VoxelForge/Standalone/Research/WorldAssemblerM3/Linux31/REPORT.md

Linux31 report

FI column finding

The earlier L29/L30 passage table ran SurfaceWorld twice. On the real FloatingIslands recipe, the old column check under-counted connected ends because it treated support at the route foot as the column cap. The route foot is near z = −116 m; the actual lower cap is near z = −122 m, inside native bedrock whose top is z = −119.5 m. The upper caps sit inside anchor-island rock near z = −62 to −69 m. There is no floating column end or unintended gap. The open passage mouth is intentional.

The audit now samples each planned cap against the native density field. Negative density is solid rock. Every main column cap overlaps native rock within 0.25 m. The style 3 short lower cap overlaps within 0.50 m.

FI style	Cap	Z (m)	Native overlap / samples	Minimum native density
0	Main upper / anchor island	−65.846	5,051 / 10,005	−14.761719
0	Main lower / bedrock	−122.390	6,149 / 12,588	−8.000000
1	Short upper	−61.574	9,725 / 11,033	−23.464844
1	Short lower	−75.637	3,915 / 11,500	−15.554688
1	Main upper / anchor island	−66.485	4,497 / 11,118	−15.613281
1	Main lower / bedrock	−122.416	6,673 / 13,735	−8.000000
2	Main upper / anchor island	−68.891	3,986 / 8,523	−15.546875
2	Main lower / bedrock	−122.382	5,396 / 10,885	−8.000000
3	Short upper	−65.805	5,095 / 10,064	−14.671875
3	Short lower	−80.054	2,498 / 10,413	−11.910156
3	Main upper / anchor island	−69.478	3,902 / 9,985	−15.597656
3	Main lower / bedrock	−122.436	6,130 / 12,453	−8.000000

The negative control moves the style 0 lower cap into the air gap at z = −96.390 m: 0/3,120 native overlaps, minimum density +22.273438, and detected=true. The audit still catches a genuinely disconnected end.

Recipe-stamped passage matrix

Each row is checked against the requested and resolved recipe stamp. The columns are capsule blocked/unsupported counts and mouth checks, open-air violations, naturalness, longest uncolumned span / 12.5 m limit, connected column ends, route bindings / failures, and walkability samples.

For all eight FI rows, FloatingIslandRouteAudit also passes: 78,064 capsule samples, 0 blocked, 0 unsupported, and 0 grade failures. The runner now asserts both capsule audits.

Requested → actual	Style	Compiler	Capsule	Open air	Natural	Long bridge	Column ends	Bindings	Walkability
surface-world → surface-world	0	clang	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	1 (0 failures)	5642/5642
surface-world → surface-world	0	gcc	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	1 (0 failures)	5642/5642
surface-world → surface-world	1	clang	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	1 (0 failures)	1490/1490
surface-world → surface-world	1	gcc	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	1 (0 failures)	1490/1490
surface-world → surface-world	2	clang	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	1 (0 failures)	1545/1545
surface-world → surface-world	2	gcc	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	1 (0 failures)	1545/1545
surface-world → surface-world	3	clang	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	1 (0 failures)	1495/1495
surface-world → surface-world	3	gcc	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	1 (0 failures)	1495/1495
floating-islands → floating-islands	0	clang	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	4.464/12.500 m	2/2	2 (0 failures)	4000/4000
floating-islands → floating-islands	0	gcc	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	4.464/12.500 m	2/2	2 (0 failures)	4000/4000
floating-islands → floating-islands	1	clang	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	3.851/12.500 m	4/4	2 (0 failures)	1047/1047
floating-islands → floating-islands	1	gcc	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	3.851/12.500 m	4/4	2 (0 failures)	1047/1047
floating-islands → floating-islands	2	clang	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	2 (0 failures)	1135/1135
floating-islands → floating-islands	2	gcc	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	0.000/12.500 m	2/2	2 (0 failures)	1135/1135
floating-islands → floating-islands	3	clang	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	4.243/12.500 m	4/4	2 (0 failures)	1061/1061
floating-islands → floating-islands	3	gcc	0/0 blocked/unsupported; mouth 2/2	0 violations	1/1	4.243/12.500 m	4/4	2 (0 failures)	1061/1061

All 16 rows pass. The runner rejects a recipe mismatch. Its negative control requests SurfaceWorld while the resolved layout is FloatingIslands; the log records requested=surface-world actual=floating-islands status=FAIL and the process exits 2. Logs and the machine-readable matrix are under Saved/Linux/l31/.

FI floor lines

The L30 rings persisted with height tint off, so they were geometry. Evaluator v19 reduces FloatingIslands Gradient displacement to transition_width / 256 (one quarter voxel at the 16 m transition width), and the layout validator uses the same bound. The v18 one-voxel cap still left a 0.251819 m radial height range and a visible contour.

On the east radial line from (64, −32), samples at radii 1–16 m all have height −119.500000 m and normal (0, 0, 1). The extended mesh raster has a 0.000479 m total height range. The readable shader now applies a cubic smooth tint (t²(3−2t)). Tint-off and tint-on captures show a flat floor without contour bands.

