Skip to main content
Configure native Adapt-1 planning through the request’s options object and its request-level evaluator fields. Match the search limits to the application’s transition unit, then inspect the completed work and evaluator identity in the response. The integration guide describes HTTP graphs and streaming expansion.
Types and ranges below describe schema version 5. Defaults for nullable options and behavior notes include observed responses. Check /version at the API host root and confirm compatibility with your deployment. The HTTP reference describes the request fields.

Select the evaluator

evaluator_source belongs at the request level, alongside relation, policy, root, and options. Use expected_model_sha256 to bind the expected model identity where the planning contract supports it, and check the returned evaluator identity. Use evaluator_source: domain_model for the installed Domain evaluator, including a compatible sequential_q_mlp. Responses identify evaluator.model_type, evaluator.model_sha256, and evaluator.model_version. Align its value objective when comparing its choices with another evaluator. This illustrative response fragment shows the reported model type and identity field. The hash is a placeholder; preserve the actual returned hash and model version.
Evaluator identity fragment
Completed successors can supply exact terminal_value scores, which take precedence over learned estimates. Keep a record of whether terminal scores or learned frontier values determined a choice. See value learning for fitting the value evaluator.

Check evaluator readiness

Confirm that the chosen evaluator exists before the first planning request. Requests without an installed model have been refused even when every reachable leaf within the requested depth supplied an exact terminal_value. The domain_model refusal names a numeric checkpoint as an alternative. The documented request schema types checkpoint only as an object or null; its supported contents and source compatibility are not yet documented. Follow the installed-model workflow until you have a confirmed checkpoint contract for your deployment. For sequential acquisition, confirm eligible feedback admission, training.min_samples, sequential.minimum_episodes, retained episode coverage, completed asynchronous fitting, and a passing validation result. Installation depends on these combined conditions. If training reports candidate_failed_validation, inspect reward variation and the validation report; constant-reward acquisition has returned that result. See sequential installation checks. For the separate value learner, complete initial fitting and installation, inspect value/state, and then verify the evaluator identity returned by planning.

Align the value objective

Use one objective and perspective, with compatible neutral values, discounts, and horizon conventions for evaluator estimates and terminal scores. Match the objective as well as the numeric range. Sequential values are learned from the configured rewards and transitions. An outcome-trained value learner estimates the supplied outcome target. If sequential rewards include intermediate benefits or costs that terminal scores omit, search can change the chosen action because it is comparing different objectives. Exact terminal values take precedence over learned frontier estimates. When the task objective is a terminal outcome, the value learner’s outcome targets provide an explicit way to set its value, neutral, discount, and terminal step. Align those settings with planning and validate decisions at the intended search depth. Choose the evaluator whose learned objective matches the task. See outcome targets and rewards and costs in the graph.

Observed terminal discounting

Observed terminal values have matched n + (t - n) * gamma^d, where n is neutral_value, t is terminal_value, gamma is temporal_discount, and d is the number of transitions to the terminal state. The one-step calculation is 0.5 + (0.6 - 0.5) * 0.8 = 0.58. These calculations illustrate terminal discounting toward neutral. For learned frontier values, confirm the backup rule and its interaction with already-discounted training targets in your deployment’s contract.

Bound search work

Treat these as separate limits. Read completed_depth for achieved ordinary depth and budget_exhausted and deadline_reached for stopping conditions. Record every reported stopping condition when multiple limits are reached. The following fragment makes the schema’s depth and work defaults explicit. It is an options object, not a complete planning request.
Planning options fragment

Extensions, proof bounds, and rechecks

Use completed_depth for completed ordinary search depth and extension counters for extension work. Report these quantities separately. Start with the documented defaults and confirm each optional mode’s trigger and accounting rules for your deployment before changing it. Extensions can reach terminal states beyond the ordinary depth limit. Read extensions, completed_extensions, and evaluations separately; these counters describe different aspects of the completed work. Record explored paths and work counters to show the depth reached and work performed by each extension mode. Compare modes across representative graphs and decision roles, and confirm their trigger and increment rules against the deployment contract.

Reuse, value scaling, and exploration

Configure temporal_discount in planning and construct training targets through value observations. To blend a continuation assessment with an observed outcome, calculate the target explicitly and preserve its inputs. See value observations for target construction. For a controlled comparison, hold planning options and the transition interface fixed while changing the evaluator. Use exploration_probability: 0 when evaluating the planner’s selected action without exploration. Track model identities independently of cache counters.

Inspect the result

Record the requested options, returned action, legality check, evaluator identity, completed depth, extensions, evaluation and transition counts, cache counters, and stopping diagnostics. evaluator.learning_updates reports evaluator updates within the call; it is separate from search work. Observed calls returned evaluator.learning_updates: false. Planning evaluates the supplied continuations with the selected evaluator. After executing the chosen action, admit its measured outcome through the appropriate learning write. A fit started earlier can finish while planning requests are running. See frozen-use rules. Native planning routes use /api/v1/domains/{domain_id}/planning/adversarial and the corresponding /stream WebSocket. The route name supports tasks with only maximizing decisions as well as opposing decisions. HTTP and contracts explains authentication and request addressing.