Creator models
Updated: 2026-10-17. Phase 0 steps 1 to 3 are built and described under "What exists"; steps 4 and 5 are design only. The idea: an enterprise tool where a brand licenses a creator's voice and runs it through Casper skills, with the creator in control and paid per use.
What a "creator LLM" actually is here
Not a model trained from scratch. A creator model is four things, all of which Casper already has a format for:
| Part | Casper form | Notes |
|---|---|---|
| The creator's content | A dataset, kind: text, source.origin: licensed, source.consent: recorded, with a consent record alongside |
Captions, scripts, transcripts, reviews, replies. Uploaded or exported by the creator, never scraped |
| The voice | A skill: a style card distilled from the dataset plus retrieval over it, run on a frontier model | Reproduces how they write and what they would and would not say. Fine-tuning an open-weights model is a later option once volume and consent justify it |
| Control | The policy layer | Every output is a proposal the creator approves, the same way a merchant approves a budget change. Approved outputs feed the dataset back |
| Proof | Evals and the leaderboard | Fidelity: the creator blind-rates outputs against their own writing. Outcome: creator-voice copy versus generic copy on the same campaign, measured by the Meta connector |
Why it belongs in Casper rather than beside it
The wedge is margin-led budget allocation, which decides where to spend. Creator models decide what the ad says. Same connectors, same policy layer, same leaderboard, and the brand already has the store and ad account connected. The dataset format's consent field was designed for this: a dataset whose consent is pending is not listed, and collected or licensed origin without a consent record fails validation.
Phase 0 scope
- Consent record. A
consent.yamlnext todataset.yaml: creator identity, what is licensed (text only at first; no voice cloning, no face), permitted uses, revocation terms, revenue share, signed date. Validation requires it when origin islicensed. - Voice skill.
casper voice build datasets/<creator>writes a skill folder:SKILL.mdcarrying the style card (tone, cadence, vocabulary, hard no's) generated from the dataset with citations back to it, and a retrieval tool over the dataset. - Fidelity eval. Held-out posts: the skill rewrites a brief, the creator ranks real against generated blind. Pass bar is set by the creator, not by us.
- Approval loop. Outputs are proposals; the creator approves in the hub; approvals append to the dataset with provenance.
- One pilot. One creator, one brand, one campaign, outcome measured against generic copy.
What exists (2026-10-17)
Built against the synthetic creator only. No real creator data was used anywhere in this repo. Maya is invented; her sixty posts were written for the repo by a script so the pipeline could be built and shown.
| Step | Where | Notes |
|---|---|---|
| Consent record | packages/dataset-spec: ConsentRecordSchema, readConsent, starterConsent |
consent.yaml next to dataset.yaml. licensed.media accepts only text; voice and face are refused by the schema. Validation requires the file when source.origin is licensed, and a text dataset must declare id, text, posted_at, channel in JSONL. The hub shows the record in the dataset sidebar |
| Demo creator | datasets/demo-creator-maya, generated by packages/cli/src/demo-creator.ts via casper dataset build-demo |
Origin synthetic, consent not applicable. Columns add topic, brief (what a brand would have asked for; the judge uses it) and synthetic likes |
| Voice skill | casper voice build <dataset> <skill> in packages/cli/src/voice.ts; output in skills/voice-maya |
Deterministic style card (tone, cadence, structure, vocabulary, hard no's, examples cited by post id) plus the creator connector's search_posts tool. Input brief, output text, no write scope. A real dataset's consent record becomes a Licence section the prompt enforces |
| Fidelity eval | eval.yaml (three mechanics cases) and casper voice judge |
Judge holds out five posts, regenerates from their briefs with those posts out of retrieval, writes a blind sheet and a key. casper voice score tallies a person's picks. No LLM judge |
Not built: the approval loop (step 4) and the pilot (step 5). Also open: a creator-written voice.yaml of stated hard no's and off-limits topics that the measured card merges in, since measurement can only say what never appeared, not what the creator would refuse.
Risks to settle before a pilot
- Likeness and endorsement law. AI-written endorsements still need FTC disclosure and the creator's sign-off per use. Legal review before any brand pays.
- Platform terms. TikTok and Instagram forbid scraping. Only creator-provided exports, which is also what makes consent real.
- Scope creep into voice and video. Text first. Audio and face cloning are a different risk class and are out of scope until the text product has paying users.
- Who the customer is. Brands pay; creators must earn enough to opt in. The revenue share has to be decided before the first pilot, not after.