Skip to main content
Adapt-1 is the direct programmatic surface of Core. An application can send observations and outcomes, receive structured results, and let later calls use the state formed during earlier work. Default path: choose the route that returns the result your application needs and use the service defaults. No generated language, task-specific training dataset, or learner tuning is required to begin.

​
One operating loop

Not every workflow uses every step. A read-only query does not need an external action. A stateless result does not retain adaptation. A continuing investigation may accumulate evidence without action selection. The loop describes responsibilities, not a mandatory pipeline.

​
The shortest mental model

Five nouns explain most Domain workflows:
A current observation is external information available now. Retained Domain state is evidence-shaped state Core may reuse inside a Domain. State gateway continuity refers to retained evidence and constraints when a Domain contract is unnecessary. External state belongs to the external system and remains outside Core. An event is one record; an episode is an application-defined boundary that can contain many events; one Domain can span many compatible episodes.
A Domain tells Core what information means. Events and outcomes provide the evidence from which task-specific state can form. A query asks for one current result. The application decides how to use that result and returns any eligible consequence.

​
The application boundary

The external system, the application, and Core have different responsibilities.
01 · External source
Environment or data source

Owns the external state and exposes the current observation available to the application.

02 · Application boundary
Application or optional harness

Maps public inputs, supplies available choices, commits the value actually used, executes externally, validates, and measures consequences.

03 · Structured Core call
Core

Interprets the request inside the selected contract and retained state, then returns the contract-defined result and applicable evidence or diagnostics.

04 · Return path
Consequence and later call

The external consequence remains outside Core. The application can return an eligible event, outcome, correction, or feedback signal that later calls may use.

​
What stays stable and what changes

The public contract does not require a fully authored ontology. With Discovery, input projections, fields, combinations, temporal relationships, state-dependent action values, and other learner-facing structure can be formed from evidence. A Domain can still name available actions without naming the correct action, define a coordinate system without supplying target coordinates, and define a successful native outcome without supplying the hidden action-effect mapping that produces success.

​
What the application owns

An adapter may translate an image, sensor packet, or external record into the selected request schema. It should expose only information available at that point in the workflow. It should not insert the preferred answer, a hidden target, or the consequence relation Core is expected to form.

​
Domains and multiple tasks

One Core instance can retain several distinct Domains. Every Domain operation addresses one domain_id, so their contracts, evidence, and learned state do not blend. Reuse one Domain when episodes share the same public meanings, result contract, consequence semantics, and state lifecycle. Use separate Domains when those meanings or boundaries differ, even when the inputs originate from the same product or dataset.

​
API gateways

Adapt-1 exposes several gateways into the same substrate: These are public access paths, not separate internal architectures. An application can use one gateway or compose compatible workflows.

​
Learning relationships are not task labels

A Domain can support more than one evidence relationship. Keep these questions separate:
  • What is supported by the evidence now?
  • What observable state follows an input or action?
  • How would a candidate move a bounded public state toward one objective?
  • Was a committed choice useful?
  • Should a later outcome revise earlier state-changing choices?
  • Which typed mechanism best explains the observations?
The answers may use ordinary Domain reasoning, transition learning or Transition Discovery, Structure Discovery, bounded transition scoring, policy feedback, sequential learning, typed hypotheses, or a stateless route. The Domain learning guide explains the distinctions.

​
Language boundary

Core does not need to turn its work into prose. Adapt-1 can return contract-specific values such as ranked hypotheses, predicted observations, policy scores, intervals, coordinates, support records, conflicts, and missing evidence. Application code can consume those results directly. Adapt-1 and Rei Chat are separate products. A language layer can articulate a completed Core result when prose is useful, but generated language is not required for the underlying workflow.

​
Start here

Choose a workflow

Decide whether the result is stateless or should use a continuing Domain.

Choose a learning setup

Distinguish zero-start learning from state earned in a separate acquisition phase.

Use Discovery

Let Adapt-1 form learner-facing structure from admitted evidence instead of authoring the ontology first.

Design an authored Domain ontology

Define explicit reviewed task-facing structure when you need it.

Adapt-1 quickstart

Authenticate and make the first direct call.

Explain a result

Inspect support, missing evidence, conflicts, abstention, and learner state.