Fr0zkaandClaude Opus 5.5
ba017077d1
The shared passage carve took the minimum of piecewise-linear capsules (scallops at joins) and added wall relief on regular 8 m / 4 m octaves, which read as evenly spaced ribs. Interior sweep joins are now Catmull-Rom smoothed and the relief uses irregular rotated 43/29/19 m scales; the natural wobble remains. New axial ripple audit (0.125 m radius samples, 0.5-8 m periods): beaded control and L34 geometry fail, all 24 passage rows pass. FI mouth tread apron tapered. 30/30 determinism, 16/16 invariants, ring audits LOD0-8 pass; cost within noise. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CSFYgsdW8ejZSyiMf4qRSX
125 lines
8.3 KiB
Markdown
125 lines
8.3 KiB
Markdown
# VoxelForge L35 report
|
||
|
||
## Rib diagnosis and fix
|
||
|
||
The ribs are in the geometry, not just the lighting. L34 owner-SW style 0 measured a 7.438 m
|
||
radius period at 0.0361 m (3.61 cm). Its LOD0 mesh also measured a 4.750 m normal period with
|
||
0.02682 axial amplitude (1.537°), above the 0.0255 limit. The large period tracks the old 32-voxel
|
||
noise octave (8 m at 25 cm per voxel); the 4.75 m normal period is near its 16-voxel octave (4 m).
|
||
The 8 m layout chunk is the same length by coincidence: the relief is sampled directly in world
|
||
coordinates. Neither the roughly 2 m route samples, 25 cm voxel size, nor 2 m portal fade matches
|
||
the measured periods.
|
||
|
||
The shared passage carve used the minimum of piecewise-linear capsule segments, which added local
|
||
scallops at joins. The regular wall-relief octaves added the longer beat. L35 Catmull–Rom smooths
|
||
eligible interior sweep joins, excluding short spans and the route ends. It replaces the old relief
|
||
with irregular rotated 43/29/19 m scales and reduces the terrain-band contribution. The natural
|
||
wall wobble remains. The evaluator version is 21.
|
||
|
||
On owner-SW style 0, L35 measures a 7.062 m period at 0.0109 m (1.09 cm); coherent amplitude is
|
||
0.0044 m (0.44 cm). Its mesh-normal scan passes at 8 m with 0.02496 axial amplitude (1.430°).
|
||
The periodic residual stays below both the 0.035 m ripple limit and the 0.0255 normal limit.
|
||
|
||
The FI slab is the shared passage support tread under the bore. It is not `LandingExtensionQ`, a
|
||
tile-boundary seam, or a separate support layer. The flat projecting apron in L34 was too broad; the
|
||
support tread is required for capsule and probe walkability. L35 tapers and grains its entry shoulder
|
||
and keeps a shallow supported walk core. The capture now shows a narrow, slightly flatter tread at
|
||
the mouth joined to the crater by irregular ground. The support and capsule audits pass.
|
||
|
||
## Ripple and world audits
|
||
|
||
The axial audit samples side-wall radii every 0.125 m across 32 m windows and scans periods from
|
||
0.5 to 8 m. The synthetic control is 2 m radius spheres spaced 1.5 m apart; it measures 0.0599 m
|
||
and is detected above the 0.035 m threshold. The L34 real geometry measures 0.0361 m and fails.
|
||
After the fix, all 24 recipe-stamped passage rows pass: four styles for both owner recipes and the
|
||
upper-SW fixture, with Clang and GCC. Walkability, support, open-air, route bindings, mouth density
|
||
profiles, and strict layout checks pass in those rows.
|
||
|
||
The branch-locked FI floor and SW ceiling ring audits pass at LOD0 through LOD8. Coverage is 1.000,
|
||
each profile has at least 3.16 m relief, and no flat run or following step was found. The largest
|
||
normal change is 5.34°. The L30 ring control still fails on its 1.875 m flat followed by a 0.241 m
|
||
step; the L31 flat control fails with only 0.000046 m height range. These are intended negative
|
||
controls. Ring logs and per-LOD windows are under `Saved/Linux/l35/ring-*`.
|
||
|
||
The recipe-stamped interface/hole and transition suite passes 16/16; hole checks report no hole
|
||
columns or air samples, and strict validator controls reject 10/10 malformed layouts. L13 and L9
|
||
pass 4/4. The separate mesh-mouth profile supplement reports 3 pass, 6 no-coverage, 0 fail; the
|
||
density mouth profiles are covered by the 24/24 passage suite.
|
||
|
||
## Captures
|
||
|
||
The 12 standard 1600×900 JPEGs are in [`captures/`](captures/). They use the L34 fixed cameras,
|
||
eye height and framing, and are all below 44 KB (limit 300 KB): four views each for `owner-fi`,
|
||
`owner-sw`, and `upper-sw-fixture` (`ground-mouth`, `rock-entry`, `mid-tunnel`, `arrival`).
|
||
|
||
## Determinism and hashes
|
||
|
||
The L25 command and 24 m crossing pass at W1/W8/W12 with Clang and GCC for both recipes. The
|
||
upper-SW fixture passes at the same worker counts with both compilers. Result: 30/30 stable runs.
|
||
Clang and GCC W12 hashes match. Both recipes report recipe hash
|
||
`a2985985e6985a2a`, plan hash `a5e29887e839c5d6`, and manifest hash `6c89e165364e56df`.
