Skip to main content
Adapt-1 updates retained state from observations and outcomes so applications can use what it has learned in later predictions, decisions, explanations, or control commands. The workflows below specify application inputs, returned results, retained evidence, and learner updates.
How workflows update state
  • Context: store or correct records through the State gateway.
  • Prediction: enable the relevant predictor and send complete observations it can admit.
  • Action policies: enable the applicable learner and return outcomes tied to the executed choice and its decision context.
  • Trajectories: configure acquisition or refinement and report measured executions or candidate comparisons.
Each learner needs eligible evidence. A stored record alone does not confirm a predictor or policy update; inspect learner admission and update diagnostics.

​
What you can build

Decisions and workflows

Learn which choices work in context, from routing and recommendations to scheduling and multistep workflows.

Prediction and discovery

Forecast outcomes, classify observations, track changes, and discover causal structure through experiments.

Context and reasoning

Diagnose faults and compare hypotheses using measurements, scene geometry, and retained operational evidence.

Control and interaction

Learn game actions and continuous controls, including manipulation and contextual precision adjustments.

​
Decisions and workflows

InputsIssue and product fields, current queue conditions, and available routes.Adapt-1 returnsA route selected using experience from related requests.Feedback and retained stateResolution, reassignment, and delay revise which routes work in each context.Connect decisions to outcomes
InputsJob attributes, permitted tools or services, and a defined quality, cost, or latency objective.Adapt-1 returnsA tool or service selected for the current job.Feedback and retained stateMeasured results from the tool actually used revise later selections.Define the decision contract
InputsCurrent user or task attributes, eligible options, and a defined response or completion outcome.Adapt-1 returnsA recommended option selected from retained feedback in related contexts.Feedback and retained stateThe response to the presented option updates its value in that context. Keep the option, decision context, and observed outcome linked.Learn from recommendation outcomes
InputsQueue state, available capacity, deadlines, and permitted job or resource choices.Adapt-1 returnsThe next scheduling choice among the options supplied by the application.Feedback and retained stateCompletion, delay, and resource use revise scheduling choices. Sequential learning attributes later outcomes to earlier assignments that changed the operating state.Connect scheduling choices to later outcomes
InputsThe current task state, completed prerequisites, available operations, and the goal and episode boundary.Adapt-1 returnsA next operation selected using retained task experience and the current observation.Feedback and retained stateExecuted operations and their resulting states update the workflow. Delayed or terminal outcomes can revise earlier choices.Build a multistep action workflow
InputsCurrent measured state, permitted actions, a goal, constraints, and a bounded planning horizon.Adapt-1 returnsIn a configured planning workflow, a candidate plan evaluated using learned predictions of successive states. The evaluation accounts for consequences across multiple actions.Feedback and retained stateBefore-and-after observations from the actions actually executed update the transition predictor used by later plans. The application handles execution and returns the measured transitions.The application can execute one action and query again after observing the result. Define the horizon in the Domain’s transition units, including the observation boundary and any other participant’s response.See how planning uses learned transitions or choose the horizon and decision boundary. The Planning guide also covers learned values, adversarial decisions, and control sequences.
InputsCurrent operating measurements, permitted choices, and observed outcomes across changing workloads or equipment conditions.Adapt-1 returnsA decision informed by retained policy state for the current operating conditions.Feedback and retained stateWith change detection and historical-regime matching enabled, Core can recognize a shift and reuse a retained regime when it matches new evidence. Matching and reuse depend on the configured support requirements.Learn across changing conditions
InputsDecision-time market signals, position state, and permitted entry, hold, reduce, or exit choices.Adapt-1 returnsA selected trading policy with its learned outcome support.Feedback and retained stateThe executed choice, position events, and attributed outcome revise later policy selections. The application supplies portfolio limits and order execution.Follow the trading workflow
InputsCurrent failure state, prior attempts, and allowed retries, handlers, or escalations.Adapt-1 returnsA permitted recovery choice based on the observed state.Feedback and retained stateRecovery outcomes revise later choices; delayed credit applies when earlier actions affect the final result.Choose how to learn from recovery outcomes

