Glossary

Every term the documentation relies on, each with its one-line meaning and the chapter that owns it. Terms the API spells differently from ordinary prose are noted, so a YAML file and a sentence never disagree.

Regime

The configuration the world is currently in — the set of conditions the agent acts under. Prose sometimes says configuration; the API always says regime (regime_id, RegimeProvider, the regime: section). They are the same thing. → Regimes and schedules

Configuration interval

A maximal stretch of steps recorded under one regime label — the unit the metrics segment a run by. Written \(T_k\). → Anatomy of a run

Shift (boundary)

The step at which the recorded label changes. Adaptation and recovery are measured from one; a shift never ends an episode. → Regimes and schedules

Axis

One named condition a regime sets — an arrival rate, a capacity multiplier, a distribution over failure kinds. Any schedule key that is not regime_id, duration_steps or alias_of is an axis. → Regimes and schedules

alias_of / origin

A schedule entry that repeats another regime's axes under a new label, giving forgetting its second encounter. Every trace row records the origin — what its label repeats — and forgetting pairs encounters through it. → Regimes and schedules

Provider

The object that answers regime_id(step) for a run. Built in: constant (a stationary control) and scheduled. → Regimes and schedules

Window (rollout window)

The harness's administrative slice: window_steps of experience per algorithm update. A window boundary ends nothing. → Anatomy of a run

Episode

The environment's own unit, ended only by the environment (terminated/truncated). A persistent world may never end one. → Anatomy of a run

World instance

One provisioned environment living for a whole run. In a persistent world, reset destroys and rebuilds it, which is why the harness never resets on its own initiative. → Anatomy of a run

Tier 1 / Tier 2

The two levels of the environment contract. Tier 1: any gymnasium.Env, supplying its own reward. Tier 2 adds three probed capabilities — a per-step record in info["trace"], a regime report in info["regime"], and describe() — and gains the composable reward system. → Rewards as data

Step record / StepContext / EpisodeStep

The description of one step. A StepContext is the read-only slice a reward component scores; an EpisodeStep is the serialised row written to the trace, carrying the reward, its per-component breakdown, the label and the origin. → Rewards as data

Trace

The run's record: one EpisodeStep row per step, JSONL on disk. The evaluation layer reads traces and nothing else. → Anatomy of a run

Component

One term of a composed reward: a small class with a name, a cadence, a default clip and weight, and reads — the StepContext fields it consumes. → Rewards as data

Cadence

When a component fires: STEP every step, EPISODE_END once, on the episode's final step — which is why a window shorter than an episode legitimately sees only per-step terms. → Rewards as data

Specification (reward spec)

The reward as data: named components with weights, clips and gates, serialisable to YAML, selected by configuration. → Rewards as data

Gate

A specification rule that zeroes one component's contribution unless another's is positive — for terms that should pay only alongside real progress. → Rewards as data

Anchor

What adaptation and recovery are measured against: segment (the interval's own peak and settled tail — the default) or ceiling (the theoretical bound). Numbers computed under different anchors are not comparable. → The metrics

Ceiling / upper bound (\(R_{\mathrm{ub}}\))

The most a reward specification could pay, derived from its components' clips and weights — per ordinary step and per episode-final step. Anchors the performance gap. → The metrics

Settled reward / asymptote

The mean over the final fifth of an interval — where the policy ended up, as opposed to how it got there. Recovery measures settling towards it. → The metrics

Forgetting

First encounter's mean minus the second's, paired through the origin. Positive means worse; negative is backward transfer. → The metrics

Plasticity / effective rank

A diagnostic for the capacity to keep learning: the effective rank of the policy's representation, recorded per step when capture is enabled. Falling rank with flat reward is a policy losing plasticity. → The metrics

Family (algorithm_id)

The arm a run belongs to in a comparison — usually an algorithm, sometimes one algorithm under two configurations. Cross-family tables group on it.

Sweep / cell

A matrix of runs written as one document — what every cell shares, the families, the seeds — expanded into ordinary experiment configurations. One cell is one run.

Backend

A training algorithm registered by name, selectable in configuration (discrete_hill_climbing, ppo, ppo_ewc, ppo_lcm).

Bundle

An exported policy as one directory: a manifest recording the environment, spaces, reward specification and provenance, beside the checkpoint — enough to serve the policy without the training run.

MORPHEUS

The persistent enterprise simulation platform this package was built alongside — live business worlds that inject failures and shift configuration while an agent remediates incidents. One adapter speaks to it; nothing else in the package knows it exists. → Background