# Simpler references under the stronger motor

Register before new execution. Eight new first attempts and six reused controls.
This is development qualification of one inspected source ancestry (00976), one
corridor and two already inspected seeds (96162, 96167), not a generalization test.

## Question and intervention

Original SONIC weights with fixed IK-lowered reference complete clear, original
beam and beam lowered by 20 mm in both seeds. Does the unmodified local-duck source
retain those complete tasks with the same motor? Separately, can a nominal motion
complete the clear-space task? Extra posture construction may remain necessary,
or the stronger motor may make the simpler source sufficient in these conditions.

The sole behavioral intervention is the selected fixed candidate from the same
four-candidate native bank. Six new local-duck attempts cover three layouts and
two seeds. Two new nominal attempts cover clear space and the same seeds. Reuse
the six original-motor IK-lowered runs from the completed motor comparison. Their
results and acquisition costs remain old data. No nominal obstacle task is tested.

Freeze original checkpoint bytes beside the SAME scene evaluation config used by
the controls; release defaults would change mass randomization. Preserve all
task/scene geometry, goal, floor, body model, observation noise, motor gains,
action scaling, histories, clock and scoring. Sources are unchanged; no tuning,
phase shift, reference rescaling, new generation or post-result exclusion.

## Gates and execution

Verify loaded policy tensors and the 640D reference / 64D token / 930D history /
29D action contract. Before task actions, require exact actual entry, recorded
dynamics, actuator gains and all four native libraries against the corresponding
original-motor IK control. Initial proprioception/history must match. Reconstruct
the selected candidate's first query; reference equality to the other candidate
is diagnostic, not required, because the reference is the intervention. If first
references match, first actions must match with the unchanged motor. Audit every
later query against its declared library, actual root orientation and clock.
Compare first-physics Torch RNG; instrumentation must not consume it.

Retain the 1 s standing transition, 50 Hz control, ten future reference samples at
100 ms spacing, 200 Hz physics and the original 500-tick / 10 s deadline. Run
sequentially, clear local duck first as an infrastructure check; do not expand
the cap based on outcomes. A task failure is a result and does not cause a retry.
Infrastructure mismatches abort the launch. At most one separately preserved
infrastructure-repair retry per affected case; never silently overwrite a trial.

## Outcomes and analysis

Primary outcome is the unchanged full task: whole-body passage and required duck,
15 upright recovery ticks followed by a fresh 50-tick stationary hold inside the
0.5 m three-dimensional goal region, speed <= 0.1 m/s, no forbidden environmental
contact > 1 N at any substep and no pelvis-height fall, within 10 s. Foot-floor
support is allowed; foot-obstacle contact is forbidden. No imitation termination,
hidden reset or self-contact-absence claim. Clear tasks use their original score.

Report every task, stop cause/body, success-conditioned completion time and paired
gains/regressions of local duck versus lowered (six pairs); nominal clear outcomes
are a separate two-case qualification, not a six-layout competitor. Distinguish
earliest forbidden-force threshold crossing from peak force. Replay every issued
token and action with the frozen codec. Count full action-history duplicates
across layouts; physical executions do not imply independent source demonstrations.
Report native attempts, task control steps, physics samples and wall time. Keep
tracking error and source geometry as diagnostics, separate from task success.

## Decision branches

If local duck retains all six complete tasks, retain it as the simpler supported
fixed option here; do not claim equivalence or general robustness. If it loses
tasks, keep IK-lowering for those requests and report the loss before considering
new reference construction. Nominal clear failures preclude assuming the cheap
ordinary option is qualified. A successful nominal result has no obstacle claim.

Then register public integration using a NEW motor-specific ledger, not the old
scene-motor ledger. Compare against strongest fixed candidates and move to a
separate clearance/entry panel. Public choice, online continuation, independent
sources/scenes, partial perception and hardware remain separate untested steps.
No training or public-selector promotion belongs to this eight-attempt panel.