Files
VoxelForge/Standalone/Research/WorldAssemblerM3/Linux26/REPORT.md
T
Fr0zkaandClaude Opus 5.5 37a223046c wip(L26, side branch): every passage style walkable in both recipes (0/0/0); SW bridges look wrong
All 4 styles x both recipes pass capsule, open-air and naturalness audits, and full-view hashes are
deterministic. But the SW passage crosses 355 m of open SurfaceWorld sky as a ~2 km fragmented rock
ramp, which is not natural. L27 replaces long bridges with a shared ceiling-to-ground rock column.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CSFYgsdW8ejZSyiMf4qRSX
2026-10-21 02:39:44 +02:00

145 lines
10 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# World Assembler L26 — passages walkable in rock
**All four styles pass for both recipes.** Clang and GCC report zero blocked capsule samples, zero
unsupported samples, zero grade failures, zero open-air violations, passing naturalness, and passing
route bindings. The L24 open-air control and L23 smooth-wall control were detected.
## Where support was missing
In the owner SurfaceWorld recipe, the default Spiral passage has 1,083 control points. Its unsupported
bridge range is segments 98–1077, a 1,957.395 m gap. Segment 98 is at about `(0,-5,-142.7) m`; segment
1077 is at `(-4.6,-1.9,-497.4) m`. Both are inside the SurfaceWorld stratum (`-632..-120 m`), below
its native terrain. The gap starts after the TunnelNetwork transition; it is neither TunnelNetwork
cave air nor a stratum seam. The first L25 failure sample had only 2/5 native support probes, and
following samples had 0/5. Their dispatch path was `SurfaceWorld-native`.
Core now scans the native field sign and adds a deterministic bridge only over contiguous unsupported
ranges. The shared density field gives the bridge at least 5 m width, 3–5 m planned thickness and
surface relief. The audit measured two connected ends, at least 5 m width, 256/256 cross-sections,
3.5–6 m thickness and 2.5 m relief on the default Spiral. All four SurfaceWorld styles use connected
bridges; Floating Islands already has supporting rock and needs none. The capsule clearance gate also
checks the resolved player capsule, fixing FI styles 0/1/3 at their bent mouth segments.
The first bridge export still contained no passage mesh because tile readiness classified all nearby
tiles as native-only. Resolved passage and bridge bounds now keep those tiles shape-dependent. The final
local export changed `shape_free_desired` from 27/27 to 4/27 and contains the composed bridge density.
## Eight-row audit
Counts are from the Clang W12 run; GCC W12 matched every row. `B/U/G` means blocked, unsupported and
grade failures. Route bindings are required bindings passed / required bindings.
| Recipe | Style | Capsule samples | Support samples | B/U/G | Open-air samples / violations | Wall relief | Bindings | Bridge segments / length |
|---|---:|---:|---:|---:|---:|---:|---:|---|
| SurfaceWorld | 0 | 301,140 | 32,765 | 0/0/0 | 3,887,808 / 0 | 0.472 m | 1/1 | 30–293 / 2,004.318 m |
| SurfaceWorld | 1 | 284,949 | 31,046 | 0/0/0 | 3,707,328 / 0 | 0.449 m | 1/1 | 33–357 / 1,945.857 m |
| SurfaceWorld | 2 | 275,877 | 30,072 | 0/0/0 | 3,664,064 / 0 | 0.410 m | 1/1 | 98–1077 / 1,957.395 m |
| SurfaceWorld | 3 | 287,154 | 31,298 | 0/0/0 | 3,741,792 / 0 | 0.552 m | 1/1 | 34–373 / 1,980.607 m |
| Floating Islands | 0 | 50,589 | 4,407 | 0/0/0 | 695,616 / 0 | 0.779 m | 2/2 | none |
| Floating Islands | 1 | 57,015 | 5,060 | 0/0/0 | 742,016 / 0 | 0.806 m | 2/2 | none |
| Floating Islands | 2 | 61,173 | 5,318 | 0/0/0 | 813,632 / 0 | 0.767 m | 2/2 | none |
| Floating Islands | 3 | 54,369 | 4,325 | 0/0/0 | 743,456 / 0 | 0.681 m | 2/2 | none |
Every row passed naturalness (0 m flat run; relief threshold 0.2 m), route binding and strict layout
validation. The validator accepted with zero issues and rejected its malformed controls 9/9. Each L24
open-air negative control and L23 smooth-wall negative control was detected.
