# Current KimoNav continuation

Latest completed experiment: **distinct clearance and shadow admission, 20 September 2026**.

- [Current plan](2026-09-20-clearance-admission/CURRENT_PLAN.md)
- [Durable state](2026-09-20-clearance-admission/STATE.json)
- [Report and every outcome](../artifacts/clearance-admission-20260920/REPORT.md)
- [Website report](https://linjiw.github.io/kimonav/clearance/) and [plan](https://linjiw.github.io/kimonav/plan/)

Twelve new fixed executions cover three unmeasured beam offsets (−10/−30/−50 mm),
two execution seeds and local duck versus IK lowered. Each reference succeeds
5/6, no paired task gains or regressions. Both fail at −50 mm with seed 96171
and succeed there with seed 96173. All six pairs match initial state, recorded
dynamics, gains, libraries, initial history and first-physics RNG. All 4,603
issued tokens/actions reproduce exactly. Forty-five focused tests pass.

Geometry shadow decisions admit only the two −10-mm requests: task coverage
2/6, three rejected requests with a measured successful alternative. The finite
ledger rejects all six, including five measured-solvable requests. These shadow
scores are not new public rollouts or physical stopping. The existing public
ledger remains unchanged on its three historical contexts. Do not import these
fixed outcomes into public qualification without a separately frozen study.

Seed access: the registered case-sensitive search was supplemented after
registration. Earlier student-run IDs 96171/96173 are optimizer seeds; all 16
matching native commands actually use execution seed 96161. No prior exact
execution use was found in inspected commands. Preserve the role clarification;
no globally untouched data or independent-scene claim follows.

Next bounded work: specify a simple calibrated geometry-margin competitor and
an execution-envelope baseline with declared calibration/uncertainty inputs.
The new outcomes are diagnostic/calibration data for any rule designed now;
freeze separate evaluation. Never use future outcomes, seed labels or privileged
randomized dynamics as runtime inputs.

Then remove the literal 00976_bank_ parsing assumption through explicit stable
bank metadata, verify unchanged old-bank behavior and qualify an additional
motion source with the current motor/standing protocol. Candidate 00265 exists
but has 32 historical evaluation rows under different conditions; it is not
untouched or already qualified here. Do not continue only varying this beam.
Combine geometry constraints and supported entry variations after capability
qualification, with fixed controls and a qualified detour when available.

No training, hardware or moving replacement in this increment. Preserve repaired
84 / reserved 20 / excluded 16 separation and stable motion–text identities.
Adjacent research imports use frozen snapshots; do not edit that repository or
interrupt its independent GPU jobs.