Privacy

The sandbox is for synthetic test data. The initial live image implementation has the local retention limits below in its deployed code. Signing activation, observed purges and supplier processing guarantees require separate evidence. Broader protocol features remain planned. Start with the runnable quickstart and sandbox support matrix.

A check needs to see the request a model will receive, and a generation check needs the output. This page lists what you send, how long Reflections keeps each kind of data, and how to send less when your product requires it.

Initial live image implementation

The initial live image journey accepts full text inputs and one inline PNG. It does not accept end-user IDs, redacted or withheld prompt modes, retention: "none", uploaded files, reference exports or sessions. These limits apply to Reflections-managed API records and copies:

Data	Local retention
Raw check request text	Deleted when the check becomes terminal, including denials; expired requests are also purged
Private check and generation details, including execution and metadata	While pending, then 24 hours after the terminal boundary; a check remains linked to its pending generation. Detail reads return `404` after this window
Idempotency key and full-body retry digest	24 hours from creation
Unclaimed submitted PNG and derivatives	At most 24 hours from intake
Image awaiting talent release	Seven-day review window, with at most 24 further hours; a decision can shorten that deadline to 24 hours after the decision
Evaluator observations and diagnostics	30 days after the terminal boundary; no raw provider response or evaluator prose is retained
Compact provenance, including request, terms and artifact digests	Through the signed license end plus 3 years
Signed receipts and usage records	7 years

Public receipts and receipt verification expose a defined signed projection and current use status. They do not expose prompt text, execution job IDs or private metadata. Digest lookup requires the owning account’s live key. Private files can expire before their provenance or receipt records; a retained receipt does not restore the file. Reflections’ deletion of its own records does not prove deletion by an upstream evaluator or model provider. Supplier agreements, retention settings and processing evidence must be verified separately before activation. No upstream account-level zero-data-retention guarantee has been verified for this implementation.

Broader protocol: what you send

Data	Sent in	Used to
Prompt, system text, rewritten prompt, scripts, conversation context	Check `inputs`	Find the talent and check the talent’s rules
References, voice samples, masks	Check `inputs`	Match faces and voices, and check licensed changes
Outputs	Generation `artifact`	Check the license’s rules on the bytes
Audio windows, transcripts, frames	Session samples	Check a live session
End-user ID	`end_user.id`	Attribute each use, and apply per-end-user limits

Send the end user as a stable ID scoped to your platform, never an email or a name. Reflections stores only an HMAC of it, and authorization tokens carry that HMAC in the eu claim. Never put personal data in metadata.

Broader protocol retention limits

Data	Kept
Prompt text on `allow`	Deleted once the check is decided; the receipt keeps the `request_digest`
Prompt text on `deny` or `escalate`	30 days, for review and disputes
Artifacts on `allow`	Deleted within 24 hours; `sha256` and fingerprint kept for the license term plus 3 years
Artifacts held for release	Until release or expiry, plus 30 days
Embeddings of references and voice samples	Not stored unless they match; matches kept 90 days
Receipts, usage records, statements	7 years
End-user HMACs	13 months
Talent face and voice templates	While protection is on; deleted within 30 days after

Uploaded files expire after 24 hours unless a held generation needs them, and events are kept 30 days. With prompt_handling.retention: "none", prompt text is deleted as soon as the check is decided, whatever the decision. A reviewer of an escalation and anyone handling a later dispute then see the structured use and the digests, not the text.

Broader protocol prompt modes

prompt_handling.mode sets how much prompt text a check receives.

Mode	What Reflections receives	What it costs you
`full` (default)	The text as your model receives it	Nothing extra; every rule is checked before you generate
`redacted`	The text after you removed personal data	Your redaction step; the check decides on the redacted text
`withheld`	No text: only structured fields and media digests	Rules that need the text are checked on the output instead, so you may pay for a generation that is then denied

Redacted. Remove personal data about anyone other than the talent: email addresses, phone numbers, street addresses, account numbers, and the names of private people. Keep the talent’s name and every word that describes what the output should show, because the talent’s rules are checked against them. Privacy modes gives the patterns. Withheld. With no text, Reflections cannot find a talent from the prompt, follow an indirect reference or match a reference image it only knows by digest. Match names on your side with the name index and pass the talent in talent. Scope terms such as channels, territories and purpose are still checked at the check, because they come from use. The talent’s own rules are checked on the output.

