predict

Record a falsifiable prediction so it can be resolved later and fed into calibration.

Past predictions become the foundation of assess and calibration: an instinct earns trust by being measured.

Signature

predict({
  prediction: string,
  domain: string,
  confidence?: number,
  basis?: string
}) → { id: number, domain: string, confidence: number }
ParamTypeRequiredDescription
predictionstringyesWhat you expect to happen. State it falsifiably.
domainstringyesShort label that groups predictions for calibration.
confidencenumber 0..1noHow sure you are. Default 0.5.
basisstringnoWhy you think this. Useful when you later review surprises.

Example

curl -X POST https://mneva.dev/v1/predict \
  -H "x-mneva-key: $MNEVA_KEY" \
  -H "content-type: application/json" \
  -d '{
    "prediction": "the migration will need a hotfix",
    "domain": "schema",
    "confidence": 0.4,
    "basis": "the v2 backfill is doing two passes — easy to miss a NULL"
  }'

Response:

{ "id": 2, "domain": "schema", "confidence": 0.4 }

When the migration runs, resolve the prediction with resolve.

Picking a domain

Domain is the calibration key. calibration returns per-domain accuracy and confidence gap; assess matches on domain to find adjacent predictions for the track record block.

Good domains are nouny and re-usable:

  • deploy
  • schema
  • user-flow
  • auth
  • rate-limits
  • migration
  • pricing

Avoid one-off domain names — deploy-2026-05-22 would never accumulate enough resolved predictions to be useful. Same domain, repeated, lets calibration build a track record over time.

A given prediction can only have one domain. If a prediction crosses two areas, pick the one most likely to repeat.

See also

Was this page helpful?