Terminology

Purpose

This page defines the canonical vocabulary for this wiki. Agents MUST use these terms when writing durable wiki pages, source pages, entity pages, question pages, and repository-local documentation.

The wiki’s primary reader frame is time-series modeling and world models. Domain-specific terms from reinforcement learning, healthcare, recommender systems, education, robotics, causal inference, and observability MAY appear, but they SHOULD be mapped back to the canonical terminology here when ambiguity is possible.

Defaults

• Time series means ordered observations over time. Sampling MAY be regular or irregular.
• Time-series samples are assumed to be numerical by default. When the wiki intentionally uses time series more broadly, agents MUST explicitly name the modality: textual logs or events, images as in video, structured records, graph snapshots, audio, or another non-numeric form.
• Time-series model means a model over numeric time series by default.
• Agents SHOULD explicitly name the modality only when it is not numeric, such as event stream, categorical event sequence, graph time series, text-conditioned time series, or image/video trajectory.
• Agents MUST NOT use vector/time-series model. Use multivariate time-series model instead.
• Agents SHOULD NOT use tabular time series. Use tabular data when the object is a table without temporal sequence semantics, and time series when temporal order is central.
• Agents MUST NOT call vector a modality. A vector is a representation format; the underlying modality is usually numeric features, categorical IDs, graph structure, text, image/video, audio, or events.

Canonical Terms

Preferred Term	Meaning In This Wiki	Notes And Avoided Ambiguity
time series	Ordered observations over time, regular or irregular	Does not require uniform sampling. Samples are numerical by default; explicitly name textual logs/events, images/video, structured records, graph snapshots, audio, or other non-numeric samples when using the term broadly.
multivariate time series	A time series with multiple numeric variables or channels	Preferred replacement for ambiguous phrases such as vector/time-series model.
irregular time series	Measurements or events with non-uniform timestamps	Common in EHR, observability, and behavioral logs.
event stream	Ordered events, usually discrete or categorical, with optional timestamps	Use when the data is not primarily numeric measurements.
categorical event sequence	Event stream whose core values are categories, codes, IDs, or symbols	Common for education, recommender, and clinical-code data.
trajectory	Time-ordered states or observations, often with actions or control inputs	Common in world models, control, and RL.
state	Compact representation of the system at time t	Clarify whether it is observed, estimated, or latent.
observation	Data measured or made available to the model at time t	Do not assume observation equals true system state.
latent state	Learned or hidden state representation used by a model	Also called hidden state or belief state in some fields.
action	Controllable input chosen by an agent, policy, recommender, clinician, operator, or controller	Do not use for arbitrary passive events.
control input	Numeric or actuator-like action in a dynamical system	Preferred when discussing control systems or continuous actions.
intervention	Action or manipulation whose effect on future observations is of interest	Stronger causal connotation than action.
exogenous variable	External variable not controlled by the modeled agent/policy	Do not call an exogenous shock an action.
event	Something observed to happen at a time	May be passive or controllable; clarify before treating as action.
context	Variables that condition dynamics, actions, or outcomes	May be static or time-varying; clarify when needed.
modality	Semantic data type or representation family	Examples: numeric features, categorical events, graph structure, text, image/video.
world model	Learned predictive representation of environment or system dynamics used to reason about future states and action consequences	In the LeCun/AMI framing, the model should learn abstract state structure from observation, ignore irrelevant low-level details, support memory, reasoning, and planning, and usually operate in latent or representation space rather than by pixel/token reconstruction alone. MAY be passive or action-conditioned; specify which.
passive dynamics model	Predicts future observations from past observations without a controllable action channel	Often overlaps with forecasting.
action-conditioned world model	Predicts future observations or states conditional on actions, control inputs, or interventions	Preferred target phrase for datasets with actions; domain terms such as treatment MUST be mapped to these canonical terms when used.
forecasting	Predicting future observations from historical observations	Not necessarily causal and not necessarily action-conditioned.
counterfactual prediction	Predicting what would happen under an alternative action or intervention	Use only when causal assumptions or intervention semantics matter.
confounding	Dependence between actions/interventions and prior state/context that biases causal interpretation	MUST be noted for observational treatment logs.
logged decision data	Historical records of context, chosen action, and outcome	Preferred for recommender/bandit logs with weak next-state dynamics.
next-state dynamics	Relationship from current state/observation and action to future state/observation	Core world-model framing.

