CLEAR-LeWM v0.5.0: minimal repair and precise completion

What changed

CLEAR-LeWM v0.5 unifies the public benchmark around two auditable modes.

• Moderate is a minimal LeWM-compatible correction: it removes initially solved pairs, repairs Reacher joint topology and TwoRoom collision/data defects, and otherwise preserves the released PushT and Cube predicates.
• Strict inherits Moderate data hygiene and physics, then evaluates precise task completion: T-only PushT pose, symmetry-aware cube pose, physical Reacher endpoint, and legal precise TwoRoom navigation.
• Canonical seed 0, 1, and 42 manifests are included for all four tasks and both modes.
• Evaluator code, task guides, GitHub Pages media, result validation, and checkpoint provenance are versioned together.

Official LeWM reference

Task	Moderate seed 42 (model / random)	Strict seeds 0,1,42 (model mean / random mean)
PushT	88% / 3%	70.33% / 5.00%
Cube	51% / 15%	26.33% / 6.00%
Reacher	40% / 5%	43.00% / 5.00%
TwoRoom	81% / 6%	58.33% / 1.67%

All checked-in model results use pinned official high-epoch LeWM checkpoints, 100 fixed pairs per cell, pure CEM 300 x 30 with top-k 30, solver batch size 1, and strict 303/303 tensor loading. Paired random controls use the same manifests and policy seeds.

Reproducibility

Use manifests/v0.5/, results/v0.5/, and checkpoints/official-v0.5.json. Exact gates are normative in EVALUATION_SPEC.md. Previous releases remain reproducible from their Git tags; their numbers must not be mixed with v0.5.