Discovery examples

These examples cover numeric and categorical Transition Discovery, Structure Discovery, unknown action effects, Sequential Discovery, and causal-binding discovery. Numeric learning settings are example parameters.

Run one complete Discovery tutorial

Create a fresh Domain, submit the included observations, inspect before/after responses, and query retained state from one Python file.

Choose a declaration below, then use the shared create → ingest → inspect → query loop. These payloads extend the complete tutorial; HTTP and contracts describes schema coverage. All examples use:

https://rei-neuroadapt-api.reilabs.org/api/v1

Export a bearer token before running them:

export ADAPT1_API_KEY="YOUR_SECRET_TOKEN"

The hosted API derives tenant ownership and effective session identity from the bearer token. Request examples use session_id: ignored for compatibility. For a complete declaration with both predictor inputs and causal bindings initially unresolved, see Autonomous causal discovery. The causal-bindings tab below shows the mixed variant with reviewed predictor inputs.

Choose a declaration

Numeric
Categorical
Structure
Reviewed fields
Causal bindings

This Domain declares the numeric target and lets Adapt-1 discover the stable input projection.

{
  "domain_id": "process-output-demo",
  "session_id": "ignored",
  "description": "Predict an observed process output from available measurements.",
  "schema": {
    "event_types": ["observation"]
  },
  "learning": {
    "enabled": true,
    "transition": {
      "enabled": true,
      "event_types": ["observation"],
      "targets": [
        {
          "path": "values.process.output",
          "type": "number"
        }
      ],
      "required_support": 3,
      "numeric_model": "auto",
      "autonomous_projection": {
        "enabled": true,
        "minimum_observations": 3,
        "minimum_availability": 0.8,
        "maximum_input_paths": 8
      }
    }
  }
}

Send complete observations in order. The first events can return projection_accumulating. Continue while storage succeeds and inspect the projection until it becomes ready.At query time, omit the target:

{
  "session_id": "ignored",
  "question": "Predict the process output.",
  "context": {
    "values": {
      "process": {
        "temperature": 30,
        "pressure": 1.2
      }
    }
  },
  "return_fields": ["transition_prediction", "learning_state"],
  "update_memory_state": false,
  "allow_exploration": false
}

Verify:

learning_state.subsystems.structured_transition.autonomous_projection.status == "ready"
learning_state.subsystems.structured_transition.autonomous_projection.input_source == "discovered"
transition_prediction.status == "predicted"
transition_prediction.support_count >= required_support

Use a categorical target when labels have no numeric geometry.

{
  "domain_id": "case-disposition-demo",
  "session_id": "ignored",
  "description": "Predict case disposition from stable observed fields.",
  "schema": {
    "event_types": ["resolved_case"]
  },
  "learning": {
    "enabled": true,
    "transition": {
      "enabled": true,
      "event_types": ["resolved_case"],
      "targets": [
        {
          "path": "values.case.disposition",
          "type": "categorical"
        }
      ],
      "required_support": 3,
      "discrete_model": "auto",
      "autonomous_projection": {
        "enabled": true,
        "minimum_observations": 4,
        "minimum_availability": 0.9,
        "maximum_input_paths": 12
      }
    }
  }
}

Eligible event:

{
  "session_id": "ignored",
  "event_type": "resolved_case",
  "values": {
    "case": {
      "severity": "high",
      "repeat_count": 3,
      "verified": true,
      "disposition": "review"
    }
  },
  "metadata": {
    "episode_id": "cases-001"
  }
}

Query context omits disposition.

This Domain lets Adapt-1 inspect eligible scalar fields, evaluate useful combinations and lags, and induce typed target rules.

{
  "domain_id": "case-structure-demo",
  "session_id": "ignored",
  "description": "Discover observable structure associated with case outcomes.",
  "schema": {
    "event_types": ["resolved_case"]
  },
  "learning": {
    "enabled": true,
    "structure": {
      "enabled": true,
      "event_types": ["resolved_case"],
      "feature_paths": [],
      "targets": [
        {
          "path": "values.outcome",
          "type": "categorical"
        }
      ],
      "representation": {
        "enabled": true,
        "maximum_order": 2,
        "maximum_lag": 3,
        "minimum_support": 12,
        "episode_path": "metadata.episode_id"
      }
    }
  }
}

Inspect:

learning_state.subsystems.autonomous_structure.status
learning_state.subsystems.autonomous_structure.representation.selected_paths
learning_state.subsystems.autonomous_structure.representation.active_representations
induced_structure.hypotheses[].conditions
induced_structure.hypotheses[].supporting_memory_ids
induced_structure.hypotheses[].counterevidence_memory_ids

A reviewed base surface can remain explicit while representation discovery continues over combinations and lags derived from it.

