# Fixed foot-pose constrained lowering qualification

Registered before the first physical attempt on 19 September 2026 UTC. The previous local-versus-sustained study retained the same torso–beam failure. This panel asks whether a single more deeply lowered fixed reference helps, before implementing actual-state reference replacement.

One previously inspected source, 00976, remains the source family. Compare fixed local duck and the new fixed lowered candidate × original clear/beam tasks × seeds 96162 (retained failure), 96165 (previous success) and 96166 (fresh execution seed): 12 first attempts. Retain every physical outcome without retuning or retry; at most one preserved infrastructure repair retry per case is allowed. No training or hardware execution. Shared GPU use is allowed within available memory; wall-clock timings are not a performance endpoint.

The 8 cm requested extra lowering is motivated by the prior difficult seed's approximately 8 cm precontact height error. It follows the existing local-duck lowering profile. An initial strict 8 cm kinematic candidate failed its foot-pose constraints at source frame 42 and did not enter physics. The adopted treatment projects that requested depth using both original full foot poses, unchanged root XY/orientation and unchanged upper-body joints. The 12 leg joints may change by at most 0.6 rad, with 0.03 rad range margins where available and no worsening of an inherited near-limit value. A constrained per-frame solve is followed by a seven-frame minimum filter and finite Gaussian smoothing (sigma one source frame, radius three), and an exact fixed-depth re-solve. This is one constrained geometric construction, not a physical amplitude grid. The achieved peak extra lowering is 48.108 mm, not the requested 80 mm. Source-space added joint steps are at most 0.10 rad/frame; full foot-pose residual limits are 0.1 mm and 0.0001 rad.

Preserving feet along their original stance and swing paths does not establish contact, balance, friction or support compatibility. No force-derived ground truth enters the controller. Unknown support remains unknown. The source conversion and unchanged one-second nominal-standing transition are applied before the same native 50 Hz loader. All original bank candidates are retained unchanged, with a fourth named candidate added. Both trial arms use the same library size, motor, clean relative orientation, actual proprioception/action history, exact geometry and original clock. Selection is fixed; the public ranking is diagnostic only. There are no moving seams or online reference replacements.

Primary outcome is the existing complete scene task: entire-body passage, required duck event under the beam, upright recovery, 0.5 m 3D destination region, terminal speed at most 0.1 m/s, 15 qualifying upright ticks followed by a fresh 50-tick hold, no prohibited environment contact above 1 N, no inherited pelvis-height fall, and the original 500-tick/10 s deadline. Keep the earlier all-nonfoot/self-contact qualification separate. Do not alter the beam or score to rescue a result.

Audit all references and token/action outputs, initial-state/dynamics pairing and all physical score components. Compare the four overlapping local controls with the previous duration panel. Verify native resampling's foot-trajectory differences; the tiny source IK residual is not the native interpolation error. Report paired success and contact failures, completion time only on common successes, and descriptive precontact root/joint realization. Observed foot load and slip remain diagnostics. Episode frames are correlated, not independent samples.

Promote only a complete-task rescue with retention of matched successes. If lowering is geometrically constructible but not executed successfully, do not credit future feedback, admit failed trajectories as successes, or train a student on a claimed rescue. Freeze the next intervention only after this comparison is complete.