## Captures
All are 16:9 JPEGs below 300 KB in `captures/`:
| Capture | View | Size | SHA-256 |
|---|---|---:|---|
| `sw-passage-outside.jpg` | Midspan from outside, 1920×1080 | 90,937 B | `28378eeee755faf7446c0b2ce0eb8d2368c79b42f7de4c15e8b2f4d58fb7754d` |
| `sw-passage-inside.jpg` | Eye-height at first bridged segment, 1600×900 | 47,519 B | `ccadc33c36366a0ca7b2fcda0a21e4cea67bda107f0e20015171fd620d495249` |
| `sw-natural-bridge-wide.jpg` | Wide midspan view, 1920×1080 | 70,192 B | `f4a66fec1adebfbdf5806bb2e7c5a6470e3e8d9dcd60f495d930bbb1d0c5669e` |
| `fi-pillar-exterior.jpg` | Floating Islands pillar exterior, 1600×900 | 35,857 B | `b977ff34acee1294a24c2465b1c9d3d78014cff4b8eceb667f317a0a0b46cc67` |
## Determinism and regressions
The L25 hash command was rerun with the L26 binaries for both recipes, W1/W8/W12, Clang/GCC:
```sh
timeout 1800s taskset -c 0-11 env LP_NUM_THREADS=12 prlimit --as=6442450944 -- \
Saved/Linux/bin/VoxelForgeCoreHeadless-linux26-<clang|gcc> \
--world-assembler-prototype --stream-benchmark --stream-pipeline --stream-hash --force-generate \
--clip-radius=3 --max-clip-level=8 --far-horizon-lod8-step=0 --workers=N \
--recipe=<surface-world|floating-islands> Standalone/data/owner_world.json
```
All six runs per recipe matched. Hash vectors are density / mesh / payload / topology; full detail is
960 tiles and whole view is 2,370 tiles.
| Recipe | Full detail | Whole view | Exact query: density / unclamped / detail / collision |
|---|---|---|---|
| SurfaceWorld | `ace56c8f205238ad / a9c603a1248de6b8 / ea55f1f4fcc6a894 / 5ea7bc78a1b678ae` | `307d5df80e0165b4 / 1d61f30e5733beca / 3aaef136176c4b12 / aa5ff7c2a821196f` | `c949c4df2ce6f9d9 / c949c4df2ce6f9d9 / 114c501e2d830d94 / e6011d3a248d4849` |
| Floating Islands | `d18e722b37a2479e / 0e410a8e80a3e7f7 / 0751926f0c67d8a4 / f01208bc2f7841ec` | `51decfb71d282506 / 69448ec1c7fc954f / e2ff37c98fe8e494 / 5beeb7f94fdbf3d4` | `7b7c50ac24ccadcc / 7b7c50ac24ccadcc / efcba20c0c70c415 / e604823a249029bf` |
The 24 m crossing passed for both compilers and recipes. Each processed 314 moving tiles, with identical
compiler hashes. Moving-tile vectors (density / mesh / payload / topology): SurfaceWorld
`3561ea23485e359f / e8331620fb0fb908 / 765f7dad67c852aa / 3ccc09ac81c05d71`; Floating Islands
`e3d0ea833bca4215 / 6b3d0f148ed83d93 / d6ed85f16acd895e / 9745683036522461`.
SurfaceWorld away from the passage was compared directly between L25 Clang and final L26 Clang at
`--camera=10000,10000,-100,0,0`. Full detail, whole view, all LOD0–8 rows and exact-query hashes were
identical (12/12). Full-detail vector: `1c6abc5e9c657e98 / a1624a3c19bbe554 / bf98934c9d846f3a /
bbdc2a6b821c66d6`; whole-view vector: `40c2be0c7bbdaf28 / 7d9ac2a85020dc7d / c378e48bab2d194f /
81014ec4b888107a`; exact-query vector: `482500abbe835b11 / c4948c7179963784 / efcba20c0c70c415 /
137e8f6c0b03a624`.
