chunk hugs the middle and rarely ventures past 3 or 5. That tendency has nothing to do with personality, but naive models conflate the two and it bleeds into every downstream estimate.
V3 models ERS as its own latent variable. The first handful of answers calibrate your response style, and every subsequent trait update is conditioned on it. Someone who always picks "strongly agree" ends up with trait estimates close to someone who picks plain "agree".
Smarter question selection
Question selection now targets trait uncertainty rather than color uncertainty. We pick the question that maximizes information gain across whichever traits are still wobbly, with a small coverage bonus so all 8 traits get a look.
Net effect: stable results in roughly 24 questions instead of 50+.
What you'll notice
Honestly? The results page looks the same. You still get a 5-color distribution and an archetype match.
The estimates underneath are sturdier though. They're less sensitive to your response-style quirks, anchored in real psychological constructs, and better calibrated — random answers actually produce a uniform distribution now, which wasn't always the case before.
The math
If you want the full breakdown — Bayesian updates, weight matrices, information gain — the technical deep-dive has it.
Try it
The new model is live. Take the test and see how it feels.
Other Articles You Might Find Interesting
