Chapter 25 — Failure-Diagnosis Manual

The big picture

A tiny story

A doctor hears, “The patient cannot move.” That could mean muscle, nerve, medicine, pain, or a disconnected monitor. Replacing everything may remove the symptom while destroying the diagnosis.

A still LeWM agent can mean collapsed representation, action-deaf dynamics, flat cost, poor CEM samples, zeroed denormalization, or a wall blocking a valid command. Start with one frozen failing specimen, not a favorite theory.

The real rule

Walk backward from the last physical fact:

1. Did the environment receive the intended raw action?
2. Did denormalized actions vary and survive clipping?
3. Did CEM’s proposal and elites change?
4. Did candidate costs vary?
5. Did rollouts respond correctly to action?
6. Did representations keep the needed distinction?
7. Were data windows, timing, and actions valid?

Log both sides of every interface. “Selected action” is incomplete unless it says normalized, clipped, repeated, block-level, or environment units. Freeze planner randomness and the candidate batch.

For action use, hold one history fixed and compare recorded, swapped, zeroed, and permuted actions against matching environment transitions. Output change is stronger evidence than a nonzero gradient, but change in the wrong direction is still wrong.

For self-fed failure, replay the same recorded actions in recorded-context and self-fed lanes. Mark the first divergence and slice by doorway, contact, and support. Do not involve CEM until this simpler comparison is interpretable.

The trick that can fool us

Two final videos show an unmoving dot.

Agent One: candidates and rollouts vary; CEM updates correctly; costs are flat because the current embedding was duplicated into the goal slot. Oracle cost separates the frozen candidates.

Agent Two: costs separate and CEM chooses a strong action; the wrong normalization column converts it to zero before environment execution.

The symptom matches; the earliest observable breaks do not.

Other discriminating tests:

• flat cost: verify goal identity, broadcasting, terminal spread, shuffled goals, and oracle task progress;
• coverage vs bug: cross a support atlas with known-transition replay and a tiny repeated-transition fit;
• tiny overfit: add complexity in order—one transition, two actions from one context, a short sequence, then doorway/contact;
• SIGReg dominance: log each loss and module gradient, spread, prediction, and a controlled weight sweep;
• small batch: repeat statistics across batches and random projections, count distinct examples per relative time position, and vary composition.

More projection directions cannot create missing observations. Gradient accumulation does not enlarge the population that a batch-wise SIGReg test actually saw.

Experiment receipt

Archive source revisions, checkpoint hash, composed config, dataset indices, raw observations, normalized and raw actions, candidate batch, costs, and random states before repair.

The frozen SIGReg weight is 0.09; the method text says 0.1 unless stated otherwise. The PushT high-weight ablation is task-specific, not a universal safe interval. A large printed loss or gradient is not “dominance” until held-out dynamics and behavior change under controlled weights.

The diagnostic tree is a tutorial construction around the original paths in LeWorldModel v3 and frozen jepa.py, train.py, and eval.py, commit 8edfeb3. Oracle repair implicates the replaced layer under that test; it never proves every unchanged layer correct.

Quick check

1. Why must fixed-action replay come before planner tuning?
2. What can disprove “CEM is broken” for a motionless agent?
3. How does an earliest observable break differ from a unique root cause?