Strict validation passed on both compilers and recipes. L13 route parity passed on both compilers/recipes:
8/8 page pairs, 14 higher-LOD pairs, one edit pair, 782,026 exact-query samples and zero mismatches.
For L9 proofs, the passage-centered view has no eligible native blocks. At the same 10 km away-from-route
view, SurfaceWorld checked 4,848 blocks / 6,452,688 samples with 1,296 fallbacks and zero violations
on both compilers. Floating Islands had its expected 21,504 unknown fallbacks and zero violations.
Final headless binary SHA-256: Clang `fc0a5a74f17566384e094cb2429d8925147dd3eb32fb925417393e00c9bb19d5`; GCC
`54843bef3a64087e55d3eb9a76fb0856162b6ad9aeb03d6e1c204c3e91d21ca5`.
Clang built with FP contraction off; GCC used its default FP mode. Both built and started without the L3
FP-contract self-test rejecting them. Core evaluator version is 13; FI geometry remains 7 and layout
schema V3 is unchanged. Changes are shared Core, standalone audits, Linux build output configuration,
and ARCHITECTURE §8.44 / CODEMAP §3.27. No Unreal files changed, and no async path was added. The L4
`FLazy`, L7 cache, L9/L12 proofs, L11 readiness/crash handling, L13 binding, L18/L19 ceiling,
L21/L22 layout and bands, L23–L25 passage/helix, §8.10 and CLAUDE.md async rules remain intact.
## Failed and retried runs
- The first build command used an invalid `clang` variant; corrected to Clang/GCC variants. Early build
iterations R3–R6 had compiler errors and were corrected before final binaries were produced.
- An unsupported `--world-assembler-audit` option exited 2; the correct option is
`--audit-resolved-layout-passages`. One early invocation omitted that flag and performed no passage
audit. Early bridge probes also segfaulted while crossing stratum dispatch; final audits use bounded
resolved-layout queries.
- Support iterations R22–R30 exposed successive failures: FI style 0 had 767, then 27, then 5 blocked /
33 unsupported samples; an offset experiment caused 746 unsupported samples; a control-point capsule
reference left 7 blocked samples. Correcting the capsule reference cleared those. R30 still had SW
bridge sections 254/255, four SW style-1 mouth blocks, and an invalid SW style-2 bridge thickness
sample. R31 corrected them; final matrix is clean.
- The first final audit batch labeled FI JSON files as Floating Islands but omitted `--recipe`, so those
invocations used SurfaceWorld. All 16 style audits were rerun with explicit recipe flags. An initial
proof run used `--force-generate`, which disables L9 block auditing; it was repeated without that flag.
The validator-only command also exited 2 until the required streaming flags were included; all four
corrected strict-validator runs passed.
- `FloorProfileBlend` still reports its L25 alternate-style failures: SurfaceWorld style 1 is 2.544 vox
over the 2.000 threshold and style 3 is 3.645 vox. Both values are identical in the L25 logs; they are
pre-existing and are not capsule, support, open-air or naturalness failures.
- Initial capture attempts used a wrong yaw and missed/partially missed the FI export volume. After
correcting the view and centering the export, the final captures above were rendered. Several inside
camera placements showed wall-only or void-only views; the selected interior frame is at the first
bridged segment.
No run exceeded 30 minutes or remained stuck for 15 minutes. Exports and logs are under `Saved/Linux/`;
this round added only this report and the four small JPEGs under `Standalone/`. An unrelated pre-existing
untracked `Linux18/exports/` directory was left untouched. No Unreal edits, commit, push or stash were made.
## Orchestrator review (Claude) — side branch, not merged
- The audits pass (8/8 rows at 0/0/0), but the SW "natural bridge" is a ~2 km rock ramp spiralling through the
open SurfaceWorld sky from the ceiling (≈−143 m) to the ground (≈−497 m). `sw-natural-bridge-wide.jpg` and
`sw-passage-outside.jpg` show fragmented floating slabs with thin shard spikes: not natural, and visible
from far away.
- Decision (delegated by the owner): one shared mechanism for every strate. When a passage must cross a large
open void inside a strate, it descends inside a natural rock column (ceiling-to-ground, like a
stalactite/stalagmite pair joined), with the spiral tunnel inside, generalising the L25 FI anchor root.
No long bridges through open air. Implemented in L27 on `wip/l25-helix`.