Creative canon
Characters drift because nothing in the toolchain remembers them.
Prompt text is not memory. It is a description that gets re-interpreted on every call, by a model that may have been swapped since the last one. Canon replaces the description with a reference and a state that is inherited rather than re-typed.
01 / Inside the workflow
Why characters drift
Ask a generation model for "Anaya, late twenties, ivory kurta, rust dupatta" a hundred times and you get a hundred plausible women. Each is a fresh interpretation of a sentence. None of them is the same person, because there was never a person - only a description being re-rolled.
This is not a flaw that better prompting fixes. It is a consequence of where the state lives. If the only record of who Anaya is exists as words in a prompt, then identity is reconstructed from scratch on every call, and reconstruction has variance. Across a forty-episode season that variance is the show falling apart.
Canon moves the record out of the prompt and into the production. The shot does not describe Anaya. It references her.
- The character exists as a sentence
- Every generation re-interprets that sentence
- Variance accumulates across episodes
- A model change restarts the interpretation from zero
- There is nothing to compare a result against
- The character exists as approved references and state
- Every shot inherits the same references
- Results are evaluated against those references
- A model change does not alter who the character is
- A drift is a detectable mismatch, not a vague feeling
02 / Inside the workflow
What canon holds
Canon is not only appearance. Half the continuity failures in serialized work are not visual at all - they are a character knowing something they have not been told yet.
- Identity references
- Approved face, body, expression range and costume states. Generated results are evaluated against these rather than against the last shot.
- Voice and language
- The approved voice reference, pronunciation guidance for names and regional phrasing, and which language tracks that voice has to survive.
- Knowledge state
- What each character currently knows, when they learned it, and from whom. This is what catches a character referring to a letter before they have read it.
- Relationships and history
- Who trusts whom, what happened between them, and which of those events the audience has seen.
- Possessions and physical state
- Who holds the brass keepsake right now; whether the bandage introduced in episode three is still on the wrist in episode four.
- Locations and props
- Approved architecture, layout, lighting references and the props that live in each space.
- Brand and product canon
- For branded work: logos, product geometry, packaging, substantiated claims, prohibited contexts, colours, spokesperson permissions and market restrictions.
03 / Inside the workflow
How a change propagates
Canon changes. A costume is revised, a location is rebuilt, a brand updates its packaging mid-campaign. The question is not whether change happens but what it is allowed to touch.
In Tosheo a canon change is versioned and scoped. It affects authorized downstream work and leaves accepted work alone. Approved output is not silently regenerated because an upstream reference moved - that is one of the trust commitments the site makes and it is enforced here.
- 01The change is proposed against a version
Not against "the character" in the abstract, but against the specific version the current work inherited.
- 02Downstream scope is computed
The system lists what would be affected: which episodes, which shots, which language tracks, and which of them are already accepted.
- 03A human approves the scope, not just the change
Approving a new costume is a different decision from approving the re-render of eleven accepted shots. Both are shown.
- 04Accepted work is preserved by default
Unless the approver explicitly includes it, already-accepted output keeps the version it was made against, and the story reason for the difference is recorded.
Anaya injures her left wrist in episode three, scene two. A bandage is added to her costume state.
- Change
- Costume state gains a left-wrist bandage
- Scope
- Episode 3 scene 2 onward, all language tracks
- Not in scope
- Episodes 1 and 2, already accepted and unaffected
- Removal condition
- Story event only. The bandage does not disappear because a later shot forgot it.
Every shot generated after that point inherits the bandage until the story explicitly removes it. A shot that comes back without one is a continuity mismatch the QA stage can actually name, rather than something a viewer notices and the team does not.
04 / Inside the workflow
Why this is what survives a model change
Generation is a supplier, not the product. Models will change frequently; production state must survive them.
Tosheo product thesis, brief v5
Model families are superseded on a timescale shorter than a season. A production whose identity lives inside one provider's workflow has to be restarted when that provider ships a new version, or when its pricing, rights terms or availability change.
Canon is provider-neutral by construction. It holds references, state and approvals - none of which belong to a model. Routing a later episode to a different provider changes how the pixels are made. It does not change who Anaya is, what she knows, or what she is wearing.
05 / Inside the workflow
What canon does not promise
Questions
Before you bring it into production
Is creative canon the same thing as a series bible?
A series bible is a document people read. Canon is state the production system reads. The overlap is real - both record who the characters are and how the world works - but a bible cannot block a render, scope a change, or be compared against a generated frame. If you already have a bible, it is a good input to canon rather than a substitute for it.
Does canon stop me changing my mind about a character?
No. It makes the cost of the change visible before you commit to it. You see which episodes, shots and language tracks would be affected, and which of them you already accepted, and then you decide. The thing canon prevents is a change happening silently.
What happens to canon if Tosheo switches generation providers?
Nothing. Canon holds references, state, relationships and approvals, none of which are provider-specific. Routing later work to a different model changes the supplier of the pixels. It does not alter the approved identity, and it does not invalidate work already accepted.
Can two productions share the same characters or brand assets?
That is one of the things we are explicitly measuring during validation: cross-solution reuse of characters, locations, brand assets, rights and delivery profiles. The architecture is built for it. Whether it works in practice is a question the shared-core gate in the brief answers with evidence, not a capability we are claiming today.
Who owns the canon?
Creators retain control of their original worlds and outputs, subject to the pilot and service terms. Accepted assets and production data remain exportable. Canon is a record of your production, held in a system you can leave with.