# Qualify the existing sustained-duck alternative

Registered before execution on 19 September 2026 UTC. This implements the first item of the [previous next-step plan](/kimonav/progress/#next): test a stronger simple fixed candidate before attributing a benefit to online reference replacement.

The library already contains related nominal, local-duck and sustained-duck motions derived from development source 00976. Only local duck has been physically tested through the current KimoNav clean-orientation, nominal-standing interface. Sustained duck widens the lowering window; it is not a new model, independently acquired skill or unseen source.

Freeze 12 first attempts: local duck versus sustained duck × original clear/beam layouts × seeds 96162, 96164 and 96165. Seed 96162 is the retained failure and 96164 is the retained success under local duck. Seed 96165 is a fresh execution seed, still on the same inspected source and corridor. Run every physical outcome without changing scientific parameters. At most one software/infrastructure repair retry per case is allowed, with the first attempt preserved. Physics failures are not retried.

Retain the scene teacher `afd649cf…`, nominal standing start, one-second entry prefix, native 50 Hz q/dq conversion, ten future samples at 100 ms offsets, clean measured-relative orientation, known-map assumptions and complete task evaluator. The beam placement, destination, 500-tick deadline, 0.5 m 3D destination region, 0.1 m/s terminal speed, upright recovery and fresh 50-tick hold remain unchanged. Environment contact >1 N and the inherited pelvis-height fall rule remain active. Historical self-contact qualification is separate.

The treatment selects the existing sustained candidate instead of local duck. Neither arm chooses a new motion online or replaces a moving reference tail. Both run through the same native library executor; the public selector's ranking is retained only as a diagnostic. Native source snapshots used in the previous panel remain frozen. A small mode-dispatch extension enables the new fixed candidate without altering old modes.

The primary comparison is paired complete-task success. Keep clear-scene regressions and component/contact failures visible. Compare completion times only for pairs in which both succeed. Verify physical initial-state/dynamics equality and exact local-control continuity with the four overlapping prior cases. Reconstruct every issued reference and separately replay the frozen codec.

An exploratory diagnostic will compare measured and planned root/joint geometry before contact, using named-joint FK and explicit kinematic substitutions. It can localize representation discrepancies but is not a physical causal intervention. No substitution geometry or reduced reference error becomes a task-success label.

If the sustained candidate rescues the hard seed while retaining matched successes, retain it as a stronger development baseline. A fixed reference's success does not qualify arbitrary online seams or robustness to observation/scene changes. If it fails or causes regressions, preserve that result and prioritize measured-state construction or motor capability. No student training, scene alteration or automatic fallback follows this test.