Building with Core

Ways to connect Core to products, tools, and applications.

Core can be used across products, tools, and applications. The surrounding application or integration decides how information enters, how results are used, and what actions follow.

The examples below illustrate common ways to connect Core. They are not limits on what Core or Adapt-1 can support.

The application boundary

Input boundary

External systems

Records, sensors, parsed content, events, or interactions supplied through the chosen interface

→

Reasoning substrate

Core

The continuing instance forms and revises task-relevant structure

→

Result boundary

Application

Inspect, validate, display, route, or execute the current result and its supporting evidence

This is the application contract around Core, not an ordering of Core’s internal processes. Supported workflows can return structured outcome feedback through the same declared boundary, while application-side validation and outcome measurement remain outside Core.

Applications can supply input directly through the selected interface. When a source format differs from a route’s request contract, an optional adapter can translate the boundary format; it should not contain the task solution. Core has no native visual component. A detector or visual encoder is an upstream observation source, not part of Core.

A fair adapter may expose literal or deterministic facts: coordinates, colors, tracked components, available operations, exact deltas, or public action meanings. It should not label the current goal, choose the action, or encode the episode’s hidden mapping. Record the adapter as part of the system being evaluated.

Recurring application patterns

Pattern	What must persist	Typical Core-facing result
Hidden-state monitoring	Object identity, recurrence, activation, and uncertainty	Structured current-state prediction
Domain reasoning	Entities, relations, hypotheses, rules, and missing evidence	Ranked hypotheses and evidence packet
Outcome adaptation	Which relation or policy worked under which conditions	Updated scoped preference or policy score
Long-running personalization	Constraints, corrections, durable preferences, and fading weak signals	Context and structured reasoning for later work

These are application patterns, not product boundaries. A single deployment can combine them.

Domain contract

In Adapt-1, a domain declares the stable grammar of a task: object and relation types, signals, event and native outcome semantics, action or plan interfaces, candidate hypotheses, rules, and the expected result shape.

It does not contain the current episode’s hidden mapping, target coordinates, correct procedure, or solution path. Structure still has to form from the observations and outcomes presented during operation. This lets a fresh domain begin without accumulated task history while keeping public interface semantics separate from the answer.

For action-producing tasks, preserve state -> committed decision -> exact action -> transition -> native outcome as an observed trace. Scoped feedback can assign credit to the intended relation/policy; a reusable causal mechanism remains a stronger claim that requires later predictive evidence.

State revision

Early state	Revision	Later work begins with
Candidate structure has partial support	Repeated evidence strengthens it	The supported structure remains active
Several interpretations remain plausible	Their associations stay active together	Multiple alternatives remain available
A weak signal is present	Weak support loses influence	The signal has reduced weight

Several interpretations can compete until the current result settles provisionally; later observations then meet an already changed state.

Persistent state can also preserve incorrect structure. Use isolation, reset controls, ordered event handling, evidence review, and explicit feedback when operating stateful workflows.

Structured outputs

Depending on the domain and gateway, an application can consume:

• state estimates and structured predictions;
• ranked hypotheses, policies, assignments, or other domain-shaped results;
• evidence, conflicts, missing information, and confidence;
• inspectable state changes for later review.

Natural language can be added after that result exists. Rei Chat demonstrates the articulation boundary; Adapt-1 exposes the structured contract directly.

Interfaces

Adapt-1

Direct access to Core for stateful reasoning, adaptive behavior, evidence, and machine-actionable outputs.

Rei Chat

A conversational demo where a language model articulates Core’s structured result.