# Capturing for the best accuracy

## What leaves the device

Capture runs in the browser and extracts facial geometry from every frame: 478 landmarks, 52 expression coefficients and the head pose. Only those numbers leave the device: no image, video or raw audio. With transcription, the text also leaves it. Extra signals (pulse colour values, voice features, hands and body) are off by default and only travel if you ask for them.

## Browsers and frame rate

- Chrome or Edge 94+, Safari 16.4+ (iOS 16.4+). Firefox is not a target: kodec has not been tested there.
- Serve the page over HTTPS or from `localhost`.
- The frame rate depends on the device, not on the platform: a recent laptop or phone usually stays close to the camera's rate, an entry-level phone captures fewer frames per second, and a phone loses rate as it heats up in a long session. Read it at run time with `onCodecEvent('capture-quality')` ([how](/docs/sdk/on-device/#try-it-on-your-phone)), not from the phone model: what limits it is the face detection on the GPU (WebGL 2) and, on phones, heat in long sessions. The device, the mode and how to check them: [Devices, quality and mode](/docs/start/devices-and-modes/).

## What Engine checks in every batch

| Condition | If it fails |
| --- | --- |
| Each frame has `featuresMeta` | Accepted and counted in `framesWithoutFeaturesMeta`; those frames give no valence or arousal axes |
| 478 landmarks per frame | 478 are expected (468 face + 10 iris). With 468, measured on 2026-08-09, Engine accepted every batch and the session never closed; a frame with no face (`landmarks: []`) has not been measured |
| 52 expression coefficients per frame | Accepted; the frame is used without them |
| Batches numbered from 1, the last one `Final` | Without the `Final`, the step stays open; if batches are missing when the `Final` arrives: `400` with `qualityCode: "KPN-EMO-MISSING-PACKETS"` and `missingOrders`, and that `Final` stores nothing: send the missing ones and the `Final` again (see [Declared limits](/docs/concepts/limits/)) |
| `step` exists in the position's plan | `400` with `qualityCode: "KPN-EMO-STEP"` |

```json
{ "status": 400, "qualityCode": "KPN-EMO-STEP" }
```

## Give it a few seconds

Some indicators are computed over a window of recent frames, so the first seconds of a capture have no value for them. In a capture of the sample, workload first appears at about 5 seconds and then arrives in windows, not in every frame (which frame comes first changes from run to run):

```json
{ "framesWithWorkload": 21, "framesTotal": 300 }
```

Plan captures longer than that: Measurement asks for at least 6 s. Engine does not enforce step durations; your interface does.

## Batches

Send batches in order, a few seconds of frames each. With a camera, a frame is about 19.7 KB of JSON from SDK 0.35.0 (478 landmarks), so a few seconds already weigh a few megabytes. From 0.34.0 the SDK cuts each batch at 300 frames or 3 MB of frames, whichever comes first (about 150 camera frames); an SDK before 0.34.0 cut only by count, up to about 9.3 MB. If your proxy limits the body size, allow at least 4.5 MB, or 10 MB in front of an older SDK (see [Proxy](/docs/api/proxy/)); if you send batches yourself, cut them by size as well as by number of frames. Resending the same batch is safe: Engine keeps the first copy.

## Recommendations from the whitepaper

These come from Neurologyca's validation programme, not from checks in the API:

- one person in view;
- a stable distance and head angle: absolute levels are not comparable across setups;
- a personal [baseline](/docs/concepts/baseline/) before comparing people or sessions.

Quality degrades before it fails: low-quality signal lowers how much a reading can be trusted rather than raising an error.

## Quality signals

| Signal | Where |
| --- | --- |
| `framesWithoutFeaturesMeta` | Every `202` of a batch |
| `framesWithoutMediaTime` | Every `202` of a batch: frames without `mediaTimeMs` (where a stimulus player was, optional per frame). Information, not an error |
| `missingOrders` | The `202` of the last baseline batch, and the `400` with `KPN-EMO-MISSING-PACKETS` of a capture's `Final` batch |
| `qualityCode` (`KPN-EMO-…`) | `400` when a batch breaks the contract (step, media, plan, or batches missing before the `Final`) |
| `processing.error` | Live stream: something failed while processing a capture; `stage` says what and `final` whether Engine will retry |
| `hci.warning` | Live stream: a warning about an indicator your context includes, once per batch; it is not an alert; DevLab does not show the level that triggers it (`threshold`), so do not build on it; whether it is sent depends on the environment |
