AI XR Analytics —
What the Models Actually Do.
This page is written for the technical evaluator asking 'is there real machine learning here, and what does it and does it not decide?'. Not marketing copy about intelligence — a plain description of what the models cluster, what they predict, what they refuse to answer, and what stays out of training.
The intelligence layer sits above the capture layer (/track-xr-user-behavior) and the researcher workflow (/analyze-xr-behavior). Its job is triage: given millions of spatial events per week, decide which moments a human should actually look at first — and give a defensible reason for each pick.
Clustering, not classification
The system groups comparable friction moments across sessions using unsupervised clustering over the derived behavior signals. No pre-labelled 'good' or 'bad' session — clusters emerge from the data, and each surfaced cluster carries the feature deltas that define it so a researcher can accept or reject the grouping.
Drop-off prediction with explicit horizon
A per-scene model predicts probability of session abandonment inside a bounded look-ahead window (seconds, not minutes). Predictions are ranked by expected retention impact and are always attached to the specific interaction or waypoint that drove them — no unattributed 'AI says'.
Comfort clustering ahead of tickets
Motion-sickness and discomfort patterns are surfaced as clusters over headset motion and disengagement signals, not classifications. The system flags patterns that historically preceded refund/support activity in analogous cohorts — with the caveat that this is a pattern signal, not a medical claim.
Refusals the layer will make
The models will not label individual users, will not score sessions on a 'quality' axis, and will not predict outcomes for cohorts smaller than the confidence floor. When a question is out of scope, the layer says so explicitly — the alternative is confident nonsense.
Frequently asked questions
- How is this page different from /analyze-xr-behavior?
- The Analyze page is about the human workflow — how a researcher segments cohorts and diffs releases by hand. This page is about the automated layer above it — what the models decide to surface without being asked, and how those decisions are made.
- Are the models trained on my data alongside other customers?
- No. The intelligence layer runs on aggregated, customer-scoped signals. There is no cross-tenant training and no raw audio, video or biometric input in any training pipeline — clustering and prediction operate on the derived behavior signals described at /track-xr-user-behavior.
- What are the confidence floors — when will the layer refuse to answer?
- Below a cohort size and event-volume floor per model, the layer returns an explicit 'insufficient signal' rather than a low-confidence answer. Floors are surfaced in the response payload so downstream automation can gate on them.
- Can I inspect why a cluster or prediction was surfaced?
- Yes. Every surfaced cluster carries the feature deltas that define it; every prediction carries the interaction or waypoint that drove it and the horizon it applies to. This is a triage layer, not a black box — researchers can accept, reject or override any surfaced item.