Docs·Concepts·Calibration

How Goable calibrates to your spots

Physics gives you a starting score; operator outcomes bend it to reality. Any spot in Goable can graduate from a generic base profile to a calibrated curve once enough outcomes accumulate. Calibration is evidence-gated, not an automatic refit, and it depends on operator outcome data that is at design-partner stage today.

Maturity note: outcome collection and the calibration workflows described below are being built with design partners and are not yet generally live. The cadences, thresholds and gates here document how the loop is designed to run, not a claim that it is refitting your traffic today.

Why calibration matters

A generic "kitesurfing" profile knows that 18 knots is workable for an intermediate. But a Tarifa school may run sessions at 16 knots while a windless Catalan cove cancels them. Same physics, different local thresholds. Calibration learns these from submitted outcomes.

The calibration loop

You score a session via /v1/score. The response carries a session_id.
The session runs (or doesn't). You report what happened via POST /v1/score/:id/outcome with outcome_type ran / cancelled / no_show / rescheduled.
A weekly Bayesian refit pipeline trawls the accumulated outcomes per-(activity × cell × dimension) and produces a candidate curve.
A held-out Brier Skill Score validation gate decides whether the new curve beats the previous one. If yes, it ships; if no, the previous curve persists. The rollback gate is designed so operators are not quietly served a worse model.
The drift monitor runs daily on the deployed curve. If skill drops two sigmas below baseline, an emergency recalibration triggers mid-week.

Hierarchical fallback

New spots don't have outcomes yet. The engine uses a 5-level spatial hierarchy (base, region, cluster, sub-spot, micro) with Bayesian shrinkage: your new spot borrows its parent's calibrated curve until it has its own ≥150 paired outcomes. The calibration_provenance block in every score response tells you which level was hit:

Sub-spot resolved with n≥100 outcomes yields hierarchical_calibration = 1.0. Falling all the way back to the base profile yields 0.85. Same scoring engine, fewer guarantees on local accuracy.

Cadence + thresholds

Weekly batch runs Sunday 02:00 UTC. A cell becomes eligible for calibration once it has ≥150 paired outcomes; below that the catalog's prior carries. Each batch ships a markdown audit (visible to ops at /ops/calibration) with the BSS Δ, drift safeguards passed, and the scoring sample size.

Inspecting calibration for your traffic

Tenants on Pro+ can see the verification dashboard at the verification page in your tenant portal: Brier Skill Score plus reliability diagrams stratified by horizon, sub-spot and cluster. The numbers come from the same scoring_audit_log × outcomes join the drift monitor watches.