# Before production

Go through it before you capture with real people in production. Each box links the page that explains it.

## Both modes

- \[ ] Consent is collected **before** `captureProtocol`, and the capture starts only from a gesture of the person. [Privacy](/docs/resources/privacy/)
- \[ ] The API key lives only in your backend, never in a page or an app. [Authentication](/docs/start/authentication/)
- \[ ] Your proxy's `authorize` uses `ownsAll`, with both checks. [Your backend](/docs/sdk/backend/)
- \[ ] You keep the pair of ids (`userProtocolId`, `protocolMechanicId`) in your backend. [Results in your backend](/docs/sdk/results-backend/)
- \[ ] Your page shows the warnings and handles every error by its code. [Errors, warnings and retries](/docs/sdk/errors/)
- \[ ] Your proxies and gateways let the event stream and the long calls through. [Timeouts](/docs/guides/timeouts/)
- \[ ] Your CSP allows what the SDK loads. [CSP and transcription](/docs/guides/csp-transcription/)
- \[ ] You know what deleting a person deletes, and you can do it on request. [Privacy](/docs/resources/privacy/)
- \[ ] You cache the token and know the key rotation: `createTokenSource` reads the key when it is created, so you restart after rotating it; share `tokens` with `createEngineAdmin`; `POST /token` is rate-limited, so do not mint one per request. [Your backend](/docs/sdk/backend/#rotating-the-key)
- \[ ] You ran a pilot on your people's devices: the device check before measuring, the session rate after, and whether deferred (Preview) suits those that cannot measure or transcribe. [Devices, quality and mode](/docs/start/devices-and-modes/)
- \[ ] Your production access is ready: the commercial agreement with Neurologyca, account, key and kit. [Your environments](/docs/start/environments/) · [Pricing](/docs/access/pricing/)
- \[ ] Your legal review covers biometric data and, with an interpretation, the indicators and any transcript sent to an LLM. [Privacy](/docs/resources/privacy/)
- \[ ] Your app does not use the readings to diagnose, on their own to decide about the person, or to share them without consent, and a person reviews every decision that uses them and affects someone. [Responsible use](/docs/guides/responsible-use/)
- \[ ] Each reading reaches only its declared reader, the person can pause or stop at any time and, in health, wellbeing, mental health or emotional support, a route to help is always visible. [Responsible use](/docs/guides/responsible-use/#what-your-app-must-do)

## On-device

- \[ ] You tested on the browsers and phones your people use. [On-device (real time)](/docs/sdk/on-device/#try-it-on-your-phone)
- \[ ] Transcription is on only in protocols with speech steps. [On-device (real time)](/docs/sdk/on-device/#transcription)
- \[ ] Your startup screen paints the messages of `onStartupStage`, offers Cancel (`signal`), waiting or going on without speech on `startup.stalled`, and a reload after the face model runs out. [On-device (real time)](/docs/sdk/on-device/#the-startup-screen)
- \[ ] If you prepare before Start: one client per page, `release()` when the page or component goes away, and speech (about 160 MB) only after the person agrees. [On-device (real time)](/docs/sdk/on-device/#prepare-before-start); which way fits your flow: [Load the models](/docs/guides/load-models/)
- \[ ] If you keep the models between captures (`keepPrepared`): one `prepare()` before the first capture and none between captures where there is no WebGPU, `client.release()` before you drop the client and between one person and the next on a shared device, and, if you receive startup reports, `capture-server` 0.19.0 or later. [On-device (real time)](/docs/sdk/on-device/#keep-the-models-between-captures)
- \[ ] If you host ONNX Runtime with a `wasmPaths` string, it ends in `/` and serves the pair of each executor. [CSP and transcription](/docs/guides/csp-transcription/#hosting-the-files-yourself)
- \[ ] If you send startup reports (`startup.report.url`), your consent and privacy notice cover that browser data. [Your backend](/docs/sdk/backend/#startup-reports-optional)

## Deferred

- \[ ] Your organization has accepted the DevLab Terms, including the [Data Processing Addendum](/legal/dpa/) (DPA) that covers Neurologyca's processing service (your organization admin accepts them; see **Your account › My agreements**), and you use the deferred mode only with people who have given their informed consent. [Privacy](/docs/resources/privacy/)
- \[ ] Your receiver deletes confirmed videos when the job ends. Your scheduler calls `receiver.sweep()` periodically (for example every hour), because the receiver starts no timers. [Your backend](/docs/sdk/backend/#the-upload-receiver)
- \[ ] You retry pending notifications: your scheduler walks `receiver.pendingNotifications()` and calls `receiver.notify(captureId)` for each. [Errors, warnings and retries](/docs/sdk/errors/)
- \[ ] Your webhook is idempotent by `jobId` and job state. [Results in your backend](/docs/sdk/results-backend/#the-deferred-webhook-as-a-contract)
- \[ ] The service reaches you only over HTTPS, with different secrets for the signed URLs and the webhook and one nonce store per purpose. [Processing service](/docs/sdk/processing/#with-our-service-or-self-hosted)
- \[ ] Your token URL hands a token only for a running job of that recording. [Processing service](/docs/sdk/processing/)

## Where you are

The **Status** view of [Your integration](/integration/progress) shows your team's milestones and the next step.
