Provider routing
Model access is becoming abundant. Production reliability is not.
The durable layer is not a model selector. It is the system that decides what to send where, refuses to cross a budget ceiling, honours the rights restrictions on a reference, and stops swapping providers once a look is approved.
01 / Inside the workflow
What the routing decision uses
- Modality
- Video, image, voice or audio. A shot may involve several, and they do not all go to the same place.
- Reference support
- Whether the provider can actually accept the identity references this shot has to inherit. A provider that cannot is not a cheaper option, it is the wrong option.
- Language
- Which languages a voice provider handles credibly, including pronunciation of names and regional phrasing.
- Quality need
- A board panel and a delivery master are not the same job and should not cost the same.
- Cost and the authorized ceiling
- The maximum from the cost-authorization gate is a hard bound, not a target the router tries to stay near.
- Rights terms
- Whether the provider terms are acceptable for this deliverable, and whether a reference asset carries a model restriction that rules the provider out.
- Recent acceptance performance
- How often work routed to this provider, for this kind of shot, has been accepted by a human lately. Acceptance is the only quality signal that means anything.
02 / Inside the workflow
Why the router deliberately stops optimising
A router that always picks the best available model produces a season that looks like it was made by six different shows. Variety is a virtue in a demo reel and a defect in serialized work.
So once a treatment is approved - a provider, a set of parameters, a look - the router locks it for the scope where consistency matters, usually the season or the campaign. Improvements are adopted deliberately at a version boundary, not silently between episode four and episode five.
- Every shot goes to whatever currently scores highest
- Look drifts as providers ship updates
- Two episodes can differ for no story reason
- Cost is unpredictable per episode
- Repair is harder: the original conditions are gone
- An approved provider and treatment is fixed for the scope
- Episodes match because they were made the same way
- A change of provider is a decision with a version boundary
- Cost per accepted minute becomes forecastable
- A repair can reproduce the conditions of the original
03 / Inside the workflow
Portability is a commercial position, not a technical one
Providers change pricing, terms, availability and quality on their own schedule. A production business whose work cannot move is exposed to all four.
Tosheo stays portable because canon, rights, approvals and lineage are held outside any provider. Moving a production to a different model changes the supplier of the pixels and nothing else. That is also why this site does not compete on model access: access is the commodity, and the governance above it is not.
Questions
Before you bring it into production
Which models does Tosheo use?
Routing is by capability rather than by brand loyalty, and the set changes as providers ship. The /models section describes the families we route to, what each is genuinely used for, and what a production system still has to control above them. It does not rank them, because a ranking we have not measured would be an invented claim.
Can I choose the model myself?
You approve a treatment, which in practice fixes the provider and parameters for the scope. That is the level at which the choice is meaningful. Picking a provider per shot is a way to make a season look inconsistent.
What happens when a provider ships a new version mid-season?
Nothing, unless someone decides otherwise. A locked treatment does not follow provider updates automatically, because an episode that silently changes look is a continuity failure with a technical cause. Adopting the new version is a decision with a version boundary and a scope.
Am I charged when a provider fails?
No. Provider failures and retries are accounted for as a Tosheo cost, not a customer cost. That is a deliberate incentive: the party choosing the router should carry the cost of choosing badly.
Does the router ever ignore the cost ceiling?
No. The authorized maximum from the cost gate is a bound on dispatch. A job that would cross it does not run and quietly warn - it requires a new authorization from a named person.

