# Public requests with the qualified original motor

Register before six new first executions. This is one bounded integration check
on the familiar source ancestry00976, corridor and seeds96162/96167. It is not a
new source/scene generalization test, training run or learned admission study.

## Question and frozen intervention

The original SONIC motor with local duck passes all six clear/original-beam/
beam-minus20mm tasks. Nominal passes two clear tasks. Connect these supported
references to the actual public task/map/entry path, retaining the same motor,
configuration, independent four-candidate bank, reference clock and full score.

Register six new public_qualified requests: three layouts × two familiar seeds.
Fourteen existing qualification executions (eight simpler-reference plus six
original-motor IK-lowered) populate the prior ledger. Six of those are selected
fixed pairs: nominal for clear and local_duck for both beam heights. All14 remain
old data; do not count ledger use or offline probes as new physical executions.
Retain local-duck clear and lowered runs as additional strong fixed comparators.

The sole behavioral change relative to each selected fixed pair is public choice
through the frozen ledger/profile. Do not alter the task, reference, motor,
observation corruption, default calibration events, dynamics, scoring or deadline.
The source snapshots record the small callback extension that dispatches the
versioned profile validator; reference packing and tracking remain unchanged.

## Admission contract

Use a new finite-context v3 ledger. Bind checkpoint, SAME common scene config,
whole bank bytes/candidate hashes, actual control/physics dt, maximum steps and
actual actuator stiffness/damping fingerprint. Original release defaults have a
different mass range and are not substituted. Loaded policy tensors, entry,
recorded dynamics, gains and all native libraries must match the selected fixed
control before any task action.

Match the measured standing state, treating configured calibration offsets
separately from actual body deviations. Match the complete operational request
and score to exactly one of the three prior task contexts, including destination,
obstacles, scene hash, hold/deadline/recovery and contact-envelope conditions.
These are finite contexts, not a continuous safe clearance envelope. An unseen
midpoint height, new destination or changed hold is explicitly unsupported.

A candidate requires at least two designated prior original-motor first attempts
and no observed failure in that exact context, then must pass the unchanged
reference geometry screen. Rank admitted candidates by existing posture cost.
Untested sustained and nominal obstacle cases remain unqualified. The ledger
contains only prior outcomes, never the eventual result of the current episode.

Selection may read task/map/entry, the motion bank and frozen qualification data.
Task IDs, motion labels and inactive compatibility reference files do not select
the candidate. Paired-control files are diagnostic checks, not selection inputs.
No safe-stop or physical impossibility claim follows from a rejected request.

## Execution and validation

Retain 1s standing transition, 50Hz planner-owned clock, ten future samples at
100ms spacing, 200Hz physics, initial proprioception/history and first-physics
Torch RNG. Require selected reference's initial query and action to match its
fixed pair. Audit every later query against the selected native library and
measured root orientation. Replay every token/action with the frozen codec.
Report complete action/physical-trace equality or differences against the fixed
pair, without changing tolerances or task outcomes to force equality.

The unchanged primary score requires whole-body passage, required duck,15 ticks
of upright recovery then a fresh50-tick stationary hold within the0.5m 3D goal
region, speed<=0.1m/s, no forbidden environment contact>1N at any substep and no
pelvis-height fall, within500 ticks/10s. Foot-floor support is permitted. This
score does not qualify self-contact absence. No imitation termination or reset.

Run sequentially, first clear request as an infrastructure check. Every request,
rejection, failure and timeout remains recorded. No outcome-dependent tuning or
physical retry. Infrastructure failures abort the launch; preserve any failure
before at most one isolated repair retry. Do not interfere with other GPU jobs.
The separate11 offline profile-rejection probes are software checks only.

## Decision and interpretation

Measure complete tasks, selected candidate, geometry-only versus evidence-filtered
choice, and paired gains/regressions relative to selected fixed controls. Report
success-conditioned times, first/peak contact separately, and acquisition cost.
Check the initial public choice with held body state/bank/goal across the maps.

The preflight predicts geometry-only and evidence-filtered selection tie in all
six cases. If confirmed, report a working connection to the stronger simple
baseline, not an admission-mechanism advantage. Do not attribute a comparison
with the older scene motor to a better planner. Exact repeat histories are not
independent new demonstrations. If public execution loses fixed successes,
diagnose the interface before changing the bank or training a model.

After this check, freeze a distinct clearance/entry evaluation before outcomes.
Count false admissions and solvable requests rejected, compare strong fixed and
reference-only geometry controls, and move toward independent sources/scenes.
No additional familiar-seed search, online replacement, perception claim or
hardware belongs to this six-request study.