Docs / Architecture

The quality system

The deterministic gates, the independent critic, fix rounds, budgets per job, and how lessons flow back into the briefs.

Nothing reaches a platform because a model said it was good. Every episode passes two kinds of check on its GPU slot before it can be published: measured gates — code that measures the finished video, the same way every time — and an independent critic, a separate model that did not write the episode. What fails goes back to the writer as a fix round; what still fails is held for you, and not charged. What went wrong becomes a lesson for the channel's next episodes.

 finished render
       │
       ▼
 measured gates ── fail ──┐
       │ pass             │
       ▼                  ▼
 independent critic ── fail ──▶ fix round (the same writer, with the reports) ──▶ measured again
       │ pass                                   │ still failing after its rounds
       ▼                                        ▼
 ready (in the library)                  held (in the library, not charged)
       │                                        │
       └──────────── lessons for the channel's next briefs ◀───────┘

The measured gates#

Each gate is code with fixed limits. A failure names the gate and what to fix.

Gate	What it checks
Posting texts	every field the platforms need is there and within their limits: a title of at most 100 characters without angle brackets, a description of at most 5,000, tags under 480 characters in all, hashtags as hashtags
Identity	the channel speaks as itself: never as a real person, never as made or approved by a brand or organisation that did not make it, and with no machine-translation wording
Video	1080 × 1920 at 30 frames a second, H.264 with AAC sound at 48 kHz in stereo, 120–175 seconds and exactly the length planned
Voice	the channel's own narrator; lines paced to be heard; the narrator speaking at most 80 % of the time, with silent stretches for the picture and the music; the last line said before the end card; the narration really audible in the final mix
Facebook edit	a valid edit of 55–89 seconds for Facebook's Reels, which then has to make sense on its own: read by a story check that wants 8/10, re-cut from better lines when it falls short. When no edit passes, the full video still ships and the Facebook and Instagram posts wait
Text	every word on screen inside the one text band, nowhere else
Units	units and symbols as the standards write them: what a viewer reads uses the standard's symbols and spacing, what the narrator says is in words a voice cannot misread
Loudness	−15.5 to −12.5 LUFS integrated, true peaks no higher than −0.5 dBFS
Picture	no black stretch before the end card, no picture frozen for 5 seconds or more, a first frame that is not black
Flashes	photosensitive safety (WCAG 2.3.1): no area flashes more than three times in any second
Pace	no shot shorter than the pacing allows
Page	stills rendered again with logging on: no script or shader errors
Captions	4.2–7.0 seconds each, at most 2.0 words and 10.5 characters a second, with gaps and breathers, covering at most 70 % of the video
Variety	the episode's setting differs from the channel's previous episodes, as the planner assigned it

Some gates apply only where a channel uses the feature:

Gate	Where	What it checks
Loop	channels whose endings loop	the last frame cuts straight into the first, and no caption sits over the seam
Originality	where the originality guard is on	no title, opening line or narration the channel used before
Product	channels made for a product	every claim about the product as its sourced brief lists it, the call to action said once mid-video, no health, safety, money or comparison claim without a source, the product's picture and name on the end card, no link in the texts

The critic#

The critic is a separate review by Claude, a larger model than the one that wrote a standard episode, with its own budget that the writer cannot touch, so no episode can go out unreviewed because its writer spent everything.

What it sees: a contact sheet of frames across the video, the captions, the narration and the measured motion of the picture.
What it scores: visual quality, readability, factual accuracy, platform policy and entertainment.
To pass: a score of 8/10 or more, no fact it found wrong, nothing a platform's rules would not allow, and no blocking issue. Anything else is a hard failure.
Entertainment below 7/10 — a static, empty or predictable video — sends the episode back once, with the critic's concrete list: show the consequences, keep the camera moving, build to the payoff. If the round leaves no hard failure, the episode ships, and the reasons are kept as a lesson.

Its verdict — the score, a summary, what failed and what it would change — is shown on the episode's page, and through the API and MCP in plain words.

Fix rounds#

A failure resumes the same writing session with the reports: the gates' failures and the critic's fixes. It does not start over, so a fix round costs far less than a new episode.
An episode gets up to two fix rounds for hard failures. Every round is bounded by the job's model budget and its time.
An episode that passed its first build below 9/10 (or below 8/10 for entertainment) may get one polish round with the critic's improvements, when budget and time allow. The passing version is kept: if the polish scores lower or breaks something, the passing version ships.
When a round's render is cut off before it finishes, the version the round started from comes back, with its own checks — never nothing, and never made again from scratch.

Budgets and time#

Every job runs within limits set before it starts:

A model budget per job, in a token only that job holds. The writer's share and the critic's share are separate. The gateway refuses a call once a share is spent: that ends the round, not the episode — the checks judge what the round made.
A time box: the writer's rounds have a little over two hours; the GPU job has a hard limit for its kind, after which it is stopped.
Narration within a character cap for the job, the workspace's day and the service's day.
Your money within its hold and your caps: the job is charged what it really used, at the plan's rate, and the workspace's and the channel's monthly spending caps apply to every job.

Budgets exist so that a run that goes wrong costs a known, small amount — and, through the hold, what it costs is shown before it starts.

Held episodes#

An episode that still has blocking points after its fix rounds is held: it is in your library with the reasons in plain words ("The critic scored it 7/10 — a short needs 8 to pass", what failed, what the critic would change), it never posts, and it is not charged. No second episode is made in its place. You can watch it, then let it go, or release it to use it anyway: releasing a held episode charges what it cost to make, shown on the button (Release for posting — N credits). An episode a person held after it was ready (from the approval queue or a review link) was already paid for, and its release is free.

Lessons#

Every episode that did not pass cleanly on its first attempt leaves a lesson: which gates rejected it, the critic's blocking points and fixes, its scores, and how it ended. The next episode of the same channel gets the channel's lessons in its brief, distilled and ordered by how often they recur — a point the critic raised on three earlier episodes comes before one raised once. Alongside them, the brief carries what the channel's own numbers taught: the openings the hook lab found hold its viewers, and the retention loop's rules for where they leave.

Lessons stay inside their channel and workspace. Nothing learnt from one workspace's episodes or platform data is given to another's.

What the system does not promise#

The gates and the critic reduce mistakes; they do not guarantee that every fact is right or that every platform will accept every video. That is why every episode is in your library before it posts, why you can ask for approval before each one, and why you can hold any of them. See Units, symbols and facts for how facts are sourced and checked.

The quality gates and the critic — the same checks, for members.
The episode pipeline — where the checks sit in an episode's job.
Originality guard and checker — how episodes stay different from each other.
Approvals on your phone — a person's yes before posting.

Something missing or unclear? Tell us and we will improve this page.