Cohorts and audiences

A cohort is the population a simulation runs. It is drawn from one of your prediction models — the model supplies the probability buckets and the per-bucket action distributions that make a high-intent agent behave differently from a low-intent one.

Everything else on this page is about making that population specific: who these people are, and how confident you are that they behave like your real ones.

Creating a cohort

In the app, Quantum → New cohort. You choose:

Field	Notes
Track	Software or content. This is fixed at creation — it changes what a session means.
Prediction model	Supplies the behavioural buckets. See Build a model.
Target	A URL for a web run, or an uploaded APK for Android. Can also be set per run.
Agent count	How many synthetic sessions. Can be set per run instead.
Audience	The archetype mix — see below. Optional, but a cohort without one is behaviourally flat.

You can also build a cohort from a document: upload an ICP description, persona research or a segmentation deck, and Moveo One proposes an audience from it, which you then edit. It is a starting point, not an oracle — read what it proposed.

Archetypes

An archetype is a described person with behavioural parameters attached. It is not a label on a chart: the parameters change what agents of that archetype actually do — how long they persist through friction, how much interruption they tolerate, how quickly they recover attention after a distraction.

Archetypes are described by facets — a closed vocabulary covering things like:

• Who they are — age band, geography, category involvement, brand relationship
• How they behave — patience, attention span, ad tolerance, ad literacy
• The situation — daypart, device posture, social setting, time pressure, whether they're paying for an ad-free tier

The vocabulary is closed on purpose. A facet value outside it is rejected when you save the cohort, rather than silently defaulting to "medium" and producing a persona that reads right and behaves wrong.

You describe the person, not the propensity

You cannot set an agent's skip rate, dwell tolerance or sound propensity directly. Those are derived from the facets you chose. Setting both would let a cohort describe someone patient who behaves impatiently, and nothing downstream could tell which one was true.

The software track lets you author your own archetypes. The content track uses a published library, which you compose into a mix rather than edit.

Audiences

An audience is a weighted mix of archetypes — "40% comparison shoppers, 35% returning loyalists, 25% incidental visitors". Weights determine how many agents of each kind a run draws.

Two things make audiences worth using rather than running one flat population:

• Every figure in a report breaks down by archetype, so a change that helps one group and hurts another is visible instead of averaged away.
• Audiences are saved and reusable. Test a redesign against the same audience you tested the original with, and the comparison means something.

You can also describe an audience in plain English and have Moveo One propose the mix. It returns archetypes from the closed vocabulary — review the weights before you save.

Calibration

Calibration is what turns "this cohort is plausible" into "this cohort matches something measured". It is optional. An uncalibrated run is a supported, honestly-labelled state — the report says "Ran uncalibrated" rather than implying more.

Calibrating fits a small number of behavioural parameters against a dataset you supply, and writes the result as an immutable layer. Layers compose into a chain, so a later refit adds to the record instead of overwriting it. A refit is always a new layer, never an edit — which means you can always see what a past run was calibrated against.

The flow

1. Upload a dataset. A CSV export of real behaviour.
2. Inspection. Moveo One reads it and proposes a column mapping. Status: INSPECTING.
3. Mapping review. You confirm or correct the mapping. Status: MAPPING_REVIEW.
4. Fit. Status: CALIBRATING, then CALIBRATED or CALIBRATION_FAILED.

Step 3 is a required human gate

Nothing is fitted until you confirm the mapping. A calibration is only as good as knowing which column is which, and a wrong guess produces a confidently wrong cohort. Moveo One will wait in MAPPING_REVIEW indefinitely rather than proceed on an assumption.

Statuses

Status	Meaning
UNCALIBRATED	No calibration. Runs are allowed and labelled as such.
INSPECTING	Reading your dataset, proposing a mapping.
MAPPING_REVIEW	Waiting for you to confirm the column mapping.
CALIBRATING	Fitting.
CALIBRATED	Fitted and accepted. The report names the layer used.
CALIBRATION_FAILED	Fitting ran but did not produce an acceptable result. The reason is shown.
STALE	The underlying model was retrained; the calibration predates it.

A rejected layer is a result, not an error. If a fit does not pass its own checks, the layer is stored with the reason it was rejected, and the cohort keeps whatever it had before. You can read why rather than guessing.

Next

• Software simulations — run this cohort against a product.
• Content simulations — run it against a creative.
• Reading results — what comes back.