# Public capability selection and one clearance probe

Registered before new native execution, 19 September 2026. This is a bounded development comparison in simulation, with one inspected motion ancestry (00976) and the existing known-map corridor.

The preceding fixed-foot lowering package completed 3/3 original-beam and 3/3 clear tasks. The geometric selector still prefers local duck, whose matched beam result is 2/3. The next uncertainty is whether prior execution qualification connects this available capability to public selection, and whether the resulting choice remains executable when clearance changes.

## Frozen comparison

- Eighteen first attempts: three modes × clear/original beam/beam lowered by 20 mm × seeds 96162 and 96167. The former is a retained development failure; the latter is a fresh execution seed, not a held-out source or scene.
- Modes: existing geometry/cost `public_select`, `public_qualified` with a frozen prior-evidence eligibility mask followed by the same geometry/cost rank, and the strong `fixed_lowered` execution control. All use the same four-candidate bank and motor checkpoint.
- The eligibility rule requires at least two attempts and all successes in the designated prior fixed-candidate panel for the clear/beam context. The ledger records the actual prior outcomes, case IDs and audit hashes. It consumes no current evaluation outcome, source answer trajectory or task identity. The rule is a simple development baseline, not a fitted probability, novelty claim or safety certificate.
- For clear space all four candidates meet the evidence gate; the old posture cost favors nominal. For the original beam, only lowered meets the evidence gate. Geometry still must pass the unchanged 20 mm sampled-separation and 0.5 m endpoint screens. An empty eligible set is a request rejection before action; it is not a successful navigation, qualified safe stop or proof of physical impossibility.
- Qualification is tied to the motor hash, native-motion hashes, public goal/map and measured standing-entry profile. Only clear space and the same beam's vertical interval [−20, 0] mm are admitted in this probe. Prior execution evidence covers only the original height. The changed height is explicitly labeled extrapolation.
- The lower-beam asset changes only the obstacle's Z translation. Task geometry and scoring gate change consistently. Floor, XY location, orientation, dimensions, goal and initial state are unchanged. This avoids the adjacent program's separate goal/floor confound. No detour candidate exists in this bank; clear space is the supported control.

## Execution and analysis

Use the preceding 1 s standing transition, native 50 Hz packing with a 10-sample/100 ms-spaced future, independent bank and clean relative orientation. Preserve the same teacher, 10 s / 500-tick clock, passage/duck/recovery criteria, 15 recovery ticks, fresh 50-tick stationary hold, 0.5 m goal tolerance, forbidden environmental contact >1 N and fall condition. No imitation termination, reference reset, new weights, online moving seam or vision input is introduced.

Record all attempts, process costs and failures. No scientific retuning or physical retries. At most one isolated infrastructure repair retry per affected condition, preserving the first attempt; a physical task failure is a completed process, never an infrastructure retry.

Reconstruct scores from body/contact traces, reference vectors from the independently loaded bank and actual pre-action orientation, and tokens/actions through the frozen codec. Require paired initial states and dynamics to match. Compare new public beam executions with fixed lowered, and old/new clear executions with each other. Retain overlapping historical controls as parity checks, not independent new evidence.

Primary reporting: complete task counts per layout/mode; paired gains and regressions; selected candidates; false admissions (an admitted choice that fails the full task); contact body and time; and conditional completion times. Original and changed-height results remain separate. Describe first differences in inputs/actions to establish whether the map changes public decisions. A fixed reference succeeding does not demonstrate feedback planning; matching the fixed control is the expected integration result, not a gain over that control.

If public selection inherits original-beam improvement but fails under changed clearance together with the fixed control, prioritize execution-envelope qualification and reference construction. If public and selected fixed differ, repair the interface first. Do not tune the ledger on these new outcomes and rescore this panel. Independent sources, lateral obstacles, vision, hardware, moving replacement and the necessity of fixed-foot IK remain unresolved.