Key finding
Replay health
All 22 solver pairs completed 2 independent subprocess replays.
What This Run Shows
Key finding
Run-to-run verdict
Every tested solver pair was bit-exact under wp.config.deterministic='run_to_run'.
Key finding
Largest remaining gap
No remaining run-to-run gap was observed in the tested harness slice on this GPU/software stack.
Method
Independent replays
2 subprocess runs per supported scenario/solver pair.
Canonical hash
Hashes include each scenario's canonical core arrays, such as rigid state, particle state, losses, and gradients.
Fixed workload
60 frames, 4 substeps, 60 Hz, seed 0, world count 2.
Deterministic contacts
Contact scenarios use CollisionPipeline(deterministic=True) with the report's global Warp determinism mode.
Visual capture
Solver videos use one world and MP4 frames read through ViewerGL.get_frame().
Scenario Videos And Statistics
featherstone
240 frames encoded at 24 fps (960x540, stride 1).
mujoco
240 frames encoded at 24 fps (960x540, stride 1).
semi_implicit
240 frames encoded at 24 fps (960x540, stride 1).
vbd
240 frames encoded at 24 fps (960x540, stride 1).
xpbd
240 frames encoded at 24 fps (960x540, stride 1).
falling_cube
Falling Cube
Single free rigid body floor contact. Useful for detecting tiny contact drift after impact.
Final height mean
0.15 m
Final height range
0.15-0.15 m
Awake bodies
0
Mean speed
2.801e-04 m/s
featherstone
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration14.081 s
mujoco
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration41.97 s
semi_implicit
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration13.255 s
vbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration16.426 s
xpbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration13.808 s
mujoco
240 frames encoded at 24 fps (960x540, stride 1).
vbd
240 frames encoded at 24 fps (960x540, stride 1).
xpbd
240 frames encoded at 24 fps (960x540, stride 1).
box_stack
Box Stack
Twenty jittered boxes with many contacts. A compact stress case for ordering and reductions.
COM drift
0.002 m
Kinetic energy proxy
1924.727
Contact samples
6
Peak contacts
38
mujoco
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration48.201 s
vbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration16.806 s
xpbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration13.961 s
mujoco
240 frames encoded at 24 fps (960x540, stride 1).
vbd
240 frames encoded at 24 fps (960x540, stride 1).
xpbd
240 frames encoded at 24 fps (960x540, stride 1).
domino_chain
Domino Chain
Sequential impacts amplify small timing changes into visible propagation differences.
Fallen dominos
0
Last-domino time
not reached
First angle
14.324 deg
Tracked fall events
0
mujoco
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration50.447 s
vbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration20.5 s
xpbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration17.277 s
featherstone
240 frames encoded at 24 fps (960x540, stride 1).
mujoco
240 frames encoded at 24 fps (960x540, stride 1).
xpbd
240 frames encoded at 24 fps (960x540, stride 1).
arm_7dof
7-DOF Arm
Franka arm tracks deterministic sinusoidal PD targets from a fixed base.
Max target error
0.14 rad
Mean target error
0.14 rad
Peak PD torque
4.751 N*m
Peak contacts
0
featherstone
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration30.527 s
mujoco
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration58.379 s
xpbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration36.849 s
featherstone
240 frames encoded at 24 fps (960x540, stride 1).
mujoco
240 frames encoded at 24 fps (960x540, stride 1).
xpbd
240 frames encoded at 24 fps (960x540, stride 1).
humanoid
Humanoid Standing
Unitree H1 holds a standing pose; drift or falling exposes articulated-contact robustness issues.
Final root height
0.723 m
Fell worlds
0
Max COM XY drift
0.198 m
Peak contacts
28
featherstone
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration24.991 s
mujoco
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration50.329 s
xpbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftbody_q (0)
Dominant extra driftnone
Max core diff0
Duration20.746 s
semi_implicit
240 frames encoded at 24 fps (960x540, stride 1).
diffsim_ball
Diffsim Ball
Differentiable particle target with wall and floor contacts; core includes loss and initial-velocity gradients. Diffsim ball target gradient inspired by ``example_diffsim_ball``.
semi_implicit
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftfinal_particle_q (0)
Dominant extra driftnone
Max core diff0
Duration14.044 s
semi_implicit
240 frames encoded at 24 fps (960x540, stride 1).
diffsim_cloth_com
Diffsim Cloth COM
Small differentiable cloth target; COM uses an atomic accumulation before backpropagating gradients. Small cloth gradient scenario inspired by ``example_diffsim_cloth``.
