Concepts

Tessera trains a behavioral model on your own event stream: one embedding per entity, optional prediction heads, and low-latency scoring APIs. The platform never assumes your business domain — everything domain-specific (entity types, event types, attribute meanings, labels) is data you declare. These pages teach the six ideas the whole product is built on, each with a worked example from a different industry:

Page	Idea	Worked example
Entities and events	Everything is an entity emitting typed events	Gaming sessions
Schema contracts	You approve exactly what the model may see	Healthcare operations
Presence	Absence is information, never a silent zero	Logistics shipments
Labels and masking	Labels are targets, never inputs	Streaming subscription churn
Leakage gates	Builds fail closed when the future leaks in	E-commerce
Fingerprints and reproducibility	Every artifact is content-addressed	—

Glossary

Entity

Anything in your world that accumulates history: a player, an account, a device, a shipment, a patient visit. You declare entity types; Tessera keeps one chronological event ledger — and ultimately one embedding — per entity. See Entities and events.

Event

One typed, timestamped fact about an entity — the canonical record {entity, event_type, timestamp, attributes} that every connector, ingestion call and upload normalizes into. See Entities and events.

Schema contract

The versioned, fingerprinted document that scopes model visibility: which event types are included and, per attribute, its declared type, its role (feature, target_only, identifier, metadata, excluded) and encoding. An attribute you never reviewed is never a model input. See Schema contracts.

Presence

The five-state resolution of every contracted feature attribute on every event — present, missing, unknown, corrupted, not_applicable — so the model sees explicit presence tokens instead of silent zero-fills. See Presence.

Label

A customer-defined training target: a binary, multiclass, or regression outcome attached to an entity at the moment it became known (observed_at). Labels are structurally targets, never inputs; unobserved is never treated as negative. See Labels and masking.

Maturation

The waiting period (maturation_days) before an outcome is trusted for supervision: a label observation only trains model states that are at least that many days older than the observation itself. See Labels and masking.

Leakage gate

The build-aborting checks that keep information from the future — or from the label itself — out of model inputs: deny name patterns, per-attribute denies, identifier value scanning, and the temporal gate. Gates fail closed: a violating dataset build aborts with a typed report instead of shipping a poisoned model. See Leakage gates.

Fingerprint

The sha256 content address (over canonical JSON) that identifies every immutable artifact: contract versions, leakage policies, dataset builds, encoder sets, checkpoints, evaluation reports, promotion bundles. Loaders re-hash on read and refuse to serve on mismatch. See Fingerprints and reproducibility.

Deployment slot

The stable, API-addressable unit scores are served from (prod, shadow-1, …). Bundles rotate under a slot; the slot name never changes. See Serve.

Promotion bundle

The immutable, fingerprint-pinned package of everything serving needs — model checkpoint, contract version, leakage policy, encoders, tokenizer version, evaluation report — created when no gate is blocking and every failed advisory quality gate has been explicitly acknowledged under project policy. The measured gate block remains verbatim; acknowledgement provenance is additive. See Promotion gates and Fingerprints and reproducibility.