Continuity QA
A continuity error is a mismatch against something specific.
That is the whole reason canon exists. Without an approved reference, "she looks a bit different" is an opinion. With one, it is a comparison, and a comparison can be checked, ranked and routed to a repair.
01 / Inside the workflow
The two kinds of continuity
Visual continuity is what people expect: does this look like the same person, the same room, the same jacket. It is the easier half, because a reference exists to compare against.
Narrative continuity is the half that ruins serialized work and gets checked least. A character who knows something they were not told. A prop in two places. An injury that heals between episodes. A relationship that resets. These are invisible to any check that only looks at pixels, and they are the errors an audience actually notices.
- Face, body and expression against the approved set
- Costume state for this story day
- Location architecture, layout and lighting
- Prop presence, condition and position
- Product geometry, packaging and colour for branded work
- What each character knows, and when they learned it
- Who currently holds which possession
- Physical state carried from earlier episodes
- Relationship state after the last reveal
- Events the audience has and has not seen
02 / Inside the workflow
Compliance is checked in the same pass
Separating creative QA from compliance QA means two reviews, two backlogs and one of them being skipped. They run together.
- Dialogue against the approved script, including improvised drift
- Pronunciation of names and regional phrasing per language track
- Caption accuracy, timing and reading speed
- Synthetic-media disclosure present where it is required
- Brand rules: prohibited contexts, colour use, spokesperson permissions
- Substantiated claims: the claim shown is the claim that was approved
- Delivery specification: the technical requirements the buyer set
03 / Inside the workflow
Why a check never accepts anything
Automation can propose, generate, check, route, and assemble. It cannot silently greenlight spend, approve rights, accept a final cut, publish externally, or represent a person or brand.
Human-control rule, brief v5
A check produces a warning with a location, a reason and a reference it failed against. It does not produce a verdict. The reason is partly accountability and partly accuracy: plenty of flagged shots are correct, and plenty of clean shots are wrong in ways no check was looking for.
Quality in Tosheo is defined as accepted output. Generated seconds have no value until an authorized human accepts them for the intended delivery, and that definition is what the whole review stage is built around.
A generated shot has Anaya referring to the letter's author by name, in episode two.
- Check
- Narrative continuity, knowledge state
- Reference
- She has not read the letter as of this scene
- Severity
- Contradiction, blocking for this shot
- Proposed action
- Script revision, then regenerate this shot only
No visual check would have caught this. The shot is well made and the performance is good. It is simply impossible within the story, and the only reason the system can say so is that the knowledge state was recorded when she did not yet know.
Questions
Before you bring it into production
Does the QA stage catch every continuity error?
No, and we will not claim it does. It catches mismatches against things that were recorded. Something nobody put in canon cannot be checked against canon. What the stage changes is the proportion of errors found before delivery rather than after publication, and it makes the ones it finds specific enough to repair cheaply.
What does a warning actually do?
It attaches to the shot, names the reference it failed, and proposes an action. It does not reject the shot and it does not accept it. A reviewer decides, and their decision is recorded against the version they saw.
Can I dismiss a warning?
Yes, with the dismissal recorded. Sometimes the reference is wrong, sometimes the story changed, sometimes the flag is a false positive. What is not possible is a warning silently disappearing.
How is QA quality measured?
Continuity and compliance warnings per accepted minute, accepted shots per generation attempt, and repair success without collateral change. Those are the metrics the brief defines for production quality, and they are what we will publish once we have run enough production to have honest numbers.
Who reviews the language tracks?
Native review per track is recorded as part of the track, not folded into the primary master review. A pronunciation problem in Tamil is not something a Hindi reviewer can be expected to catch, and pretending otherwise is how localization quality quietly degrades.