Choose an execution pattern

Use this guide to choose what one command means and when Adapt-1 receives another observation. Your application defines the control interface, executes returned commands, and measures their consequences.

​

Choose the control interface

Control width and sequence length are separate choices. For example, eight two-coordinate commands contain sixteen scalar values. Set both from the interface your application can execute and measure; see configuration.

​

Check that numeric changes reach the executor

Decode nearby numeric proposals before a long acquisition run. Thresholds, clipping, and actuator resolution can map different values to the same executed control. For an illustrative binary actuator with requests in [0, 1] held over eight ticks, a threshold decoder might turn it on for all eight ticks when the requested value is at least 0.5. A duty-cycle decoder could instead hold it on for round(8 * value) ticks: Duty cycles still have finite resolution. Choose a mapping that suits the actuator and lets useful proposal changes produce distinguishable executions. Preserve the proposal, decoded control, and measured consequence so repeated behavior can be traced to the decoder. Keep the decoder and its version fixed within comparisons. If its meaning changes, check compatibility with retained executions before reuse.

​

Check the execution window

For a fixed command duration, horizon * command_duration gives the maximum unpadded duration of a proposal. Calculate the full execution window from the controls actually applied, including neutral padding, then add any separately defined assessment period. Compare that total with the time the task needs. Record whether an attempt ended because the task completed, the environment terminated, or the command sequence ran out. Repeated progress followed by sequence exhaustion calls for checking the horizon and timing contract before increasing the attempt budget. Choose compatible dimensions and horizon during configuration.

​

Target commands

Use target commands when a fixed controller can execute a requested value while Adapt-1 learns which values to request and how to order them. Request a bounded sequence from the measured initial state and execute it with the declared controller. Return the outcome with the applied commands and measured successor states. Define the coordinate conventions and hold duration before learning. If the executor applies a coordinate transform, validate its measured inputs and preserve the same transform and correction-composition order during comparisons and retained use.

​

Rate commands

Rate commands specify how a controlled value changes over an interval. The starting observation must include the state needed to interpret those rates. Record the actual applied values and duration so the resulting transition can be attributed to the execution. A whole-attempt proposal commits its commands before execution. Contextual refinement can adjust the acquired sequence from the observation available at that boundary. Keep actuator response, bounds, and timing fixed across candidate comparisons.

​

Delayed responses

An outcome may continue changing after a command ends. Include the relevant measured state and define when the result is assessed. Use a fixed assessment interval when completion requires a condition to remain satisfied. Keep command holds, neutral padding, and assessment timing consistent across comparisons. Removing a command follows the declared time convention; see structural stages.

​

Short execution windows

Use short blocks when the environment needs another observation before the full task is over. Each block remains an ordered numeric sequence. For example, a proposal that executes six commands before termination produces six action rows and seven state rows. Preserve the full proposal separately from the executed prefix. Your application owns episode identity, block index, command duration, and reset semantics. Use a distinct request ID per new proposal and keep the same compatible learner Domain across windows. Define how the measured outcome is attributed to each executed block in your integration. Cross-block credit requires an explicit supported contract. The acquisition guide gives the observation body. For a named discrete action after each observation, use Sequential Discovery.

​

Compose routines through continuous state

When the application supplies a sequence of objectives, each acquired routine must accept the state produced by its predecessor. Keep each routine’s Domain, selected base, context encoding, and executor version together. Adapt-1 acquires each routine’s controls under the application’s objective order and handoff rules.

​

Correct an acquired behavior

Preserve a useful base, then learn bounded corrections from measured pre-action context. Include variables that distinguish conditions requiring different controls. Use the contextual refinement workflow to compare candidates on the same conditions. Optional context normalization, a quadratic basis, and response guidance can help when corrections vary with context. Keep decoder scaling fixed and inspect the executed commands after clipping. Include representative conditions in training comparisons, with separate confirmation contexts. Compare the retained correction with its unchanged base on reserved cases. To check persistence separately, reconnect to the same Domain and repeat evaluation without feedback. The retained-use guide covers the API calls and the manifest needed for reuse.

​

Choose feedback that describes the outcome

Use binary completion when that is the dependable measurement. Use measured stage progress or graded error when the application can supply it consistently. Keep score direction, scale, and assessment time fixed within a stage. Check completion separately from the training score. Supply step_outcomes only when outcomes are measured at each executed command; otherwise return the measured final outcome. Continue with the configuration guide to define the numeric schema and create a trajectory Domain.