Reading a result: ranges, versions and reasons
Every number Constellation gives you has three parts: what to expect, how far it could land from that, and what produced it. When it cannot give a number, it says why instead.
Expected, the range and the low case
The Platform's screens never print p10, p50 or p90. They use three words, and the API keeps its names:
| On screen | In the API | Means |
|---|---|---|
| Expected | p50 | The middle forecast: as likely to come in above as below |
| Range | p10 to p90 | 8 in 10 land inside it, 1 in 10 below and 1 in 10 above |
| Low case | p10 | The 1 in 10 bad outcome: only 1 in 10 come in lower |
A pass shown as "Expected 39.6 GB, 38.6 to 40.6 GB" (illustrative values) is this in the API:
"value": { "p10": 38.6, "p50": 39.6, "p90": 40.6 }
Risk is judged in the low case. A pass is under your contract minimum when its low case is, and a recommendation's gain (gain_gb_p10) is what it keeps in the low case.
In energy and wind yield, "P90" means the conservative value, exceeded 9 times in 10. The API's p90 is the reverse, the high end of the range; its conservative value is p10. That is why the screens use words instead.
Totals
A total's expected value is the sum of its passes' expected values. Its range is not the sum of their ranges: adding every pass's low case assumes they all fall short at once, which is rarer than 1 in 10. So Passes shows a total's expected GB, and under it the floor "if every pass has its low case", which the total is very likely above.
Versions
Every result says what produced it, so a number can be traced and reproduced. The line under it reads Priced <time> · details; details opens the rest:
| On screen | In the API |
|---|---|
| Priced at | The run's created_at |
In details: workflow name and version, such as Pass yield v3 | workflow and workflow_version on the run, and workflow_version on each item |
In details: network config version, such as network v6 | network_config_version on the run |
In details: each model lane's release, such as snr: <release>, or Physics only on Explore | Each breakdown entry's source: model:<family>:<release>, physics:clear-sky-budget or climatology, with :plan_required:learned_corrections after it where the model was not asked |
| An amber chip on the line, only where a lane fell back off its model, with its reasons in the tooltip | The same source with :unavailable:<code> after it where the model did not answer |
| A neutral Demo rain chip, where passes carry the demo fleet's planted rain; it is not a fallback | A weather source, or a step_sources entry, of simulated:demo-rain-cell |
A single forecast from GET /predictions carries model_release_id and predicted_at instead. Keep these with anything you export.
Reasons
Each pass's Why on Passes is its ledger, explain in the API: every step's value, unit and source. A source that starts with unavailable:, such as unavailable:invalid_answer, means that step could not price the pass; the Platform shows it as Not priced. See Models and their tiers for every source. See Items.
Declines
Where Constellation cannot forecast something, it declines it with a code rather than guess:
GET /predictionsanswersentity_errorsbeside its items, each with theentity_idit is about and acode, such asbelow_min_elevation. A200can carry both. Branch oncode, notmessage. See Entity errors.- A workflow run lists the antenna and satellite pairs it left out in
declined, each with its reason. - Each model card lists its codes under Declines, for example Link SNR.
Related
- Data modes: the Live, Audit, Sim, Scenario or Demo fleet label beside a view's title
- Accuracy: how often results land in their range
- Glossary: the Platform's, the API's and the industry's names side by side
Next: Recommendations and delivery.