Phase 182, 182.C.24.01. Which controls are full-parity on mobile and which get
a review-and-approve experience. Targets the V1.1 native apps train
(LILITH_NATIVE_APPS_RELEASE_NAME); the desktop question is answered in
ADR-0007.
The split rule#
Not "simple versus complex". Two questions decide each control:
- Is the phone better at this? A camera, a microphone and a video player
are phone-native. A storyboard timeline is not, and the reason is not screen
size —
182.C.23.04's controls are continuous and spatial, and a finger is a coarse pointer with the target under it. - Is it a deciding operation? Approving audio, accepting a take, releasing
and withdrawing all carry an
If-Matchprecondition because somebody else may be deciding the same thing (182.C.25.03). A deciding operation is not harder on a phone — it is riskier, because the thing that justifies it is the part that fits worst on the screen.
The thing the item's wording misses#
Enrolment is not an authoring control, and it is the strongest reason to have a mobile surface at all.
The subject — the person whose face this is — is not the customer. They may have no Studio account, no reason to open a laptop, and no relationship with us beyond agreeing to this. What they do have is the phone the reference photographs were probably taken on. Their flow is: see what is being asked, see the exact assets it covers, agree or refuse, and later withdraw.
That flow is short, high-stakes, and entirely mobile-shaped. It is also
completely separate from authoring: 182.C.26.02 binds the proof to the
digests of the exact assets, so adding a reference later is a new
enrolment needing new proof rather than an edit — which means the subject's
surface is a sequence of small, final acts, not an editor.
So mobile carries two flows, and only one of them is in this item's title.
Per stage#
Stages from creator-workflow.md.
| Stage | Mobile treatment | Why |
|---|---|---|
| Reference preparation | full parity | the photographs are usually on the phone already; upload, quality feedback and pack review all fit |
| Consent / enrolment (subject) | full parity, primary surface | the subject may never see Studio; see above |
| Script locking | review-only | long-form text editing on a phone is possible and miserable; reading and approving a locked script is neither |
| Pronunciation | review-only | hearing a pronunciation is phone-native; editing IPA or a lexicon entry is not |
| Audio approval | full parity | listening is the whole task, and approving it is one deciding act with an If-Match |
| Scene direction / storyboard | not on mobile | continuous, spatial, multi-track; a coarse pointer over an occluded target |
| Take review | full parity, primary surface | watching a video and judging it is what a phone is for |
| Correction | review-only | requesting a correction is fine; choosing correction parameters belongs with direction |
| Export | full parity | a scoped download decision, not an editing act |
| Deletion / withdrawal | full parity | a subject who wants their likeness gone must not need a desktop to say so |
Rules that follow#
An approve button without the measurements is a rubber stamp with a phone
attached. Mobile may carry an approval only where it also carries what
justifies it: the measurements, and the windows nobody could measure
(182.C.17.06). "No findings" reads as "no problems" on a small screen faster
than anywhere else.
Non-waivable is non-waivable on a phone. The twelve gates in review-and-escalation.md do not acquire a mobile exception, and the four waivable ones still need their bounded expiry and their second approver. A surface where waiving is easier is a surface where waiving happens.
Review-only is a real state, not a degraded one. A control marked review-only shows the value, shows what it affects, and offers no edit — rather than offering a cramped editor that produces worse decisions than no editor.
The subject's flow is not behind the creator's login. It is reached by an
invitation, it shows exactly the assets the proof will cover, and it can be
revisited to withdraw. Withdrawal is in force from the instant it is made
(182.C.38.05), which is the same reason it cannot require an app the subject
never installed.
What this does not decide#
182.C.24.02 (resumable uploads, constrained-device previews, deep links, job
recovery), 182.C.24.03 (biometric leakage through the platform's own
surfaces), 182.C.24.04 (accessibility) and 182.C.24.06 (automation) are
implementation items that follow from this split rather than choices inside it.
What a device test cannot reach (182.C.36.09)#
Added 2026-09-11 with the review flow
(apps/oshun/mobile/e2e/flows/shell/human-video-review.yaml) and the verifier
that keeps its test ids and the screen in step. The automation covers the
supported critical path — open a take, see what was measured, see what could not
be, play only when asked, and find the decision controls — because
the split rule makes review the path a phone has.
The limitations below are the reason this item asks for them to be written down rather than closed. Each is a thing the flow deliberately does not assert.
| Limitation | Why automation cannot reach it |
|---|---|
| The take actually playing | Maestro asserts a view is visible, not that frames advanced. Whether a person moves convincingly is the judgement under review, and no assertion available on a device makes it. |
| Camera capture for enrolment | The system camera cannot be driven from the flow, and a simulator's synthetic frames are not a face. The enrolment screen is 182.C.26 and this is one reason its automation will be partial. |
| A real notification arriving | Delivery goes through APNs/FCM. The flow can open the deep link the notification would carry — checkDeepLink is unit-tested on both branches — and cannot prove the notification was delivered, only that a tap on one lands correctly. |
| A background upload surviving an app kill | The transfer is handed to the OS (sessionType: 'background'). Killing the app mid-transfer and resuming is a device-lifecycle event the runner does not orchestrate; the resume DECISION is unit-tested against a server offset instead. |
| Reduced motion, VoiceOver and TalkBack as the OS applies them | The flow runs with default settings. motionPolicy and the control spec are unit-tested, and a screen reader reading the outcome words out loud is a manual check (182.C.36.08). |
| A constrained device's decoder | Which renditions a given phone can play is hardware. choosePlayback is tested both ways; which way a particular device takes is not. |
Two of these — playback and the decoder — are the same fact in different
clothes: a device test can prove the surface offers the take honestly, and
cannot prove the take is any good. That is what the evaluator stack is for,
and it is why 182.C.40's evidence is not a Maestro run.