Discover transition inputs

Discover transition inputs, inspect admission, and query with the target withheld.

Use this when the target is known and Adapt-1 should discover eligible transition inputs. For a complete runnable example, start with First learned result.

Transition Discovery forms the executable input projection from eligible events while the target remains explicit. It can also discover compatible before/after causal bindings in a mixed setup where the transition inputs themselves stay authored.

Transition projection applies to a declared relationship:

1. observable fields available before the result
2. observable numeric, categorical, boolean, or set target

The application declares the target. Core observes eligible events, identifies stable scalar paths, rebuilds buffered evidence when the projection changes, and exposes the discovered contract in learner state.

Configuration

{
  "learning": {
    "enabled": true,
    "transition": {
      "enabled": true,
      "event_types": ["observation"],
      "input_paths": [],
      "targets": [
        {
          "path": "values.process.output",
          "type": "number"
        }
      ],
      "required_support": 3,
      "autonomous_projection": {
        "enabled": true,
        "minimum_observations": 3,
        "minimum_availability": 0.8,
        "maximum_input_paths": 8,
        "discover_causal_bindings": true,
        "minimum_causal_variables": 2
      }
    }
  }
}

Projection fields

For normal use, configure the fields shown in the production examples:

Field	Meaning
enabled	Enables autonomous transition projection
minimum_observations	Controls how much complete evidence is required before discovery can become ready
minimum_availability	Prevents sparse fields from entering the discovered projection too easily
maximum_input_paths	Bounds the discovered input surface
discover_causal_bindings	Enables automatic before/after binding discovery when causal learning is used
minimum_causal_variables	Requires enough shared numeric variables for a causal binding

learning.transition.targets remains required. If input_paths is empty, autonomous projection must be enabled. A non-empty input_paths list remains authoritative.

Keep post-outcome measurements and target-derived values out of the pre-outcome learner view. Discovery can only work with the evidence boundary you provide.

Ingest events

curl -sS -X POST \
  "https://rei-neuroadapt-api.reilabs.org/api/v1/domains/process-output-demo/events" \
  -H "Authorization: Bearer $ADAPT1_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "session_id": "ignored",
    "event_type": "observation",
    "values": {
      "process": {
        "temperature": 24.5,
        "pressure": 1.2,
        "output": 52.6
      }
    },
    "metadata": {
      "episode_id": "setup-001",
      "step": 1
    }
  }'

Storage and learner admission are separate. Check learner_eligibility on every write.

Response	Meaning
accepted: false, rejection_reason: projection_accumulating	Event is buffered while no stable projection exists
schema_changed: true	The discovered projection changed and buffered evidence was rebuilt
input_source: discovered	Current paths came from autonomous projection
input_source: declared	Current paths were explicitly configured
sample_count_delta > 1	A projection revision admitted buffered events together

Inspect state

Request learning_state on /query or /explain:

{
  "return_fields": [
    "learning_state",
    "transition_prediction"
  ]
}

The projection state is available at:

learning_state
  .subsystems
  .structured_transition
  .autonomous_projection

Useful fields:

Field	Interpretation
status	disabled, accumulating, or ready
revision	Number of discovered schema revisions
buffered_events	Raw events available for rediscovery and rebuild
input_paths	Current executable transition inputs
input_source	declared or discovered
causal_binding_source	declared or discovered
causal_intervention_path	Current intervention field for causal observations
causal_variables	Variable names bound across before and after records

Query rules

Pass the current values at the same paths Core discovered. Omit the withheld target.

curl -sS -X POST \
  "https://rei-neuroadapt-api.reilabs.org/api/v1/domains/process-output-demo/query" \
  -H "Authorization: Bearer $ADAPT1_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "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
  }'

Possible abstention reasons include:

Reason	Correct response
projection_accumulating	Send more complete eligible events, then inspect projection state
Missing input paths	Supply every discovered path in query context
insufficient_support	Add compatible observations or lower support only with a measured reason
no_matching_group	Check group_by_paths and current group values
Distance or model abstention	Inspect evidence geometry and avoid forcing a prediction

Categorical targets

Use type: categorical for labels with no numeric ordering:

{
  "targets": [
    {
      "path": "values.case.disposition",
      "type": "categorical"
    }
  ],
  "discrete_model": "auto",
  "autonomous_projection": {
    "enabled": true
  }
}

Do not encode categories as arbitrary numbers. Preserve their string or boolean semantics.

Mixed causal discovery

For causal events, explicit pre-outcome inputs are often safer while before/after binding can remain automatic.

{
  "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"
      }
    ],
    "autonomous_projection": {
      "enabled": true,
      "discover_causal_bindings": true,
      "minimum_causal_variables": 2
    },
    "causal_graph": {
      "enabled": true,
      "event_types": ["intervention_result"]
    }
  }
}

Each event should contain repeated numeric variable names under 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.

Causal binding discovery identifies a compatible structured before/after contract. It does not prove causal identifiability, remove confounding, or replace an intervention protocol.

Move from discovery to a pinned contract

1. Run discovery on representative training data.
2. Read the discovered input_paths and causal bindings.
3. Check availability, semantics, leakage, stability across folds, and behavior under missing fields.
4. Create a new Domain or update learning configuration with approved explicit paths.
5. Evaluate frozen on held-out events.

Explicit paths prevent later projection changes. Keep automatic discovery enabled only for the portions that should continue adapting.

Discovery overview

Review the three 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.