> ## Documentation Index
> Fetch the complete documentation index at: https://docs.reilabs.org/llms.txt
> Use this file to discover all available pages before exploring further.

# Configuration reference

> Look up forecasting and planning defaults, bounds, and settings for an existing Domain.

## Futures

All fields below belong under `learning.transition.futures`.

| Field | Default / valid values | Effect |
| - | - | - |
| `enabled` | `false` | Turns on the additional futures forecast. |
| `conditioning` | `"adaptive"`; also `"prior"`, `"kernel"` | Adaptive weighs conditioning using resolved forecasts. Prior and kernel are explicit controls for comparison. |
| `numeric_transport` | `"none"`; or `"linear_residual"` | Enables validated numeric-level transport. |
| `max_support` | 256; 1-1024 | Maximum support rows contributing probability atoms. Numeric-center fitting can use the full retained matching set. |
| `max_branches` | 16; 1-128 | Displayed atoms and live planning branches; smaller budgets can leave more mass unresolved. |
| `bandwidth` | 0.25; positive | Evidence locality. Smaller values give more local weighting. |
| `prior_strength` | 2; positive | Regularizes adaptive conditioning. Larger values make its adjustment more conservative. |
| `max_scores` | 256; 1-4096 | Recent scored outcomes retained per calibration group. |
| `max_calibration_groups` | 128; 1-1024 | Maximum histories for distinct evidence masks/groups/actions. |
| `max_pending_forecasts` | 128; 1-4096 | Maximum unresolved issued forecast receipts. Revisions occupy slots until resolved. |
| `future_id_path` | `"metadata.future_id"` | Path identifying revisions of the same eventual outcome. |

Related transition settings:

| Field under `learning.transition` | Default | Effect |
| - | -: | - |
| `required_support` | 2 | Minimum matching evidence before prediction. |
| `max_samples` | 512 | Retained transition observations; separate from `max_support`. |
| `max_distance` | 0.75 | Maximum accepted normalized evidence distance. A permissive value is not a substitute for representative observations. |
| `group_by_paths` | Empty | Exact data partitions before conditioning. |
| `numeric_min_skill` | 0 | Validation requirement for numeric fitting/transport. |
| `numeric_ridge` | 0.00000001 | Regularization for the numeric fit. |

## Probabilistic planning

All fields below belong under `learning.transition.planning`.

| Field | Default / valid values | Effect |
| - | - | - |
| `enabled` | `false` | Enables transition planning. |
| `probabilistic` | `false` | Uses distributions instead of the point-planning path. |
| `actions` | Empty list; up to 256 | Candidate action values; must correspond to observed action values. |
| `actions_path` | Unset | Alternative to fixed actions: a path containing the candidate list at the projected state. Do not set both forms. |
| `state_bindings` | Empty | Maps each predicted target to the next state's input path. |
| `goals` | Empty | Conditions defining desired states. |
| `objectives` | Empty | Numeric maximize/minimize/target preferences, with explicit scale and weight. |
| `preconditions` | Empty | Conditions checked before expanding a candidate action. |
| `constraints` | Empty | Conditions the projected outcomes should satisfy. |
| `horizon` | 3; 1-32 | Number of action steps searched. Start small, then confirm intermediate-state support. |
| `beam_width` | 32; 1-1024 | Candidate sequences retained during search. |
| `discount` | 0.95; 0-1 | Weight of later utility. Independent of the sequential reward learner's discount. |
| `risk_weight` | 0; 0-1 | Mixes expected utility with utility in the worst-outcome tail. |
| `tail_probability` | 0.1; greater than 0, up to 1 | Fraction of worst outcomes used for the tail utility. |
| `max_failure_probability` | 0; 0-1 | Maximum modeled constraint-failure upper bound; unresolved mass contributes when constraints exist. |
| `max_unresolved_probability` | 0.05; 0-1 | Maximum unsupported or truncated probability mass accepted in a plan. |
| `max_rollouts` | 4096; 1-100000 | Hard expansion budget. Exhaustion returns an abstention. |
| `policy_weight` | 0; nonnegative | Optional first-action preference from contextual/sequential policy utility, separate from forecast probabilities. |

Leave `uncertainty_penalty`, `exploration_bonus` and `minimum_confidence` at zero
in probabilistic mode. Those point-planner controls are incompatible here; use
the explicit probability/risk controls instead.

The `goals`, `preconditions` and `constraints` condition operators are `eq`, `ne`,
`lt`, `lte`, `gt`, `gte`, `between`, `contains`, `subset` and `superset`.
Conditions use a literal `value` or a `value_path`. Objectives use `direction`
(`maximize`, `minimize`, `target`), `weight`, `scale`, and a target where applicable.

Larger retention, support, branch and rollout budgets increase work and state
size. Increase the budget that is actually limiting the experiment rather than
setting every limit to its maximum.

## Update an existing Domain

Prefer a fresh Domain for the first integration test. When updating an existing
one, retain your complete intended learning declaration and submit it to
`POST /domains/{id}/rules` with the same session.

Do not assume that posting only `{"learning":{"transition":{"futures":...}}}`
will deep-merge all omitted settings. Read back the effective configuration and
verify retention, targets, inputs, action mapping and other enabled learners.
Do not blindly repost a response containing read-only diagnostics.

Changes that invalidate retained transition evidence can be rejected. Use a
separate Domain for a genuinely different target schema. Planning-only changes
do not require discarding transition experience.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.