Appendix F — Glossary

Define core terms with a formal definition, one-sentence explanation, everyday analogy, and common misconception.

Course progress Course outline 48 of 48 lessons available

Part 0 — Reading Guide: What Exactly Are We Going to Learn?

01 Chapter 0 — Before You Begin available now

Part 1 — World Models: An Internal Sandbox for the Agent

Part 2 — Turning Images into State: The LeWM Architecture

06 Chapter 5 — Trajectory Data: To the Model, the World Is Not an Image Collection available now
07 Chapter 6 — The Visual Encoder: Issuing a “State Passport” for Every Frame available now
08 Chapter 7 — The Dynamics Predictor: Moving Time Forward in the Mind available now
09 Chapter 8 — A Complete Forward Pass: Follow One Batch from Start to Finish available now

Part 3 — Preventing the Model from Cheating: Prediction Loss and SIGReg

Part 4 — Putting the Model into Action: Planning in Latent Space

Part 5 — Engineering Reproduction: From Paper to Running System

Part 6 — What Has LeWM Actually Learned?

Part 7 — Why “Accurate Prediction” Can Still Produce “Poor Planning”

Part 8 — From Reproducer to Researcher

Appendices

The big picture

FormalWhat the term names.
PlainA short story to carry.
BoundaryWhat it does not prove.

An analogy opens the door; the definition and misconception guardrail stop it from changing the machine.

A tiny story: the translator’s booth

One researcher says target and means the encoded next observation. Another hears goal and points to the planner’s image. A third hears teacher forcing and invents an EMA teacher that original LeWM does not have.

These words live in different lanes. Every row therefore keeps four windows: formal definition, plain explanation, analogy, and misconception.

A translator booth keeps formal definition, plain explanation, analogy, and misconception in four separate windows.

An analogy helps memory; it does not supply assumptions or evidence.

Model and representation terms

Term	Formal definition	Plain explanation	Analogy	Misconception
World model	Learned state/dynamics model predicting temporal or action consequences for reasoning/control.	It rehearses “what happens next?”	Rehearsal room.	It must render video or contain complete readable physics.
Latent state	Learned vector compressing an observation/history for prediction, analysis, or planning.	A compact internal state card.	Passport, not photograph.	Every coordinate is physical; one frame is automatically Markov-complete.
Target representation	Predictive-loss reference; in v3, the same trainable encoder encodes the shifted real successor without detach.	Encoded next observation to match.	Training destination ticket.	Goal image, EMA teacher, target network, or stop-gradient.
Action conditioning	Action/action block modulates a transition predicted from latent context.	Ask “what if I do this?”	Left/right simulator buttons.	Different symbols always produce different, correct causal branches.
Representation collapse	Degeneracy that removes task-useful variation or effective structure.	Different situations look identical inside.	One passport for everyone.	Any low dimension is collapse; collapse equals overfitting.
SIGReg	Finite random 1-D projections compared with a standard Gaussian as anti-collapse pressure.	Check many shadows of the latent crowd.	Inspect a city from many angles.	Finite shadows certify Gaussianity, topology, physics, or utility.

Prediction and control-loop terms

Term	Formal definition	Plain explanation	Analogy	Misconception
Prediction loss	Discrepancy between predicted future representation and encoded observed successor.	Did predicted next meet real next?	Forecast versus arrival.	Low one-step loss guarantees rollout or planning.
Teacher forcing	Sequence training conditioned on recorded context, not only earlier predictions.	Training walks from the real trail.	Supplied preceding piano bar.	Requires teacher network, EMA, or stop-gradient.
Rollout	Repeated dynamics under actions, often feeding predictions into later steps.	Build a future path step by step.	Copies from earlier copies.	Necessarily video, real data, or a feasible path.
Open-loop control	Execute a committed sequence without observation-based revision during that interval.	Follow the plan before looking again.	Printed driving route.	It has no model or never works.
Closed-loop control	Later actions use observations after earlier actions affect the world.	Look again and update.	Navigation reroute.	Feedback guarantees success or repairs wrong models/costs.