semi_implicit
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftfinal_particle_q (0)
Dominant extra driftnone
Max core diff0
Duration10.41 s
semi_implicit
240 frames encoded at 24 fps (960x540, stride 1).
diffsim_spring_cage
Diffsim Spring Cage
Differentiable spring-cage target; core includes loss and spring rest-length gradients. Spring rest-length gradient inspired by ``example_diffsim_spring_cage``.
semi_implicit
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftfinal_particle_q (0)
Dominant extra driftnone
Max core diff0
Duration9.834 s
vbd
240 frames encoded at 24 fps (960x540, stride 1).
vbd_cloth_patch
VBD Cloth Patch
Pinned cloth patch with ground contact, using VBD bending-aware coloring. Pinned cloth patch evaluated with the VBD solver.
Particles
154
Pinned particles
22
Height range
0.664-1.286 m
vbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftparticle_q (0)
Dominant extra driftnone
Max core diff0
Duration17.605 s
vbd
240 frames encoded at 24 fps (960x540, stride 1).
vbd_soft_body
VBD Soft Body
Pinned tetrahedral block with ground contact, using VBD particle/tet solve paths. Pinned tetrahedral block evaluated with the VBD solver.
Particles
160
Pinned particles
32
Height range
0.945-1.24 m
Tets
360
vbd
2/2 subprocess replays matched byte-for-byte with CUDA graph enabled.
ExecutionCUDA graph
Unique hashes1
Dominant core driftparticle_q (0)
Dominant extra driftnone
Max core diff0
Duration18.015 s
Solver Matrix
| Scenario | Solver | Verdict | Unique hashes | Dominant core signal | Max core diff | Max extras diff | Duration |
|---|---|---|---|---|---|---|---|
| 7-DOF Arm | featherstone |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 30.527 s |
| 7-DOF Arm | mujoco |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 58.379 s |
| 7-DOF Arm | xpbd |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 36.849 s |
| Box Stack | mujoco |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 48.201 s |
| Box Stack | vbd |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 16.806 s |
| Box Stack | xpbd |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 13.961 s |
| Diffsim Ball | semi_implicit |
BIT-EXACT | 1 | final_particle_q (0) | 0 | 0 | 14.044 s |
| Diffsim Cloth COM | semi_implicit |
BIT-EXACT | 1 | final_particle_q (0) | 0 | 0 | 10.41 s |
| Diffsim Spring Cage | semi_implicit |
BIT-EXACT | 1 | final_particle_q (0) | 0 | 0 | 9.834 s |
| Domino Chain | mujoco |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 50.447 s |
| Domino Chain | vbd |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 20.5 s |
| Domino Chain | xpbd |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 17.277 s |
| Falling Cube | featherstone |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 14.081 s |
| Falling Cube | mujoco |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 41.97 s |
| Falling Cube | semi_implicit |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 13.255 s |
| Falling Cube | vbd |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 16.426 s |
| Falling Cube | xpbd |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 13.808 s |
| Humanoid Standing | featherstone |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 24.991 s |
| Humanoid Standing | mujoco |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 50.329 s |
| Humanoid Standing | xpbd |
BIT-EXACT | 1 | body_q (0) | 0 | 0 | 20.746 s |
| VBD Cloth Patch | vbd |
BIT-EXACT | 1 | particle_q (0) | 0 | 0 | 17.605 s |
| VBD Soft Body | vbd |
BIT-EXACT | 1 | particle_q (0) | 0 | 0 | 18.015 s |
Observed Gaps And Scope
Observed gaps in this run
These notes are generated from the actual solver results above rather than a static message.
- No observed run-to-run determinism gaps remain in this tested report slice.
What the report does and does not prove
The current harness is precise about repeated local replays, but it intentionally stops short of broader claims.
- This report only checks repeated runs on the same machine, driver, GPU, and Warp/Newton build; it does not prove cross-GPU or cross-driver reproducibility.
- The verdict is based on final core state hashes. Videos and scenario extras are diagnostic evidence, not the primary pass/fail signal.
Controls in this harness
wp.config.deterministicis set before heavy Warp/Newton imports and propagated to subprocesses.- Each scenario owns a seeded NumPy RNG, so pose jitter and control trajectories are reproducible.
- The comparison boundary is process-level: no Python or GPU state is reused between replays.
- Only final core state decides the verdict; scenario extras are diagnostic and reported separately.