{
  "domain_id": "reviewed-structure-demo",
  "schema": {
    "event_types": ["resolved_case"]
  },
  "learning": {
    "enabled": true,
    "structure": {
      "enabled": true,
      "event_types": ["resolved_case"],
      "feature_paths": [
        "values.signal_a",
        "values.signal_b"
      ],
      "targets": [
        {
          "path": "values.outcome",
          "type": "categorical"
        }
      ],
      "representation": {
        "enabled": true,
        "maximum_order": 2,
        "maximum_lag": 2,
        "minimum_support": 12,
        "episode_path": "metadata.episode_id"
      }
    }
  }
}

The explicit base paths stay authoritative. Active rules can still update from evidence.

This pattern keeps the pre-outcome transition inputs explicit while allowing Adapt-1 to discover compatible before/after causal bindings.

{
  "domain_id": "intervention-effect-demo",
  "session_id": "ignored",
  "description": "Learn observable intervention effects from before and after measurements.",
  "schema": {
    "event_types": ["intervention_result"]
  },
  "learning": {
    "enabled": true,
    "transition": {
      "enabled": true,
      "event_types": ["intervention_result"],
      "input_paths": [
        "values.before.temperature",
        "values.before.pressure",
        "values.intervention"
      ],
      "action_path": "values.intervention",
      "targets": [
        {
          "path": "values.after.output",
          "type": "number"
        }
      ],
      "required_support": 3,
      "autonomous_projection": {
        "enabled": true,
        "minimum_observations": 4,
        "discover_causal_bindings": true,
        "minimum_causal_variables": 2
      },
      "causal_graph": {
        "enabled": true,
        "event_types": ["intervention_result"]
      }
    }
  }
}

Each event contains repeated numeric variable names under before and after, plus an intervention string matching one of those variable names. Inspect causal_binding_source, causal_intervention_path, causal_variables, and causal_graph before using the result.

Learn unknown action effects

A Discovery Domain can expose public measurements and legal opaque actions while leaving their effects unresolved. Adapt-1 learns from the observed changes produced by executing those actions.

Supplied by the application	Formed from admitted evidence
Public measurements, target paths, and event meanings	Executable predictor inputs and supported transition predictions
Stable legal action identities and the executed action	Action-conditioned evidence about observed effects
Compatible intervention records	Discovered intervention and variable bindings, then learned causal relations
Observable outcomes and real episode boundaries	Retained evidence for subsequent predictions and applicable policy learning

This setup supplies no action-effect mapping, hidden topology, reversibility rule, or one-to-one assignment between actions and effects. The target identifies what to learn from the observations. Leave transition input_paths empty and enable autonomous_projection. When the records support causal learning, also enable discover_causal_bindings and causal_graph, with the intervention binding and causal-variable declaration unresolved. The autonomous causal example shows this declaration and its event contract. Use the following loop:

Capture the public state before the action.
Obtain a legal action from the configured selector or experiment schedule, then execute that exact action.
Observe the resulting public state and outcome.
Submit the pre-action measurements, executed action, and observed consequence with their true episode and step.
Inspect learner admission, discovered inputs, bindings, and model versions.
Query from the next public state and inspect predictions or policy attribution from the retained evidence.

action_path identifies the field containing the executed action. The causal intervention binding identifies evidence of an intervention on a measured variable. An opaque action ID does not automatically name that variable. Inspect the discovered causal bindings separately from action-conditioned prediction.

Learn choices from measured outcomes

Use Transition Discovery to predict an immediate reward from current measurements and a permitted choice. For example, a monitoring service can issue one of two condition classifications from numeric readings, then receive a correctness score for the classification it issued. The scoring system keeps the reference category and returns the measured score for that choice. Declare the measured reward as a numeric transition target, leave input_paths empty, and enable autonomous_projection. Include the chosen action in completed events. Measurement fields can use stable names such as x_00 and x_01; preserve each field’s meaning, type, and units across records.