Who is responsible

The protocol assigns the following responsibilities; this text is not evidence that a particular supplier agreement or production processing setup has been qualified. For prompts and outputs, Reflections is a processor for your platform under a data processing agreement. For talent identity data, such as face and voice templates, Reflections is a controller acting for the talent.

Example: a birthday message

A Northlight user asks for “Maya Okafor wishing my sister Dana Whitfield a happy birthday, her number is 555-0142”. Northlight removes the private name and the number, and asks Reflections not to keep the text:

curl https://app.reflections.ai/v1/checks \
  -H "Authorization: Bearer rk_test_…" \
  -H "Reflections-Version: 2026-10-15" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "talent": ["tal_9t0wxyhk5zba2q8v1m3n7p4r6s"],
    "facet": "fct_v1m3n7p4r6s9t0wxyhk5zba2q8",
    "inputs": [{ "role": "script", "text": "Happy birthday, [name]! Have a great run today." }],
    "use": {
      "modality": "speech", "kind": "impersonation", "purpose": "personal",
      "commercial": false, "channels": ["in_app"], "territories": ["US"]
    },
    "end_user": { "id": "usr_48213" },
    "prompt_handling": { "mode": "redacted", "retention": "none" }
  }'

const check = await fetch("https://app.reflections.ai/v1/checks", {
  method: "POST",
  headers: {
    Authorization: `Bearer ${process.env.REFLECTIONS_KEY}`,
    "Reflections-Version": "2026-10-15",
    "Idempotency-Key": crypto.randomUUID(),
    "Content-Type": "application/json",
  },
  body: JSON.stringify({
    talent: ["tal_9t0wxyhk5zba2q8v1m3n7p4r6s"],
    facet: "fct_v1m3n7p4r6s9t0wxyhk5zba2q8",
    inputs: [{ role: "script", text: "Happy birthday, [name]! Have a great run today." }],
    use: {
      modality: "speech", kind: "impersonation", purpose: "personal",
      commercial: false, channels: ["in_app"], territories: ["US"],
    },
    end_user: { id: "usr_48213" },
    prompt_handling: { mode: "redacted", retention: "none" },
  }),
}).then((r) => r.json());

import os, uuid, requests

check = requests.post(
    "https://app.reflections.ai/v1/checks",
    headers={
        "Authorization": f"Bearer {os.environ['REFLECTIONS_KEY']}",
        "Reflections-Version": "2026-10-15",
        "Idempotency-Key": str(uuid.uuid4()),
    },
    json={
        "talent": ["tal_9t0wxyhk5zba2q8v1m3n7p4r6s"],
        "facet": "fct_v1m3n7p4r6s9t0wxyhk5zba2q8",
        "inputs": [{"role": "script", "text": "Happy birthday, [name]! Have a great run today."}],
        "use": {
            "modality": "speech", "kind": "impersonation", "purpose": "personal",
            "commercial": False, "channels": ["in_app"], "territories": ["US"],
        },
        "end_user": {"id": "usr_48213"},
        "prompt_handling": {"mode": "redacted", "retention": "none"},
    },
).json()

The check is decided on the redacted script, the text is deleted once it is decided, and the receipt keeps only the request_digest. Northlight fills in the sister’s name only in what it sends to the voice model, and still submits the output as a generation, where the spoken words are checked.

Privacy modes

Full, redacted and withheld prompts in practice.

Resolution

Matching names locally with the name index.

Privacy in the Standard

Data minimization on both sides.

Get started

Quickstarts

Concepts

Integration guides

Build

Go to production

Initial live image implementation

Broader protocol: what you send

Broader protocol retention limits

Broader protocol prompt modes

Who is responsible

Example: a birthday message

Privacy modes

Resolution

Privacy in the Standard

​Initial live image implementation

​Broader protocol: what you send

​Broader protocol retention limits

​Broader protocol prompt modes

​Who is responsible

​Example: a birthday message

​Related

Privacy modes

Resolution

Privacy in the Standard

Initial live image implementation

Broader protocol: what you send

Broader protocol retention limits

Broader protocol prompt modes

Who is responsible

Example: a birthday message

Related