Train

From approved data to a promotable model, across the training API surfaces:

Page	Surface	What it covers
Datasets	POST /datasets	point-in-time builds, windows, splits, manifests, leakage aborts
Labels	POST /labels	source kinds, maturation, stats, immutability rules
Training studies	POST /training-studies	the default way to train: a fixed grid of runs, a validation-only winner, one evaluation, the fallback
Training runs	POST /runs	one configuration: tiers and overrides, metrics, cancellation, evaluation gates
Event models	POST /event-labels, /event-datasets, /event-models	scorers of single events from their own fields: event labels, point-in-time builds, training, gates, bundles and activation (scoring follows with WP-V17)
Entity baselines	POST /entity-baseline-models	a simple fallback model of a dataset when no sequence model qualifies: admission, candidates, gates, bundles, batch scores
Promotion gates	GET /gate-definitions, GET/PUT /promotion-policy	gate facts, severity, acknowledgement, audit provenance
Model cards	GET /models/{id}/card	verified lineage, training, evaluation, promotion, and limitations evidence
Governance evidence	GET /models/{id}/card:export, POST /dossiers, GET /orgs/{id}/model-inventory	authored governance fields, pinned card exports, review dossiers, the organization inventory
Two-person approval	PUT /dual-control-policy, POST /dual-control-requests:approve	a second person for acknowledged promotions, contract approvals and fingerprint overrides
Baseline evaluations	POST /evaluations:baseline	the pre-registered bar from the internal quality safeguards and your comparators, before any model
Proof center	POST /protocols, POST /evaluations, POST /backtests	declaring the bar in advance, judging a dataset build against it, and lead-time backtests
Model file retention	GET/PUT /model-file-retention, /model-files	deleting archived models' files whole: retention, explicit requests and their cancel, holds, the drain

The lifecycle: labels declare what to predict → datasets assemble supervised, leakage-gated training data → a study (or one run) trains and ends in an evaluated model version → models allowed by the server-resolved promotion verdict become bundles you activate and serve. Advisory quality failures require an explicit, audited acknowledgement; integrity failures always block. Each evaluated model's model card assembles that evidence into one traceable, current-state dossier. The quickstart walks the whole path on sample data.