​
Prediction and discovery

InputsMeasurements available before the result, the value to predict, and completed records from the same process.Adapt-1 returnsA forecast for the declared target, with available support or abstention.Feedback and retained stateObserved values update the predictor. Discovery can identify useful fields and temporal relationships when configured for the task.Build a first predictor
InputsMarket measurements available at prediction time and a defined future value or change to estimate.Adapt-1 returnsA prediction of that observable market target.Feedback and retained stateThe subsequent market observation updates the predictor. A portfolio rule can consume the forecast under the application’s own decision and execution contract.Separate forecasting from policy learning
InputsObservable attributes, the category to predict, and completed examples with their observed labels.Adapt-1 returnsA categorical prediction with supporting evidence or abstention under the configured learner.Feedback and retained stateConfirmed labels revise the classifier. Structure Discovery can learn useful field combinations and rules as evidence accumulates.Configure categorical predictions
InputsOrdered measurements, an event label or observable value to predict, and the application’s alert criteria.Adapt-1 returnsA predicted event class or process value that the application evaluates against its alert criteria.Feedback and retained stateConfirmed event labels or completed measurements update the predictor. Temporal context supplies earlier observations when the current measurement is insufficient.Use history in telemetry decisions
InputsA continuing stream of measurements and observed outcomes, with the relevant run or episode boundaries.Adapt-1 returnsCurrent predictions and induced rules, with learner diagnostics for inspecting how their support changes.Feedback and retained stateNew evidence revises predictions and rule support as conditions shift. Retained structure can be reused when a recurring configuration returns.Inspect changing task structure
InputsOrdered observations of signal positions, bounds, strength, and any other fields required by the interval template.Adapt-1 returnsPersistent interval or scalar estimates under the configured output key.Feedback and retained stateRepeated observations update retained tracks and their extents. The configured persistence and reporting thresholds determine which intervals appear in the output.Configure the adaptive interval policy
InputsPre-trial conditions, permitted settings or interventions, and a declared objective.Adapt-1 returnsA candidate setting or intervention selected for the next trial.Feedback and retained stateYield, error, cost, or another measured outcome revises which choices work under those conditions.Connect trials to retained learning
InputsObservable variables or object-presence fields, permitted interventions, measured outcomes, and a defined intervention protocol.Adapt-1 returnsLearned causal structure and predictions of intervention effects under the configured causal and transition learners. Define the causal result being evaluated: graph edges and numerical relationships in a CausaLab-style setup, or object roles in an IVRE-style task. Specify how model outputs or hypotheses produce the required causal result.Feedback and retained stateIntervention records revise the learned structure and predictions. The experiment design determines which causes can be identified. When Adapt-1 also chooses experiments, evaluate whether those choices help distinguish the remaining causal explanations.Distinguish causal questions and configure the learner

​
Context and reasoning

Use confirmed findings, scene observations, and retained operational evidence to produce a defined reasoning result.
InputsOrdered scalar measurements, episode and step identifiers, a permitted history bound, and the prediction target or available choices.Adapt-1 returnsA prediction or decision informed by recent history. For example, the same equipment temperature can call for a different response depending on whether it has been rising or falling.Feedback and retained stateTemporal context exposes earlier observations to the learner. Outcome evidence teaches which history is useful; configured Structure Discovery can select useful temporal lags. Sequential learning separately assigns later outcomes to earlier actions.Use temporal context and discover useful history
InputsSymptoms, measurements, confirmed findings, and candidate explanations configured in a Domain.Adapt-1 returnsRanked diagnoses, supporting evidence, and missing information under the Domain’s hypothesis contract.Feedback and retained stateConfirmed findings and corrections revise which explanations remain supported.Inspect diagnostic evidence
InputsCandidate hypotheses with supporting and falsifying conditions, declared consequences, and retained observations.Adapt-1 returnsRanked hypotheses with matched conditions, supporting record IDs, falsifying evidence, and selection or abstention. For authored observation hypotheses, consequence text comes from the supplied predicts strings.Evidence and retained stateNew measurements supply evidence for later evaluations. Keep each result linked to the records it evaluated. Preserve intervention evidence when testing causal hypotheses.Inspect recorded evidence and define the hypothesis workflow.
InputsScene features, object identities, coordinate frames, and the spatial relation to evaluate.Adapt-1 returnsA task-defined relationship between objects or scene elements.Feedback and retained stateIn a Domain workflow, corrections or measured outcomes revise later spatial results.Define a spatial reasoning task
InputsObservable object and scene features, coordinate conventions, and the required coordinate or pose format.Adapt-1 returnsA target coordinate, pose, or placement proposal matching the task’s output geometry.Feedback and retained stateReference corrections or measured execution outcomes revise later proposals under the selected learning contract.Choose a structured spatial output
InputsOperational records with their subject and source, followed by new findings and corrections.Adapt-1 returnsRelevant retained evidence and optional structured reasoning about its relationships to the current request.Evidence and retained stateNew records extend the available evidence. Generic corrections replace superseded content under the same record ID. Preserve the original source and store a separate attributed correction record to keep the correction history.Maintain operational context with State and preserve correction history.
InputsA saved decision ID, its complete original response, and the source records behind its evidence.Adapt-1 returnsA sealed explanation containing the earlier selection, evidence, and state identity.Evidence and retained stateExplanation by decision ID preserves that decision after later events or rule changes. Repeating its question evaluates the current state.Inspect a saved decision.

