# Distinct clearance requests: fixed capability and shadow admission

Registered before new simulator outcomes on 20 September 2026. This is a development diagnostic on motion ancestry 00976 and the existing corridor, not an independent-scene generalization study.

## Question and fixed budget

Can either existing fixed reference complete requests at unmeasured beam heights, and what coverage/errors would the unchanged admission rules produce? Execute exactly **12 first physical attempts**: beam offsets −10, −30 and −50 mm relative to the original beam × seeds 96171 and 96173 × fixed local duck and fixed IK lowered. The seed search scope and previous original-motor panels are recorded in SEED_ACCESS.json. Six distinct requests have two measured alternatives each. Repeated frames and related references are not independent scenes.

For each seed, execute heights in order −10, −30, −50 mm, with local duck before lowered at each height. Fixed controls run even when the geometry screen or public profile rejects the request. They are explicit capability probes, not public admissions. No physical failure retries, training, motion generation, phase/amplitude/gain tuning or outcome-dependent selection. Infrastructure failures must be retained and separated; stop and document any required correction before a separately identified retry.

## Held conditions and intervention

Copy the original SONIC checkpoint and SAME scene evaluation config from the completed public integration. Keep the four-member bank bytes, source/native interpolation, 640D reference / 64D token / 930D history / 29D action interface, 1 s standing transition, 50 Hz control, 200 Hz physics, original goal and 10 s deadline. Translate only beam center Z in the task, scoring gate and USD; preserve floor, dimensions, orientation and horizontal placement. The choice of local duck versus lowered is the paired intervention.

New seeds do not have old exact-state controls. Validate actual standing against the existing declared profile, including separation of configured calibration offsets from actual body deviations. Verify loaded motor tensors, common config, actuator gains and clocks before task actions. Within each new request, the lowered arm must match its new local-duck counterpart's initial state, recorded dynamics, gains, native libraries and first proprioception/history. Compare first actions only if their references match. Reconstruct every issued reference against its selected library; replay all tokens/actions with the frozen codec.

Use frozen adjacent research-module snapshots. Do not edit the adjacent working tree, interfere with other GPU jobs or import their changed task/terrain code. Execute serially when at least 3,500 MiB GPU memory is free. A resource delay is not a physical attempt. No hardware runs.

## Outcomes and scoring

Retain whole-body passage, ducking, 15 upright recovery ticks followed by a fresh 50-tick stationary hold inside the 0.5 m 3D goal region at speed ≤0.1 m/s, no forbidden environmental contact above 1 N at any 200 Hz substep and no pelvis-height fall, within 500 control ticks. Foot-floor support is allowed; self-contact absence is not claimed. No imitation termination or mid-episode reset.

Report all 12 outcomes, contact/fall/timeout causes, complete tasks per reference over six requests, paired gains/regressions and success-conditioned completion time. Best-of-two measured fixed success is a retrospective solvability reference, not a deployable oracle or proof that all motions were searched. Count distinct full action histories. Inspect actual pelvis/body/contact traces to interpret differences without changing the primary score.

## Frozen shadow decisions

Before the first task action of each run, save geometry-only choice and candidate checks using the unchanged 20 mm reference separation margin, endpoint threshold and minimum-posture-cost tie break. Apply the unchanged finite-context public ledger to the actual request. It is expected to reject all three new contexts. Validate standing/runtime separately; do not treat another rejection reason as the expected finite-context rejection. The shadow decision uses task/map, measured entry, the fixed bank and OLD ledger only; no new outcome informs it.

These shadow decisions add zero physical executions. A rejected request contributes zero complete-task coverage, not a successful stop. Combine each saved choice with its corresponding measured fixed outcome only when that candidate was executed on the same request. Choices of nominal or sustained without a measured run are unknown, not inferred success or failure. Report admitted-but-failed requests, rejected requests with a successful measured fixed alternative, unknown selected outcomes, and complete tasks over all six requests. Geometry candidate rejection despite successful fixed execution remains visible.

All shadow scores are retrospective evaluations of pre-action decisions using these fixed outcomes. They do not establish online stopping, public-interface execution on the new contexts or a continuous safe envelope.

## Decision branches

If local duck covers all requests, retain it and move to independent sources and combined height/width conditions instead of extending the height grid indefinitely. If IK lowering adds tasks, retain its conditional role and investigate the realized clearance difference. If both fail, record the tested capability gap; do not conclude that every possible motion is infeasible. Compare the geometry rule with the finite gate to expose useful coverage and false admissions, without claiming that finite rejection is a contribution.

Specify any new execution-informed acceptance mechanism only after this diagnostic. Outcomes used to fit thresholds or models become calibration data; a separate panel must evaluate that mechanism. Preserve this protocol, all failures and the prior public-interface results.