Skip to main content

Models and their tiers

Five model families price your passes, each answering one question about a link. All five are served: Link SNR, Weather attenuation, RF interference and SSA pass forecast to every self-serve account, and Satellite demand from Operator up. In a workflow, the demand, rfi and ssa lanes need Operator. In a workflow, every value a model gives carries a range, not a single number: p50 is the expected value, and p10 to p90 the range 8 in 10 land inside. See Reading a result.

The five families​

FamilyWorkflow lanePredictsServed today
Link SNRsnrA link's signal-to-noise ratio at a time, in dBYes
Weather attenuationweatherRain, gas, cloud and scintillation loss on a link, in dBYes
Satellite demanddemandThe share of a pass your satellite can expect an antenna for (p_free), and how many satellites compete for itYes
RF interferencerfiThe SNR a link loses to interference, the chance of interference and what that chance rests on, and the Sun's and Moon's part of itYes
SSA pass forecastssaA pass's peak elevation and slant range with the uncertainty of the orbit, and the chance the antenna holds the satelliteYes

"Served today" is the model catalog's status, the one GET /predictions answers by. Whether a deployment serves each workflow lane is a separate answer: lanes in GET /predictions/workflows?templates=true.

GET /predictions answers each family in its own shape, and Link SNR there gives the median only. See each model card for its inputs, outputs, declines and limits.

Physics-only, pretrained and fine-tuned​

These three words say what a forecast is built from.

KindBuilt fromWhere it applies today
Physics-onlyPhysics alone: the link budget, ITU-R propagation models and orbit geometryOn Explore, a run's snr lane and SNR from GET /predictions (the link budget alone, marked reason: link_budget_only). A run's snr source reads physics:clear-sky-budget; where the SNR model serves your deployment, it adds why it was not asked: :plan_required:learned_corrections.
PretrainedA served model, the same release for every account; Link SNR is trained on link telemetryRuns on every plan ask the served weather, demand, rfi and ssa models; on Operator and above they also ask the SNR model instead of the link budget alone (the Platform calls this learned corrections). Its breakdown's sources read model:<family>:<release>. A step the model could not price keeps its physics value, and its source says why after the baseline's, as in physics:clear-sky-budget:unavailable:invalid_answer or climatology:unavailable:no_answer.
Fine-tunedA model fitted to your own telemetry, per antennaThe nightly calibration the SSA pass forecast and RF interference cards describe, fitted from your own telemetry on Operator and above. Explore uses the same models without it.

The API does not yet report which of the three each model runs as for your account: GET /catalog/models carries no such field. Read it from a result's breakdown, as above.

The Platform shows the same thing. Under each result, the version line reads Priced <time> · details. details lists the workflow and its version, the network version, and each lane's release, such as snr: <release>, or Physics only on Explore. An amber chip shows on the line only where a lane fell back off its model, with its reasons in the tooltip. Passes that carry the demo fleet's planted rain show a neutral Demo rain chip instead: that rain is part of the demo, not a fallback, and it never reaches your own stations. On Forecasts, under Models, a model's detail shows that release as Release, read from the latest run of your live workflow.

Bronze, Silver and Gold​

The model catalog labels each family with a fidelity tier: how refined the data and evaluation behind its served release are.

FamilyTier the catalog reports
Link SNRSilver
Weather attenuationSilver
Satellite demandSilver
RF interferenceBronze
SSA pass forecastNone yet

A tier labels how a model is served, not what a plan includes: every plan that asks a model calls the same release (Explore prices SNR with the link budget alone, above), and GET /predictions takes no tier parameter. Further Silver releases, and Gold, reach Enterprise deployments as they are released. See Bronze, Silver and Gold.

  • Research
  • Accuracy: how forecasts are checked against what passes delivered
  • Workflows: how lanes combine into the numbers you act on