Reading results

A Quantum report is designed to be hard to over-read. Most of what follows is about the qualifiers — because a simulation number without its qualifiers is worse than no number.

Every figure breaks down

Three axes, applied to every metric a run produces:

• By archetype — the reason you defined an audience. An aggregate that moved 1% often hides one group up 8% and another down 6%.
• By arm — when the run used A/B arms, with a delta between them.
• By market — content runs, when the audience spans several.

Read along a row, not down a column: "does comprehension hold up in Germany?" is a comparison across markets of one metric, and that is the read the layout is built for.

n, and who isn't in it

Every figure carries the number of agents it is actually about, and the count of those excluded from it with the reason. See the gate for why exclusion is reported rather than folded into the denominator.

Two distinctions that change how a number reads:

• Excluded is not "no". An agent who never saw the ad did not fail to recall the brand.
• Absent is not zero. A page nobody reached has no drop-off rate; it is not reached. A funnel that reports 0% there is telling you something false.

Exclusion skew between arms

If arm A excluded 12% of agents from a measure and arm B excluded 40%, the two numbers are not comparable — the survivors of the heavier exclusion are a different population. The report warns when this happens. Treat the warning as the finding.

Software reports

• Conversion and goal fulfilment — did agents reach the goal you set.
• Funnel and drop-off — per page, arrivals and departures, in funnel order.
• Duration — the distribution, not just a mean. A mean over a bimodal trace is a lie about both modes.
• Response battery — task clarity, friction, confidence, intent, and abandonment reason. Asked after the session; measurement never influences navigation, so a run with the instrument off produces an identical action trace.
• Predicted vs observed — the action distribution your prediction model expected for each bucket, beside what agents actually did.

That last one is the most useful number in the report and the least obvious. The simulation draws its transitions from your model's predictions; comparing them to the observed run tests the model's own claim. A sharp divergence means either the model is wrong about this cohort or the simulation is — and you cannot calibrate against a model you never check.

Content reports

Noticing, brand recall, comprehension, intrusiveness, attitude, purchase intent and CTA awareness — each per archetype, per market, and per arm, each with its n and exclusions.

Provenance

Every run publishes where its numbers came from. Two things to look for:

• Parameter provenance — each behavioural parameter is marked fitted (measured from data) or assumed (a reasonable default), with its bounds and basis. An assumed parameter is not a problem; an assumed parameter you believed was fitted is.
• Calibration provenance — which calibration layer shaped this run, what it was fitted against, and how much was measured. A run on an uncalibrated cohort states "Ran uncalibrated" in the same place. That is a supported state, not an error.

If a report shows a green Calibrated badge but no measurement record, it predates the measurement era and is labelled as such rather than presented as verified.

Sessions

A run lists its individual sessions — the persona that drew each one, what it did, whether it converted, and its answers. Screenshots are captured per session, so you can see what an agent saw at the moment it made a decision.

Debug capture

If a run was launched with debug capture on, you can open a single agent and read its decisions step by step: the page state as it was sent, the model's raw response, the three layer scores, and which action won after blending.

That last column is the point. "What did the model say" is rarely the interesting question. "The model preferred X and the agent did Y — why" is.

Debug capture is off by default and adds a lot of data per session. Turn it on for a small run when you are investigating something specific, not as a habit.

Exporting

Reports can be exported for sharing outside the app. The exclusion counts and provenance travel with the figures — deliberately, because a figure separated from its qualifiers is the thing this whole page exists to prevent.