Lab

How we employ Haga.

Running experiments with reproducible seeds, defined thresholds, and shown failure cases — the same adversarial methodology on world-model outputs and the policies trained inside them.

Each experiment below is a narrow, auditable instrument: one question, defined thresholds, and numeric reports with shown failure cases. We widen the suite only when the evidence already in hand supports it.

Methodology — thresholds, seeds, claim boundaries, and links to technical specs.

• Policy verification

  Shipped

  Lift stress

  Tiered mass/friction stress on robosuite Lift: success degrades under severe stress, with grasp-slip as the dominant failure mode.

  success rate, nominal → severe

  1.00 → 0.26

  success rate, nominal → severe

  severe episodes: grasp acquired, then lost

  74%

  severe episodes: grasp acquired, then lost

• Policy verification

  Shipped

  Stack stress

  Tiered mass/friction stress on robosuite Stack: success degrades under severe stress, with grasp-slip as the dominant failure mode.

  success rate, nominal → severe

  0.96 → 0.20

  success rate, nominal → severe

  severe episodes: grasp acquired, then lost

  80%

  severe episodes: grasp acquired, then lost

• Policy verification

  Shipped

  PickPlaceCan stress

  Tiered mass/friction stress on robosuite PickPlaceCan: success degrades under severe stress, with grasp-slip as the dominant failure mode.

  success rate, nominal → severe

  1.00 → 0.24

  success rate, nominal → severe

  severe episodes: grasp acquired, then lost

  76%

  severe episodes: grasp acquired, then lost

• Policy verification

  Shipped

  Door stress

  Panel-scaled mass/friction stress on robosuite Door with a success-primary gate: shallower degradation curve, door-did-not-open failures documented.

  success rate, nominal → severe

  0.60 → 0.52

  success rate, nominal → severe

  severe episodes: door did not open

  48%

  severe episodes: door did not open

• World-model checker v0

  Shipped

  Physics checker

  Calibrated physics-violation detector: 1.000 recall on 399 violated trajectories, 0 false positives on 200 negatives.

  recall on 399 violated trajectories

  1.000

  recall on 399 violated trajectories

  false positives on 200 negatives

  0

  false positives on 200 negatives

• Tracking + CogVideoX cohort

  Shipped

  Physics-IQ

  Frozen VIDEO_CHECKS: real quiet (0%); discovery n=6 post-hoc 100%; held-out protocol v1 n=9 100% [0.701, 1.000] — all via `static_hover`.

  real negative control

  0%

  real negative control

  held-out flag rate (n=9 [0.701, 1.000])

  100%

  held-out flag rate (n=9 [0.701, 1.000])