# Studio flow design — Directed Human Video

Phase 182, `182.C.23.01`. A dedicated flow, distinct from simple text-to-video
and from talking-head avatar cards. Web surface, V1.0 train
([ADR-0007](../adr/ADR-0007-human-video-authoring-surfaces.md)); the mobile
split is [mobile-parity.md](./mobile-parity.md).

## Why it is not the video card with more fields

The three products are not a difficulty ladder. They differ in **what they
promise**, and a creator picks the promise before they pick the settings:

| Flow                     | Promises the words              | Promises the person        | What identity comes from                           |
| ------------------------ | ------------------------------- | -------------------------- | -------------------------------------------------- |
| Text-to-video            | no                              | no                         | nothing; a prompt describes, it does not authorize |
| Talking-head avatar      | the script, spoken              | the avatar                 | a chosen avatar, not a real person's references    |
| **Directed Human Video** | **the approved audio, exactly** | **this authorized person** | **reference images bound to an authorization**     |

`182.C.13.03` already refuses to _route_ across that boundary — an avatar lane
and an active-scene lane match on every capability a request names and produce
products a creator would never confuse. The flow's job is to refuse the
**drift**: a creator who starts in one and gradually acquires the other's
promise by filling in fields.

**So it is a different card kind, not a mode toggle.** There is no control that
turns a text-to-video card into this one, because the thing that would have to
appear is somebody's consent.

## The structural difference: this flow waits for another person

Every other generation card is a form the creator fills and submits. This one
has a gate in the middle that **is not theirs to close**: the subject has to
enrol, and `182.C.26.02` binds that proof to the digests of the exact
references. That takes wall-clock time and may never complete.

That single fact shapes the whole surface:

- The flow has a **waiting state that is not an error**. "Everything is ready
  and we are waiting for Ama to confirm" is a legitimate resting place for a
  project, possibly for days.
- **Progress is not linear.** Adding a reference after enrolment is a _new_
  enrolment, so the UI shows what is covered, not just what is uploaded.
- **Withdrawal can arrive at any point**, including after generation. The
  surface needs somewhere to show that, and `182.C.38.03` settled what it can
  and cannot undo.

A design that models this as a wizard with a spinner will make the waiting state
look like a failure, and the first support ticket will be "it's stuck".

## Entry point: an authorized person, not an idea

**You cannot start this flow from a prompt.** The first input is a cast: a
person, their references, and the authorization binding the two. Starting from
an idea and attaching a face later is precisely how an unauthorized likeness
enters a pipeline — the prompt is written, the work feels underway, and the
consent step becomes the thing standing between somebody and a finished piece.

Entry is therefore from **the cast, or from an existing project**. A creator
with no authorized cast has one action available: invite somebody.

## Shape of the surface

Five regions, in the order the data forces (`creator-workflow.md` §"Where the
flow is not a line"):

1. **Cast** — authorized people, their reference packs, per-reference quality
   findings, consent state and what each authorization covers. This is the only
   entry point.
2. **Script** — the exact words, per speaker, with pronunciation pinned. Locked
   before audio, because audio is the approved script spoken.
3. **Voice and audio** — source (recorded or synthesised), preview, and one
   explicit approval act. **Approval fixes durations**, which is why it precedes
   the storyboard rather than following it.
4. **Direction** — shots, blocking, camera, and the visible-speech window per
   line. A window declared off-camera needs its reason here, at the moment
   somebody chooses it, not in a later audit.
5. **Takes** — candidates with their measurements, what could not be measured,
   comparison, and the single acceptance act.

Correction and export hang off region 5; deletion and withdrawal are reachable
from region 1 at any time.

## What the surface must never do

- **Never present a prompt field as the primary control.** `182.C.23.05` says
  raw provider JSON is not the control surface; the same argument applies to a
  free-text prompt standing in for direction, because a prompt cannot be checked
  against what was delivered.
- **Never show a route's capability as a product capability.** The eligible
  routes, their limits and their price range belong in the pre-submission
  summary (`182.C.23.06`, shipped), which already refuses to state a price bound
  over routes that cannot be priced and refuses to caption a latency estimate it
  does not have.
- **Never let a take be accepted from a thumbnail.** Acceptance is a deciding
  operation with an `If-Match`; the measurements and the unmeasured windows are
  part of the act, not a detail view behind it.
- **Never let consent state be edited from the creator's side.** It is displayed
  there and changed only by the subject.

## States every region needs

Per `182.C.23.13`: loading, error, empty, and review — and the empty states
carry most of the weight here, because a new project is empty in four different
ways (no cast, cast without consent, consent without script, script without
audio) and each has a different next action. "Nothing here yet" in all four
places is the failure this list exists to prevent.

## What this design does not settle

The component inventory and the interaction detail of regions 2–4, which are
`182.C.23.02`–`182.C.23.05`. This document fixes the boundaries, the entry
point, the order, and the refusals; those items fill it in.
