- observable fields available before the result
- observable numeric, categorical, boolean, or set target
Configuration
The numeric settings below are example parameters.Projection fields
For normal use, configure the fields shown in the production examples:learning.transition.targets remains required. If input_paths is empty, autonomous projection must be enabled. A non-empty input_paths list remains authoritative.
Ingest events
learner_eligibility on every write.
Inspect state
Requestlearning_state on /query or /explain:
Readiness and continued learning
Inspect these stages separately:
Record supported-query coverage and outcome quality separately. Evaluate online operation by recording each prediction before submitting its outcome. For frozen evaluation, stop learner writes and verify state identity through scoring. Learning schedules describes both protocols.
With bounded retention, an accepted update can leave the retained sample count unchanged. In the recorded run below, the buffer reaches 512 events; later updates report
accepted: true and increasing model_version while sample_count_delta remains zero. Inspect admission and learner versions together. Record total observations and retained samples separately. Projection revision tracks input-schema changes, so it can also remain fixed while the predictive learner continues updating.
Read a recorded projection change
These excerpts come from a separate recorded learner run. The application log records prediction diagnostics before the observation update at each step.step, diagnostics, and update below are fields in that log. Selected values are preserved from the recording.
sample_count_delta: 2 records the admission of both buffered observations, and schema_changed: true marks the projection revision.
The following query produces predictions for both action values. This run estimated reward from observations and a candidate action; predicted_reward preserves those two estimates in the application log. The number of observations needed for readiness follows the run’s configured projection thresholds.
Query rules
Pass the current values at the same paths Core discovered. Omit the withheld target.Categorical targets
Usetype: categorical for labels with no numeric ordering:
Causal discovery
Causal discovery uses observations and interventions to infer causal relationships. Define the causal result being sought and the evidence that can distinguish competing explanations.Causal questions and evidence
CausaLab and IVRE both concern causal inference. Their unknown relationships and experiment protocols differ.
The CausaLab schedule supplies intervention measurements. Adapt-1 learns the causal relationships from their observed effects. The result demonstrates causal recovery in the evaluated numerical setting.
IVRE involves ambiguous evidence about discrete causes. A detector activating with objects A and B together can be consistent with several assignments of causal roles. Further experiments help distinguish those assignments. A correct activation prediction for that combination alone leaves the individual causes unresolved.
IVRE therefore tests causal identification and the choice of informative experiments. Assess whether the learner can identify causes from a fixed set of legal observations, then assess whether its experiment choices obtain useful evidence under the same budget. Evaluate causal beliefs separately from detector prediction accuracy and experiment-selection quality.
Autonomous causal discovery
Use this setup when both the predictor inputs and the before/after causal bindings should form from observations. The application supplies the public measurements, permitted interventions, event meanings, and observable targets. Adapt-1 forms the executable input projection and compatible variable and intervention bindings, then learns relationships from admitted intervention evidence. The following declaration leavesinput_paths, causal_graph.intervention_path, and causal_graph.variables unresolved. The numeric thresholds are example settings. They control admission and readiness; evaluate the learned relationships separately.
Create the Domain through POST /api/v1/domains:
after field as a target. Keep additional post-intervention measurements outside the candidate input surface. An earlier measurement such as values.before.output is available before the intervention and has a different role from the withheld values.after.output.
Submit the observed result of each permitted intervention through POST /api/v1/domains/autonomous-intervention-demo/events. This payload illustrates the record shape; use the measurements from the actual execution:
before and after provide candidate variable pairs. The intervention string identifies the variable deliberately manipulated. setpoint is the requested value, known before execution. Keep these meanings stable across records and preserve the actual episode and measurement boundaries.
Collect varied legal interventions that can distinguish the relationships being investigated. After each write, inspect learner_eligibility and the returned projection state. When requesting learning_state, inspect:
causal_graph relationships and their evidence separately when returned. A provenance label of discovered with empty binding fields does not establish that a causal binding has formed.
Query a proposed intervention through POST /api/v1/domains/autonomous-intervention-demo/query, supplying the current measurements and proposed command while withholding all after targets:
transition_prediction when its status is predicted, and retain the support, evidence IDs, model version, and abstention reason when returned. The query context must satisfy the discovered input contract. Evaluate predicted measurements and recovered causal relationships against their respective observations and intervention evidence.
Automatic discovery can remain enabled during continued learning. Explicitly pin inputs or bindings when the application requires a reviewed fixed contract. Learn unknown action effects applies the same boundary to opaque legal actions.
Numeric before/after contract
For a mixed configuration, declare the approved pre-outcome inputs and enable binding discovery alongside causal-graph learning. The following excerpt keeps predictor inputs explicit while compatible numeric before/after bindings form from evidence.before and after, plus an intervention string matching one of those variable names. Core can discover the paired paths and intervention field while the declared transition inputs remain fixed.
Binding discovery identifies which event fields describe each variable before and after an intervention. Causal-graph learning uses intervention evidence to estimate causal relations among those variables. Inspect the recovered relations separately from binding readiness and prediction accuracy. Causal interpretation requires a valid intervention protocol and an assessment of confounding and identifiability.
Discrete cause identification
For object-panel evidence, preserve each object’s presence and the observed detector outcome. Use appropriate categorical or boolean targets for outcome prediction. Define how supported model outputs, induced rules, or hypotheses produce the causal beliefs being evaluated. A detector predictor’s output and an object-role belief have different meanings. The automatic binding interface above requires compatible numeric before/after variables. Select a learning contract that matches the discrete experiment records. Enabling that interface alone does not establish a working IVRE causal-inference loop. When the learner chooses experiments, give it the observed experiment history and an objective tied to identification quality or efficiency. Detector activation probability alone does not measure the value of an experiment for identifying causes. Record admitted evidence, supported-query coverage, causal-belief quality, and complete task outcomes. This separates unavailable predictions, incorrect causal inferences, and uninformative experiment choices. Check all three before interpreting an unsuccessful causal-discovery run.Move from discovery to a pinned contract
- Run discovery on representative training data.
- Read the discovered
input_pathsand causal bindings. - Check availability, semantics, leakage, stability across folds, and behavior under missing fields.
- Create a new Domain or update learning configuration with approved explicit paths.
- Evaluate frozen on held-out events.
Discovery overview
Review the Discovery paths and their shared lifecycle.
Structure Discovery
Discover useful fields, combinations, lags, and predictive rules.
Sequential Discovery
Discover state-dependent action values and delayed credit across episodes.
Complete examples
Run transition, structural, sequential, and mixed workflows.