​
Control and interaction

InputsObservable game state, permitted actions or macros, and episode boundaries.Adapt-1 returnsThe next action from the current observation, or a numeric control block under its own contract.Feedback and retained stateDirect outcomes revise individual choices; later outcomes can assign credit across state-changing actions.Choose the observation and action loop
InputsMeasured object and robot state, handling goals, bounded controls, and a fixed executor.Adapt-1 returnsA sequence of control commands for the handling task.Feedback and retained stateMeasured outcomes support sequence acquisition and revision. Retained behavior can be refined for changes in the starting conditions.Choose a manipulation control interface
InputsMeasured state, a goal, bounded continuous controls, command timing, and an executor.Adapt-1 returnsA numeric control sequence committed for an attempt or execution window.Feedback and retained stateActual controls, measured states, and outcomes update retained sequences. Candidate comparisons support selection, reduction, and ordering.Acquire a continuous-control sequence
InputsA retained action or sequence, current measured conditions, and bounds on the permitted adjustments.Adapt-1 returnsContext-dependent corrections to the acquired control behavior.Feedback and retained stateMeasured candidate outcomes teach the correction. Evaluate the retained refinement on new conditions with learning stopped.Learn contextual corrections
For sequence-acquisition deployment status, see Mechanism maturity. The integration supplies observations and permitted operations. It also handles parsing, sensing, actuator mapping, and outcome measurement when those are part of the task. For independent image questions and grounded pointing, Spatial Vision provides a stateless route. The spatial Domain examples above use their own structured observations and feedback.

​
Follow an application through the loop

​
Trading research

Trading research can use transition estimation or action-policy learning according to the required result. For transition estimation, query before the next value is observed, then ingest the completed transition. The retained predictor can change as new records arrive. Use transition learning or Transition Discovery. An external portfolio rule can consume that prediction. In that setup, Adapt-1 learns the transition predictor and the application defines the portfolio rule. For policy learning, preserve the returned decision and its pre-action context. Record the action actually taken, any override, and subsequent position and exit events. Return the measured outcome to the correct decision. Use contextual feedback for an attributable choice, or sequential learning when entry and position actions affect later state and a later outcome should revise them. Keep portfolio limits, permissions, order handling, and execution in the surrounding application. In a paper or shadow workflow, distinguish simulated fills from real execution and include the chosen cost assumptions in the outcome. Check later time periods with learning frozen before interpreting a change as transferable behavior.

​
Scientific experimentation

A laboratory or process-development tool can use measured trial outcomes to update subsequent experiment choices.
  1. Describe the trial context. Supply the material, batch, instrument state, or other pre-trial conditions, along with permitted settings.
  2. Request the useful result. Ask for a predicted output, a trial choice, or a structured explanation. These use different learning relationships.
  3. Run and measure. Record the intervention actually performed and its measured yield, error, duration, or other declared outcome.
  4. Retain the change. Update the corresponding predictor, choice policy, or hypothesis support, then inspect the next result.
A prediction target describes what occurred; a reward describes how useful the chosen trial was. Keep those meanings explicit when one experiment supplies both. For causal discovery, use intervention records to test competing explanations and reserve additional measurements for checking their predictions. The guide to learning from evidence and outcomes covers experimental optimization and causal discovery. Learning schedules explains when to retain a shared history and when a newly drawn hidden process needs its own learning scope.

​
Interactive control

Choose when Adapt-1 receives a new observation and what it is allowed to return. For example, a robot interface can accept position, orientation, and gripper commands, while fixed inverse kinematics and motor controllers execute them. Adapt-1 acquires which commands to use and how to coordinate them. Positioning or aiming tasks can use different control widths and the same measured-outcome loop. Evaluate learned controls in the environment where your application will use them, including the equipment and operating conditions for physical integrations. The configuration guide defines the observation, goal, action, and timing contract.

​
Service request routing

For support routing, supply the request and queue conditions at decision time. Record the queue actually used and attribute its resolution, reassignment, or delay to that decision. Retained feedback can revise later route choices under related conditions. Use the Domain guide for the task contract and the feedback guide for attribution.

​
Choose the integration pattern

Combined workflows need compatible state scopes and explicit targets and feedback meanings. Choose a workflow compares the request and result contracts.
Configured Discovery paths can form eligible inputs, field combinations, useful temporal lags, predictive rules, and action or intervention bindings from evidence. These parts of the learner structure can begin unresolved.The application supplies structured observations, permitted operations, and the target or outcome meaning. It defines the business objective, episode boundaries where needed, and the integration boundaries for sensing, execution, and measurement.Choose what to declare
Choose which users or agents should contribute to the same retained experience.Agents share experience when the integration deliberately routes them to the same authorized scope. Use separate Domain IDs for independent Domain workflows and separate authenticated identities for independent State histories. Changing the body session_id does not create isolation.Identity and task scope

​
Evaluate your integration

Choose one useful result and a comparison you already understand: a fixed route, an existing predictor, or an unchanged routine. Collect the observations and outcomes needed by that relationship, then check whether retained state changes later results on separate cases. For a frozen evaluation, stop learning writes and verify that the applicable learner state stays unchanged. Inspect support, abstention, and the operation actually executed alongside the outcome.

Verify adaptation

Connect a learning update to a later result and the operation actually used.

Choose a learning schedule

Decide when experience is collected, when it is reused, and when learning can continue.