Skip to main content

Quantum simulations

Analytics tell you what already happened. Quantum runs your product forward: it drives a population of synthetic users — calibrated against the behaviour your own prediction model learned — through your real site, app, or creative, and reports what they did and what they made of it.

The workflow it exists for:

  1. Build a cohort — a synthetic population shaped like your real users.
  2. Point it at a target — your live product, a staging build, or an ad.
  3. Read the result per cohort, not just in aggregate.

The aggregate is where most pre-launch signal goes to die. A change that moves the average by 1% is frequently a change that helped one group and hurt another. Every Quantum figure breaks down by the groups you defined, for exactly that reason.

Two tracksDirect link to Two tracks

A cohort tests one of two things, chosen when you create it.

TrackYou are testingTargetAgents produce
SoftwareA product — website, web app, or Android buildA URL or an APKA click trace: pages reached, actions attempted, where they stopped
ContentA creative — video, image, or display adAn uploaded asset in a simulated media environmentAttention and response: did they notice it, recall the brand, understand it, find it intrusive

They share cohorts, archetypes, audiences and calibration. They differ in what a session is and what a report contains. Start at Software simulations or Content simulations.

Content track availability

The content track is enabled per account. If you don't see it in the app, contact support@moveo.one.

How an agent decidesDirect link to How an agent decides

Each agent's next action blends three signals. None alone would be enough, and the blend is what makes a run reproducible rather than a creative-writing exercise:

  • Probability-calibrated transitions. Your prediction model exports the action distribution per probability bucket — what a user at 85% likelihood actually does next, versus one at 12%. Agents draw from those distributions rather than from a generic prior.
  • Persona reasoning. Each agent carries an archetype: a described person with behavioural parameters — patience, ad tolerance, attention span, prior familiarity. A language model reasons about what that person does on this screen.
  • Model-learned feature importance. The features your model weights most heavily bias what agents attend to, so a change that touches an important feature shows up in the run.

Runs are seeded, so re-running a seed reproduces the same population and the same split.

What Quantum is notDirect link to What Quantum is not

Being clear about this saves a lot of misread reports:

  • It is not a substitute for a real A/B test. It is a way to know what to look for before you spend real traffic finding out.
  • It is not a load test. Agents are behaviourally realistic, not concurrent at scale.
  • It does not invent behaviour it never saw. A cohort is only as good as the model it was drawn from and the data that model was trained on. Every run publishes the provenance of its numbers — see Reading results.

PrerequisitesDirect link to Prerequisites

  • Events flowing into Moveo One — see Installation.
  • A trained prediction model. Cohorts are drawn from a model's probability buckets, so the model comes first.
  • For an Android target, an APK you can upload.

NextDirect link to Next