POST/twinsmember
Build a twin loop: a twin of a real system (domain weather: station temperature from a global AI weather model against live METAR readings), its subjects and its parameters with their uncertainty (the prior).
Body
name | 1..64 letters, digits, _ . - |
domain | weather |
config | {points: [{id, lat, lon}], model?, variable?, offset?, hours?, leads_h?, members?, issue_delay_h?, embargo_h?, readings?, baseline?, station_bias?, forget?, pool?} |
params | [{name, value, sd}] to replace the prior's |
GET/twinsmember
The project's twin loops.
GET/twins/:idmember
A twin loop: its config, current parameters and band, every cycle with its scores, and every calibration with its fit and held-out check.
POST/twins/:id/readingsmember
Readings as they arrived, pushed or pulled from the domain's live source; each keeps when it is for and when it arrived, and a revision is a new reading.
Body
readings | [{subject, at, value, arrived?}] |
pull | {from, to}: from the domain's source (at most 40 days) |
source | a name for pushed readings |
GET/twins/:id/readingsmember
The readings a twin holds, newest first, each with its subject, the moment it is for, when it arrived and where it came from (a name for pushed readings, the domain's source, or physical:<level>:<sensor> for an admitted measurement).
Query
subject | one subject |
limit | 1..500; default 100 |
POST/twins/:id/measurementsmember
Typed measurements from a verified sensor, held against the twin's solved band by the same admission as POST /physical/measurements/admissions: inform always, validate and recalibrate only inside the band, at or above the floor, from a sensor that reads the surroundings (placement environment), recalibrate only when measured and not inferred. A weather twin stores as a reading the temperature in celsius that may validate, and nothing else; the answer says per measurement what it may be used for and whether it was stored. A refused measurement is reported with its code and the others go on.
Body
measurements | [{subject: one of the twin's points, at: when the value is for, measurement: {v: 1, kind, value, unit, attestation, modality?, placement?, sensor_id?}}], 1..1000 |
floor | lowest attestation level accepted; default soft |
POST/twins/:id/cyclesmember
Run the twin forward from an origin on Ipnops Compute, with the baseline beside it on the readings as they stood when the cycle was issued. The parameters are the latest calibration whose data had all arrived by then.
Body
origin | a time on a 6-hour boundary |
POST/twins/:id/cycles/:cycle/collectmember
Read a cycle's jobs back when they end (read every few seconds, up to wait_s) and store every forecast: the twin's, its uncalibrated prior's and the baseline's, each marked validated or not and why. A live cycle's forecasts are sealed in the accuracy ledger.
Body
wait_s | 1..7200, default 600 |
POST/twins/:id/cycles/:cycle/scoremember
Score a cycle's forecasts against the readings that have arrived: CRPS, log score, PIT and coverage per method on shared cases, and paired skill of the twin against the baseline and against its own prior. Each forecast keeps the reading it was scored against.
POST/twins/:id/recalibratemember
Update the twin's parameters from its scored cycles not yet used (POST /calibrate/recalibrate on each forecast's own design), the current parameters as the prior; cycles issued after the data's last arrival use the result.
Body
cycles | only these cycle ids |
holdout | share of the latest observations held out, default 0.2 |