World Models

In this wiki, a world model is not merely any forecaster. It is a learned model of the relevant state and dynamics of an environment or system, used so an agent or analyst can evaluate plausible futures, consequences of actions, and constraints before acting. The LeCun/AMI-style interpretation emphasizes abstract representations learned from observation, prediction in representation space, persistent state or memory, and planning over action sequences rather than next-token prediction or frame-perfect reconstruction alone. APTAMI is the landmark source for this framing; LeWorldModel is the current JEPA-style implementation anchor in this corpus.

A world model MAY be passive when it predicts future observations from history alone, but the central agentic case is action-conditioned: it predicts how future observations or states change under actions, control inputs, interventions, or clearly mapped domain-specific analogs.

Domain-Specific Terms

Domain-specific terms such as treatment, dose, recommendation, exposure, hint, deployment, or remediation MAY be used when discussing a concrete domain. On first use in a cross-domain wiki page, agents SHOULD map the domain term back to canonical terminology such as action, control input, intervention, event, observation, or exogenous variable. For example, a healthcare treatment is usually an intervention; a numeric drug dose can be a control input; an observed medication order in confounded EHR logs is logged decision data, not automatically a clean causal action.

Domain Analogs

Canonical Term	RL / Control Analog	Healthcare Analog	Recommender / Bandit Analog	Education Analog	Observability Analog
time series	rollout history	longitudinal patient record	user interaction history	student interaction history	metric timeline
trajectory	episode, rollout	care pathway	session or user journey	problem-solving trace	incident timeline
state	environment state	patient state	user state	student knowledge state	service/system state
observation	observation, sensor reading	vitals, labs, charted measurements	context features, feedback	response, correctness, hint metadata	node/edge metrics
action	action	clinical decision	recommendation, exposure	assigned item, hint, intervention	operator action when logged
control input	actuator command	dose or device setting when numeric	ranking/exposure parameter	tutoring policy choice	autoscaling or traffic-control command
intervention	policy action, environment intervention	treatment, procedure, medication	randomized exposure, campaign treatment	instructional intervention	remediation, rollback, deployment change
exogenous variable	disturbance	comorbidity, external event	seasonality, external campaign	calendar/school context	traffic spike, hardware failure
event stream	discrete transition log	clinical event sequence	clickstream	tutor/student event log	alert/incident log
passive dynamics model	model without actions	disease progression model without treatment intervention	user-behavior forecast without recommendations	learning-progress forecast	metric forecaster
action-conditioned world model	transition/dynamics model	treatment-response model	user-response model	learning-dynamics model under interventions	controlled system model

Usage Rules

• Agents MUST prefer time series, multivariate time series, irregular time series, event stream, and trajectory over vague sequence terminology.
• Agents MUST use multivariate time-series model instead of vector/time-series model.
• Agents SHOULD use time-series model without the word numeric when the data is numeric, because numeric is the default in this wiki.
• Agents MUST explicitly identify non-numeric modalities when they are central: event stream, categorical event sequence, graph time series, text-conditioned time series, or image/video trajectory.
• Agents MUST distinguish action, control input, intervention, event, and exogenous variable.
• Agents SHOULD use domain-specific terms such as treatment only after mapping them to canonical time-series and world-model terms when ambiguity is possible.
• Agents MUST call out confounding when discussing observational healthcare, recommender, education, or logged decision datasets as if they supported causal intervention modeling.
• Agents SHOULD map domain-specific names back to canonical terms on first use when a page crosses domains.

Examples

• Say: multivariate time-series models.
• Do not say: vector/time-series models.
• Say: event stream or categorical event sequence for clickstreams, tutor logs, and clinical-code sequences.
• Say: graph time series for time-varying node/edge metrics such as service telemetry.
• Say: action-conditioned world model when future prediction depends on actions, control inputs, interventions, or mapped domain analogs such as treatments or recommendations.
• Say: passive dynamics model when there is no controllable action channel.