The registry
Build your own model
Combine a dataset, a measurement, a transform and a payout rule, with no code. Every model is checked before it is created, and a published version never changes.
Backdrop: a gigantic jet above a thunderstorm, seen from the ISS, July 3, 2025. NASA, public domain.
Bangkok 3-day excess rainfall · 1.0.0
- Feedera5-rain-bangkok @ 2026-08-11
- Measurementdaily_precipitation_mm
- Transformrolling_sum over 3 days
- Payoutstep: 100 mm pays 50%, 150 mm pays 100%
- Model Validated, private until published
Validation
Every model is checked when it is created
If a check fails, the model is not created and you are told why.
- Schema conformanceAgainst the published model schema.
- DeterminismEach model is run twice and the outputs are compared byte for byte.
- A full backtestOver every complete year in the dataset’s record.
- Sanity checksFlags rules that would always pay or never pay.
- A model cardGenerated automatically, listing assumptions, limitations, license and author.
Status
Your models stay private
Models you build are visible only to your account.
Registry routes
- GET /primitivesLists the building blocks you can use Serving
- POST /modelsCompose and validate a model Serving
- POST /models/{model_id}/runsRun it against an object Serving
- Public listingA public list of published models and their measurements, without rankings Not available yet
Versions and billing
A published model never changes. A change is a new version, and every result is labeled with the model and the version that produced it.
Model runs are billed as usage. The person who runs a model pays. Authors are not charged when others use their models.
Models only read versioned datasets, so results are reproducible.
Compose your first model
It takes about fifteen minutes in a terminal, and you do not need an SDK.