Capture the measurements available before the outcome.
Query the same measurements with each permitted candidate action and the reward withheld.
Have the application compare supported reward predictions and choose an action under its declared selection or exploration rule.
Execute or issue that choice and observe its reward.
Submit the original measurements, the choice actually used, and its observed reward as a completed event to the same Domain.
Continue in chronological order so later predictions can use the new evidence.

transition_prediction supplies the estimates; the application performs comparison and selection. Record abstentions and any fallback choice. Return the reward actually observed for the executed choice. Keep each prediction recorded before its outcome is submitted. Continued event ingestion lets the reward predictor update as new outcomes arrive. Inspect admission and learner versions even when the retained sample count or discovered input paths stay fixed; see Readiness and continued learning. When actions change later states and later outcomes should revise earlier choices, add Sequential Discovery with its ordered feedback contract.

Sequential Discovery

Use Sequential Discovery when actions change later states and later rewards should revise earlier choices. An opaque-action loop can expose public observation fields and legal actions while leaving their useful action values and action-effect structure for Adapt-1 to form:

query current public state
execute the selected legal action
observe the public next state and native reward
return episode ID, step, next state, reward, and terminal status
query again from retained Discovery state

Configure learning.sequential.enabled: true with real episode, step, next-state, reward, and terminal paths. Preserve the selected policy and decision_id through execution. Inspect sequential sample counts and selection attribution before claiming that retained state influenced a later action. See Sequential Discovery for the starting declaration and Sequential learning, advanced for the full policy, reward, training, and evaluation contract.

Executable runner pattern

The following is a zero-start online pattern shared by the Discovery paths. For separate acquisition, run the evidence-formation steps before the declared new-run boundary, then freeze the resulting state or continue adapting according to the protocol:

create a unique Domain
send eligible events or feedback
inspect path-specific admission on every write
wait for usable discovered state
query, select, or predict
return the observable consequence when applicable
verify that a later result uses the intended retained state

Use a fresh unique domain_id for an independent rerun. When a reported run starts empty, state retained across its episodes belongs to that run. For held-out frozen evaluation, split entities or episodes before ingestion, form state only on the acquisition partition, record the final learner state and version, stop event and feedback writes, and verify state identity remains unchanged through scoring.

Common setup failures

Symptom	Cause	Correction
Domain creation returns `422`	Unknown field, missing target, or transition enabled without inputs or projection	Validate the declaration against the current Domain schema
Stored events leave learner count at zero	Projection is still accumulating or the event is missing a target	Inspect `learner_eligibility.rejection_reason` and `missing_paths`
Projection stays empty	Too few complete events or candidate fields do not meet availability	Send representative complete events and review observation and availability settings
Query abstains after projection is ready	Query context is missing a discovered path or lacks compatible support	Read `input_paths`, then send the same pre-target structure in `context`
Structural learner has samples and no active rule	The current evidence is still insufficient for a supported rule	Inspect representation and induced-structure diagnostics; do not force selection
Stored sequential feedback produces no sequential sample	Episode, step, next state, reward, terminal status, or policy binding is missing or invalid	Inspect feedback admission and preserve the full sequential tuple
Temporal candidates cross unrelated records	Episode identity is absent or reused across independent sequences	Set `metadata.episode_id` consistently and start a new episode at each reset
A suspicious field dominates	Identifier, post-outcome value, or target-derived field entered the learner view	Remove it, create a clean Domain, and rerun Discovery and holdout evaluation

Discovery

Understand the Discovery boundary and lifecycle.

Transition Discovery

Review projection fields, revisions, and causal bindings.

Structure Discovery

Review discovered fields, combinations, temporal relationships, and induced rules.

Sequential Discovery

Review ordered feedback, delayed credit, readiness, and action attribution.

Get started

Concepts

Guides

Discovery

Operations

API reference

Run one complete Discovery tutorial

Choose a declaration

Learn unknown action effects

Learn choices from measured outcomes

Sequential Discovery

Executable runner pattern

Common setup failures

Discovery

Transition Discovery

Structure Discovery

Sequential Discovery

Run one complete Discovery tutorial

​Choose a declaration

​Learn unknown action effects

​Learn choices from measured outcomes

​Sequential Discovery

​Executable runner pattern

​Common setup failures

Discovery

Transition Discovery

Structure Discovery

Sequential Discovery

Choose a declaration

Learn unknown action effects

Learn choices from measured outcomes

Sequential Discovery

Executable runner pattern

Common setup failures