# VoxelForge L36 report

## Findings and fixes

The exported L35 walk meshes show real floor washboard: owner-SW repeats at 7.250 m with 2.252 cm amplitude; the upper-SW arrival repeats at 1.438 m with 2.482 cm amplitude. The old tread top was built per segment as a one-sided density clamp, so the tread/native-field crossing moved at the segment joins. The wall-only ripple scan could not see it. Removing shoulder grain alone did not remove the beat.

Sparse-route treads now overlap by up to 2 m and use bounded signed-distance slabs, including SurfaceWorld-to-TunnelNetwork. Tread seating and capsule/open-air gates remain. The exported v23 meshes measure 7.188 m / 0.067 cm on owner-SW and 6.500 m / 0.057 cm on the upper fixture. The expanded five-window density audit's maximum is 1.08 cm, below its 1.5 cm limit.

L35's FI mouth mesh had a 42.265° fade-ring normal jump and a detached 8-triangle component of 0.0035394 m². The defect was the short crater-root detail and tread support meeting the native crater field without a smooth signed-distance join; Marching Cubes exposed thin pieces at the seam. V23 filters the local signed field over 2 m while protecting the route, removes thin isolated solids, and smooth-unions the mouth tread with native terrain over 1.5 m. The same mesh check now measures 27.133° against a 35° limit and finds no component at or above 0.0001 m². One 8-triangle speck remains below that floor (0.0000266 m², 9.377 m from the mouth); it is recorded rather than counted as zero.

## Audits and captures

- Passage matrix: 24/24 recipe-stamped passes across both owner recipes, four styles, the upper-SW fixture, Clang and GCC. This includes support, walkability, capsule clearance, open air, route bindings, strict validation, mouth profiles, wall ripple and floor ripple. The 2 m / 1.5 m beaded control is rejected; both L35 floor windows fail the 1.5 cm floor limit.
- FI-floor and SW-ceiling branch-locked terrain ring audits pass at LOD0–8. Coverage is 1.000; relief is at least 3.162 m and no flat run or following step is found. L30/L31 ring controls remain negative controls.
- Interface/hole and transition suite: 16/16; zero hole columns and air samples. Strict-layout controls reject 10/10 malformed layouts. L13/L9: 4/4.
- FI mouth fragment/crease control: L35 fails at 42.265° with one qualifying fragment; v23 passes at 27.133° with no qualifying fragments. The L35 mesh is the negative control.
- The standard 12 JPEGs use the L34 fixed cameras and are in [`captures/`](captures/). Sizes are 21,976–43,435 bytes (all below 300 KB).

## Version, hashes, determinism

World evaluator version is 23.

| Recipe | Stable layout | L25 whole-view density / mesh / payload / topology |
|---|---|---|
| SurfaceWorld | `resolved-layout-v3-fnv64:f4c09b442fa40cd7` | `90ca5c612c0723bb / cac3b5f0f68c75f9 / 0a143e28341ed132 / 89f2018345744589` |
| FloatingIslands | `resolved-layout-v3-fnv64:d2922ac7e8a2b491` | `da5eba38f4b14f79 / 8c403d9749b15abb / 8a917fa5b71ab179 / 3ed411c118d1f43f` |

Owner plan/manifest hashes: `a5e29887e839c5d6 / 57ebed8456c20f06`. Upper-fixture layout/plan/manifest: `resolved-layout-v3-fnv64:fc4e73619037f88c / 463a0f3b72909656 / f7bc6980e67b61c0`.

Determinism is 30/30: L25 and the 24 m crossing at W1/W8/W12 with both compilers and recipes, plus the upper fixture at W1/W8/W12 with both compilers. Core binary SHA-256:

- Clang: `f8342a7bb48bf460550bd2d4707a67156ebf02a5e5a247edd07e8de0028f6be2`
- GCC: `37d28f3c8a816f485d2bb9e40c1cc37850e6f2d3888772573d11d7b101f6b7d5`

Ten interleaved L35/L36 pairs per recipe passed. Median whole-view time: SurfaceWorld 1.728 s → 1.752 s (+1.42%; paired median +0.66%, MAD 1.16%); FloatingIslands 16.645 s → 16.630 s (-0.09%; paired median +0.35%, MAD 0.29%). Both changes are within measured run noise.

## Failed and retried runs

- Three compile fixes: `build-l36-clang-retry2.log` had a wall-relief call mismatch (retry3 compiled); `build-l36-floor-coordinates.log` had an invalid assignment (retry1 compiled); `build-l36-sweep-closest.log` assigned to a const value (retry1 compiled).
- Passage audit revisions: `run-passage-audits.log` ended after four cases; `run-passage-audits-final.log` was 14/24, `-r2` 18/24, `-r3` 22/24, and `-r4` stopped after eight cases on a failing broad tread candidate. `-r5`, `-r6`, and the final v23 run were 24/24. Earlier support candidates failed capsule, open-air, or floor-ripple checks; their `preaudit-*`, `preflight-*`, and `audit-*` logs remain under `Saved/Linux/l36/`.
- Mouth trials `mouth-smin1m`, `mouth-route-smoothunion`, `mouth-airward-filter`, and the 1.5 m filter variants left 36–44° jumps or a residual speck. The final 2 m field filter plus 1.5 m SDF union is the passing candidate. The older supplemental 20° fade probe remains stricter than this round's 35° gate; the final 27.133° jump would still exceed 20°. The L35 42.265° control is detected by the final check.
- Initial floor-window placement at fraction .49 missed the upper-fixture route target; .457 covered it. Initial ring exports (`ring-export-*-v23.log`) used a render-distance override; early `ring-*-v23-camera*.log`, `ring-*-v23-branch32.log` and `ring-*-v23-current32.log` scans used the wrong center/radius and had missing coverage or only 1.857 m relief. Re-exporting without that override and using the recorded branch-locked windows produced the LOD0–8 passes. `ring-sw-lod0-v23-branchlocked-full.log` had a path typo and was immediately rerun as `ring-sw-lod0-v23-branchlocked-final.log`.
- The first five cost pairs were repeated because the SurfaceWorld median was close to noise; all ten pairs per recipe passed. No run exceeded 30 minutes and no step was stopped for a 15-minute stall. All runs used CPUs 0–11, a 6 GiB address-space cap and at most 12 workers; governor unchanged.

No Unreal files changed. No commit, push or stash was made.

## Orchestrator review (Claude)

- Accepted: the floor washboard is gone in owner-SW mid-tunnel (floor now reads as irregular natural ground). Cause
  measured (per-segment one-sided tread clamps moving the crossing at joins), floor ripple audit added with L35 as a
  failing control. The FI mouth shards and detached fragment are gone. Evaluator v23.
- Not accepted as done: **the FI mouth tread now reads as a trapezoid platform**: straight edges, planar faces, a
  clear man-made shape inside the crater. Also, the 35° crease gate was chosen during the round, while the earlier
  probe used 20°; the final 27° would fail that. L37 targets the shape itself.