Creators
Publish a fix you can prove.
If you know a narrow task where an open-weight model keeps failing, and you have data you're allowed to train on, Retoque gives you the training and evaluation harness, a signed package format, and a public dossier that shows the result as it is.
Where things stand
Creator publishing is new. No outside creator has published a patch yet, and no creator earnings exist. The first patch is first-party, built by Retoque to test the pipeline end to end.
Fit
What makes a good patch
- One failure, described in a sentence, that the unpatched base model actually makes.
- A pass condition a program can check, such as exact JSON fields, a passing unit test or an exact span.
- Enough varied examples to split by template or source so the held-out cases are genuinely new.
- A reason a prompt change isn't enough. The prompt-only control is always run and always shown.
Eligibility
Who can publish
- A Retoque account and an organization you publish under.
- A supported base model. Patches target one exact revision on the models page.
- Free patches can be submitted by any account and still go through review.
- Paid publication is limited to verified publishers during Genesis, while the review process is new.
Submission and review
Six gates between your data and the catalog.
- 01
Data validation
Malformed conversations, prompts aimed at the training infrastructure, secrets and personal data are rejected before any GPU time is spent.
- 02
Training
LoRA training on your train split only, with a fixed seed. Hyperparameters, dataset digest and hardware are recorded in the manifest.
- 03
Held-out evaluation
Base, prompt-only, patched and rollback arms in one harness, plus the regression suite. You declare pass thresholds before the held-out run.
- 04
Submission
You submit a version for review. Nothing is public until a reviewer approves it.
- 05
Review
A reviewer checks your rights declaration, the base model license, the regression list and the evidence state the packager assigned.
- 06
Publication
An approved version is signed, listed with its full evidence, and becomes immutable. Fixes ship as new versions.
A strong evaluation can still be blocked by a rights problem, and a rejected version can be fixed and resubmitted as a new version. Review decisions are recorded alongside the evidence, which is never rewritten.
Rights
Rights you need to have.
Private datasets stay in access-controlled storage. They are never published, never placed on a blockchain, and never shown to buyers. See the dataset policy.
Training and evaluation data
You declare where every dataset came from, under what license, and that you have the right to use it to train and distribute an adapter. Data you scraped without permission, customer data you can't share, and data under terms that forbid model training are not acceptable.
The base model license
Your adapter is a derivative of the base model. You must comply with its license, and your patch license can't grant rights the base license doesn't allow.
Attribution
Third-party material you used must be credited in the package README.
Revenue
Revenue, honestly.
platform fee · not set
Paid licenses are priced per patch version in a stablecoin and settled through the license market contract on Robinhood Chain. Each purchase records the gross amount, the creator payout and the platform fee as separate values.
The platform fee is not set yet. No commission rate, payout schedule or refund window is committed policy. Before you publish a paid version, the studio will show the exact fee, payout timing and refund terms that apply to it, and they are fixed into the version's terms digest.
Free patches are expected to be most of the catalog at first, and a download is not a sale. Nobody should plan income around Retoque yet. License contracts are not deployed in this environment.
A Retoque token, if one ever exists, would be separate. Token trading fees are not patch revenue, and holding a token gives no rights to any patch.