The objects in your integration
Creating a new environment episode does not reset a Domain. Reusing a Domain carries its applicable state forward until an explicit lifecycle operation changes it.
Decide what teaches the result
Completed observations can teach a predictor without a reward. Preserve decision-time context and the executed policy when the learner should instead revise action values. Learning relationships covers the full set of configurations.
Structure and learning schedule
Discovery forms eligible learner-facing structure from evidence. Authored structure makes selected task relationships explicit before use. Either can start empty, use previously acquired state, or keep learning during operation. Choose the schedule separately: define the starting state and whether writes are allowed during the run. During frozen use, observations and returned actions can still change while the applicable learner state stays fixed. See Learning schedules.Inspect what happened
Check three things: the request succeeded, the intended learner admitted the event or feedback, and the later result used the applicable state. A stored record and a learning update are separate signals. Use the route’s result status, evidence, and learning diagnostics. Confidence fields have route-specific meanings; use the documented support and abstention behavior when deciding whether to act. See Explain a result.Contract and application boundaries
What stays stable, and what changes?
What stays stable, and what changes?
The public task meanings, event schema, result contract, available actions, outcome semantics, and boundaries stay stable within the declared workflow. Supporting observations, discovered input and representation structure, learned relationships, active alternatives, policy or transition state, uncertainty, and support can change through eligible writes.State can form during a zero-start run or a separate acquisition phase. Decide whether later use permits further learning or only reads retained state. See Choose a learning setup.
What belongs outside Adapt-1?
What belongs outside Adapt-1?
The external environment, perception or format translation, candidate generation where required, final commitment, execution, safety checks, and consequence measurement belong to the application. An adapter should expose only information available at that point, not inject the preferred answer or hidden mapping the learner is expected to form.Keep the returned result, the application’s committed value, and the operation actually executed separately when validation or overrides can change them.
Should I use one Domain or several?
Should I use one Domain or several?
Reuse a Domain when public meanings, result units and schema, consequence semantics, and state lifecycle remain compatible. Use separate Domains when those boundaries differ, even if observations come from the same product or dataset. Each Domain operation addresses a
domain_id under the authenticated owner.Current observations and external state are not retained Domain state. State-gateway continuity is another scope for evidence and constraints when a Domain contract is unnecessary.Which learning relationship do I need?
Which learning relationship do I need?
Predicting an observable consequence, judging a committed decision’s utility, scoring movement toward a bounded objective, assigning delayed credit, and comparing typed hypotheses are different relationships. A Domain can combine compatible relationships, but they do not become interchangeable because they all return JSON.Choose how a Domain learns describes the distinctions. Explain a result describes support, uncertainty, and learner diagnostics.