Planning terms

Term	Formal definition	Plain explanation	Analogy	Misconception
Planner	Finite-budget action-sequence selection using model, cost, and constraints.	Chooses trials; model estimates outcomes.	Travel agent.	Planner equals world model; search restores discarded information.
Cost	Scalar ranking rule for predicted outcomes/actions, usually lower-is-better.	Says which candidate looks preferable.	Judge’s scorecard.	Necessarily training reward or feasibility.
Terminal latent distance	v3 squared Euclidean cost between final predicted and encoded goal latents.	Ruler from endpoint to goal.	Straight map line.	Equals shortest path, progress, temporal distance, or reachability.
Decision-metric alignment	Agreement between predicted and realized candidate rankings under a protocol.	Does rehearsal choose reality’s winner?	Two judges rank contestants.	One correlation, probe, or one-step score guarantees it.
Reachability	Whether permitted dynamics/actions reach a state, often directionally within budget.	Can I get there under these rules?	Adjacent rooms with locks.	Symmetric, visual similarity, or certified by a learned score.
CEM	Sample action sequences, keep cost elites, refit proposal for finite rounds.	Winners shape the next audition.	Repeated tournament.	Exhaustive search, global guarantee, or immune to model error.
MPC	Plan, execute a configured prefix, observe, and plan again.	Trust imagination briefly, then check.	Sailor redraws each leg.	Always one action; removes bias; is a learned controller.

Evidence and diagnosis terms

Term	Formal definition	Plain explanation	Analogy	Misconception
Support mismatch	Queries lack adequate evidence in collected/training distributions.	Planner asks about a barely learned neighborhood.	Unsurveyed map blank.	Outside means impossible; global action frequency proves local support.
Model exploitation	Search prefers actions because systematic model/cost error looks attractive.	Optimizer finds a simulator loophole.	Optimize a rules typo.	Normal exploration, deception, or proof CEM is broken.
Decodability	A specified probe/split can recover a variable from fixed representations.	Can this reader extract it?	Clerk reads coded passport.	The model uses it; it is causal; it proves understanding.
Controllability	State distinctions influenceable/reachable by admissible actions under named dynamics/horizon.	What can controls change?	Elevator floor buttons.	Action sensitivity establishes controllability or observability.
Identifiability	Under assumptions/equivalence, latent model recovery is unique up to allowed transforms.	Is the hidden explanation uniquely recoverable?	ID valid in one jurisdiction.	Interpretability or true physics in arbitrary worlds.
Causal identification	Intervention effects/structure established by design or identification assumptions.	What changed because one factor changed?	Test one circuit switch.	Correlation, probe, separated branches, or any ablation proves cause.
Planning utility	Empirical decision usefulness under a declared task, planner, and budget.	Does it help choose actions here?	Map judged by trips.	Intrinsic, task-free scalar.
Action-conditioned predictive consistency	Matched action-rollout diagnostics; ACPC uses divergence, Invariance Radius, Separation Rate.	Stable for nuisance; different for consequence.	Repaint sign versus move road.	Consistency proves truth, calibrated uncertainty, or safety.
Independent reproduction	Evidence outside authors reruns/reimplements a claim under documented version/protocol.	Another investigator checks what they ran.	Second lab repeats one assay.	Author code or TwoRoom verifies every task/method.
Violation of Expectation	Matched surprise comparison for ordinary and deliberately perturbed event sequences.	Does error react to the chosen violation?	Costume versus motion-rule change.	Surprise proves intuitive physics, causality, or broad anomaly understanding.

Try to break the vocabulary

“The teacher forces the target goal through a rollout, so low latent distance proves the state is reachable.”

Every noun is real; every connection is wrong. Teacher forcing supplies recorded context. A target representation is an encoded successor; a goal belongs to planning. A rollout imagines latents under actions. Terminal distance ranks endpoints but does not prove budgeted directed reachability.

Repair the sentence by attaching every word to one object, phase, assumption set, and evidence type. Two mechanisms hiding under one word signal a category error.

Evidence receipt

Definitions follow LeWorldModel v3, frozen official code, LeJEPA v3, passive identifiability v1, controlled identifiability v2, and TwoRoom reproduction v1. Appendix E pins frontier versions. Cutoff: 2026-08-20. Analogies are tutorial constructions; independent evidence remains TwoRoom-only under its protocol.

Three quick questions

Where do target representation and goal image enter?
Why can information be decodable but unused and non-causal?
How do distance, reachability, and planning utility differ?