|
||
|
||
| Recipe | L25 full detail: density / mesh / payload / topology | L25 whole view: density / mesh / payload / topology |
|
||
|---|---|---|
|
||
| SurfaceWorld | `ace56c8f205238ad / a9c603a1248de6b8 / ea55f1f4fcc6a894 / 5ea7bc78a1b678ae` | `b40597aa93148a45 / 8a42763cb936756d / 06c9171fdaa7875c / 9057043383ba37b1` |
|
||
| FloatingIslands | `ed550e0c8b182695 / 0e410a8e80a3e7f7 / 0751926f0c67d8a4 / f01208bc2f7841ec` | `7dd12412f1c036c6 / d074c75df2f14e7f / d4a2cf28c2873580 / 903ecae9e66ee20a` |
|
||
|
||
Exact-query hashes (density / detail / collision) are SurfaceWorld
|
||
`c949c4df2ce6f9d9 / 114c501e2d830d94 / e6011d3a248d4849` and FloatingIslands
|
||
`7b7c50ac24ccadcc / efcba20c0c70c415 / e604823a249029bf`. The upper fixture layout identity is
|
||
`resolved-layout-v3-fnv64:fc4e73619037f88c`.
|
||
|
||
Core binary SHA256: Clang `6a98169293a955f0119f1ad13b9ab949a9729948401c7f832446a6d980c9c6ae`;
|
||
GCC `8535726460807e198a0f05c7b1197ab4ca51c3fda18a87d45a1e49e754589542`.
|
||
|
||
## Cost, edits, and run conditions
|
||
|
||
Five interleaved L34/L35 pairs per recipe pass. Median whole-view time is SurfaceWorld 1.776 s →
|
||
1.807 s (+1.75%) and FloatingIslands 17.092 s → 17.251 s (+0.93%); paired changes are mixed and
|
||
within run noise. No Unreal files changed. Core builds and all runs used CPUs 0–11, a 6 GiB address
|
||
space limit and at most 12 workers. No run exceeded 30 minutes, no step was stopped, and the governor
|
||
was unchanged.
|
||
|
||
Changed files: `Standalone/Core/include/VoxelForgeCore/VoxelForgeCore.h`,
|
||
`Standalone/Core/src/ResolvedLayoutInternal.h`, `Standalone/Core/src/ShapeWorldInternal.h`,
|
||
`Standalone/Core/src/VoxelForgeFieldShared.h`, `Standalone/Source/StreamBenchmarkPipeline.h`,
|
||
`ARCHITECTURE.md`, and `CODEMAP.md`. No Unreal edits were made or compiled.
|
||
|
||
## Failed and retried runs
|
||
|
||
- The first ripple audit used a 1.25 m sphere spacing and a 0.045 m limit; its 0.0418 m control
|
||
was missed. The synthetic control is now 1.5 m spacing with 2 m spheres, and the limit is 0.035 m.
|
||
- Route resampling first failed support/open-air checks (12/24). Four-way resampling and raw routes
|
||
with the old relief failed mesh-normal scans at 5.688 m/0.03064 and 4.625 m/0.03263; old 13/9/5
|
||
voxel relief also failed at 4.688 m/0.03546. SDF smoothing without terminal exclusion blocked 143
|
||
samples; high route noise left 720/976 unsupported. Those variants were reverted.
|
||
- Removing support caused 30 unsupported samples; lowering its top caused 6; an early widened/noisy
|
||
form caused 1. Full and half support splines and full vertical smoothing each caused an unsupported
|
||
sample. Quarter-weight vertical/radius smoothing passed. The all-route 0.14 m grain trial caused
|
||
132 unsupported FI samples and was reverted; mouth-only shoulder grain remains.
|
||
- Exact Cascade cosine lobes failed FI style 3 in both compilers (9.18 cm at 6.188 m). Full-depth
|
||
seeded profiles failed at 5.938 m/0.0537 m and 3.312 m/0.0792 m. The retained shallow seeded
|
||
profile passes the full matrix.
|
||
- Initial ring probes used the wrong data, camera, or 16 m radius; they returned no coverage or
|
||
only 1.857 m relief. The canonical owner input, 32 m LOD0 windows, and covered branch-locked
|
||
LOD1 windows pass. The first SW LOD1 set had three uncovered windows; retrying the seven covered
|
||
windows passed 77/77 profiles.
|
||
- A supplemental fade-ring scan across the rough owner-FI crater still exceeds the 20° normal
|
||
limit (42.27°); the same camera/window on L34 also measures 42.27°, so this is present in the
|
||
comparison build. An old fixed center has no coverage on the current layout. These probes are
|
||
recorded separately from the passing branch-locked FI/SW LOD0–8 ring suite.
|
||
|
||
No commits, pushes, or stashes were made.
|
||
|
||
## Orchestrator review (Claude)
|
||
|
||
- Accepted: the wall ribs are gone in the captures (owner-SW mid-tunnel and upper-SW arrival walls now read smooth with
|
||
irregular wobble). Cause measured (capsule-join scallops plus regular 8 m/4 m relief octaves), fixed in the shared
|
||
carve, beaded control fails, L34 real geometry fails. Evaluator v21.
|
||
- Still open, for L36:
|
||
- **The floor still ripples.** In owner-SW mid-tunnel, the walk floor shows a regular washboard/step pattern. The
|
||
ripple audit samples only side-wall radii, so it can't see the floor (the support tread).
|
||
- **The FI mouth is not natural yet.** The slab is gone, but it was replaced by a jagged cluster of broken, faceted
|
||
pieces in the crater (owner-FI ground mouth). The 42° fade-ring normal jump that is also present in L34 belongs
|
||
to the same area.
|