Fr0zkaandClaude Opus 5.5
bf39712792
Bands = end profiles of per-control-span shell capsules (~9.35 m pitch = span spacing). Shell now smooth-unions adjacent spans over 6 m and reuses the interior chord distance; a 0.60 m clearance fades in 5-20 m from each mouth. Close-ups: bands reduced but still visible on part of the exposed shell. Cost vs L109: near-ready +8..+12% (FI 0.88 -> 1.00, TN 0.92 -> 1.01 s @W12). Audits, Clang/GCC determinism, fast tier and Linux UE build pass. Kept on its own branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
3.7 KiB
Linux 110 — open-air tunnel shell
Candidate evaluator v257, source stamp 04e9180521ab0963da5f1cabbf7c4646e32311edf1439c2dea514d3a61e7e464.
Cause and fix
The bands are the end profiles of per-control-span shell capsules. The affected 452.408 m open-air section touches 50 control spans: chord lengths are 8.520–10.076 m (median 9.353 m, mean 9.324 m). The visible band pitch follows that roughly 9.35 m span spacing. Radius delta is 0 across those spans, and the floor depth is constant, so radius/floor interpolation does not explain it. The 8 m tiles, 0.25 m voxels, and 0.75 m shell do not match the period. The pattern remains with tile skirts off; wall relief is unchanged. The chord/Catmull mismatch is not the periodic source.
The shell now smooth-unions adjacent spans over 6 m and reuses the chord distance already calculated for the passage interior. A 0.60 m shell clearance fades in from 5 m to 20 m from each mouth. Interior carving still takes precedence, and the shell is written only into host air. The repeated bands are removed from the exposed shell close-ups. Regular faceting on tunnel-facing interior walls remains visible in both L109 and v257 station captures.
Captures
The capture index flags every OPEN AIR view. The verification passes all 106 images: per before/after set, 6 mouths, 9 stations, 27 open-air views, and 11 open-air close-ups. All 22 close-ups have a measured 4.000 m standoff. Nine images in each set are marked low contrast, with visible geometry. Paired bytes changed in 24/27 open-air views, 10/11 close-ups, 4/6 mouths, and 1/9 stations.
The contact sheets cover all views. The target 4 m close-ups are before 1, after 1, before 2, after 2, before 3, and after 3.
Readiness
Five interleaved L80 pairs per cell; times are medians in seconds (LOD0–2 near-ready / whole-view-ready). All base/candidate tile counts matched.
| Strate | W8 near | W8 whole | W12 near | W12 whole |
|---|---|---|---|---|
| SW | 0.778 → 0.854 (+9.7%) | 1.767 → 1.829 (+3.5%) | 0.807 → 0.906 (+12.3%) | 1.722 → 1.843 (+7.0%) |
| FI | 0.937 → 1.010 (+7.8%) | 1.811 → 1.903 (+5.1%) | 0.883 → 0.995 (+12.6%) | 1.700 → 1.798 (+5.8%) |
| TN | 0.979 → 1.090 (+11.4%) | 1.806 → 1.932 (+7.0%) | 0.920 → 1.006 (+9.3%) | 1.709 → 1.790 (+4.7%) |
| VS | 0.797 → 0.871 (+9.2%) | 1.699 → 1.765 (+3.9%) | 0.758 → 0.835 (+10.2%) | 1.575 → 1.659 (+5.3%) |
Checks
- Owner four-strata walk audit: all 3 interfaces pass; zero blocked, unsupported, or grade-failure samples. Shell audit: 89,365 floor, 8,423 wall, and 10,527 interior samples; zero violations; 22,755 open-air shell samples.
- Clang/GCC determinism: all 18 W1/W8/W12 hash runs match across owner4 Surface World, owner2 Surface World, and owner2 Floating Islands.
- Fast tier: PASS, 259 checks plus 5 diagnostic-only results; no failed gates.
- Linux VFHostEditor: PASS, 27 actions, 69.0 s,
-NoUBA -MaxParallelActions=4. Unreal was run withoutprlimit. git diff --check: PASS. No run exceeded 30 minutes.
Failed experiments
A clearance ramp delayed to 80–120 m left 6 owner2 wall samples uncovered. A 1.2 m mouth join produced 23 wall violations. Both were rejected; the final 6 m join with the 5–20 m clearance ramp passes owner2 and owner4 audits.