Captures and hashes

All six JPEGs are recipe-stamped from FloatingIslands and under 300 KB.

Capture	Size	SHA-256
FI column top, close	85,696 B	6b651655fc2127bb5b3502a1a99c024167db2edc21bfe7241f6335a858bb8575
FI column top, wide	67,043 B	926193affd463492095b97310c62416910e12fc94e75d8548a01ad1f9601b6ec
FI column bedrock base, close	60,491 B	3e524e0f6c506413b4b5b29d8835792f8b5240f95787c6a08d12ebe6e46aadd5
FI column bedrock base, wide	61,543 B	27ef1a72874c7f8e257969c5e69d614f336edce84693ee80b98783b9f610a4b2
FI floor, tint off	43,405 B	41b9180c9696c7025e23009e6efdb2e86ace5900b08c293c92371797f9a8596d
FI floor, tint on	43,405 B	0672e2ddc93fb6a4fe4f6adfdf4cdee896bdbc90dddfb7cf65050d978f86f32b

Evaluator version is 19. Binary SHA-256:

• Clang: a796d433b7e31fa05abba970dbd486db1a904d13222d9c02b9054c394375e4f7
• GCC: c2ed12ddf4db5ce97fcd2d9c8f32e4515a4fba6d36d4d8a4b5fbb07a8ab8ae73

Determinism and cost

The L25 command and 24 m crossing each passed 12 cases: Clang and GCC at W1/W8/W12 for both recipes. All 12 hash records per log were stable across compilers and worker counts. SW matches the L30 baseline; FI changes as expected with the v19 geometry. L25 and crossing hashes match for all 12 recipe/compiler/worker combinations.

Recipe	Full-detail density / mesh / payload / topology	Whole-view density / mesh / payload / topology
SurfaceWorld	ace56c8f205238ad / a9c603a1248de6b8 / ea55f1f4fcc6a894 / 5ea7bc78a1b678ae	0e8c5af01df96b14 / 70099e89b77811ed / dbce0fdcf6764126 / ddaf81a1228f9b4b
FloatingIslands	ed550e0c8b182695 / 0e410a8e80a3e7f7 / 0751926f0c67d8a4 / f01208bc2f7841ec	52e0b9f95da4291e / 7c6e9e18346278a4 / 59160fe3df02a0ae / ef7f17088fe3253e

Cost compared the Clang L31 binary with the L30 linux30-speed-fade binary: 10 interleaved pairs per recipe, 40 runs total, with zero cache hits. Combined median whole-view time was 1.772 s → 1.784 s for SW (+0.71%) and 17.468 s → 17.482 s for FI (+0.08%). The second five-pair batch was +1.20% SW and +0.01% FI; the larger early SW deltas settled as runs warmed. The paired medians across all 10 were +1.86% SW and +0.55% FI, within the observed run variation.

Failed and retried runs

• Five low-clip floor exports (export-fi-floor.log, export-fi-floor-probe.log, export-fi-floor-without-mesh.log, export-fi-floor-probe-retry.log, fi-floor-no-camera-probe.log) stopped with std::bad_alloc. Export succeeded with --max-clip-level=8 --max-level=0.
• A first radial analyzer invocation omitted the second --range value and was rejected by argparse; it was rerun with both bounds.
• A proof/parity probe used custom camera positions and failed SW parity because the sampled tile set differed from the established default-camera audit. That probe was stopped; the default-camera proof and transition suite passed 16/16.
• The first determinism launch ended before completing the matrix. The full 24-case run passed on retry. A first cost batch had an early SW timing outlier, so five additional interleaved pairs per recipe were run and retained in the combined summary.
• Early capture framing probes cropped the FI axis or placed the camera inside rock. A v19 wide floor view exposed only the export skirt edge; the final centered column and square top-down floor views were used.
• The first GCC build attempt was interrupted at 19 objects and retried successfully. One v18 diagnostic build used the default Standalone build path; those artifacts were removed, and final v19 build outputs are under Saved/Linux/l31/.
• Two one-line post-processing commands had quoting/pair-index errors; the checked-in CompareDeterminismL31.py and AggregateCostL31.py helpers produced the final passing summaries.

No run exceeded 30 minutes or remained stuck for 15 minutes. No Unreal files were edited. Nothing was committed, pushed, or stashed.

Orchestrator review (Claude)

• Merged as an intermediate step: correct FI column-cap audit (with negative control), a recipe-stamped runner that rejects mismatches, all 16 recipe/style/compiler rows passing, smooth shader tint, determinism.
• Not accepted as the final floor fix: v19 caps the FI Gradient displacement at width/256 (≈6 cm), which removes the owner's Gradient transition rather than making it smooth; the floor is now flat (radial height range 0.0005 m). The real fix is a smooth, non-terraced undulation at the L29 amplitude (≈4 m peak-to-peak). L32 restores it.