Datasets:

ClarusC64
/

vljepa_physics_consistency_probe_v01

video_scenario stringclasses 10 values	given_frames stringclasses 5 values	physics_principle stringclasses 10 values	expected_inference stringclasses 10 values	failure_mode stringclasses 10 values	stop_condition stringclasses 10 values
ball rolling down ramp	frames 1-3	acceleration	velocity increase over time	constant speed guess	stop after velocity claim
person drops keys	frames 2-4	gravity	downward path + terminal contact	hover fiction	stop after path
toy car hits block	frames 1-3	conservation of momentum	car slow + block shift	energy creation	stop after shift
ice cube melts in sun	frames 3-6	phase change	solid→liquid time expectation	instant melt	stop after time note
swing pushed gently	frames 1-4	pendulum motion	periodic arc prediction	random drift	stop after arc
book slides on table	frames 2-5	friction	deceleration curve	perpetual motion	stop after deceleration
skateboard stops uphill	frames 1-3	gravity + friction	rollback potential	stays suspended	stop after rollback call
water splashes basin	frames 2-4	fluid dynamics	diffusion outward exaggeration	geometric shape naming	stop after spread
magnet near nails	frames 3-6	attraction	possible movement toward magnet	electric force confusion	stop after movement
tree sways in wind	frames 1-4	external force	oscillation direction	no effect claim	stop after vector

VL-JEPA Physics Consistency Probe (v01)

Tests if a model can track basic physical expectations from partial video context.

This dataset checks: • continuity over frames
• cause and effect
• prevention of impossible events
• where inference must stop

Columns: video_scenario
given_frames
physics_principle
expected_inference
failure_mode
stop_condition

Use cases: • early signal of inference drift
• baseline for robotics pre-training
• measurement of hidden confabulation

Evaluation goal: Can the model keep contact with physical reality without inventing states?

Files: vljepa_physics_consistency_probe_v01.csv