Files
VoxelForge/Standalone/Research/WorldAssemblerM3/Linux147/REPORT.md
T
Fr0zkaandClaude Opus 5.5 a0d7ab3a1a audit: interior-air gate and gap-true ripple fitter on the cave planner (L147, audit only)
Tunnel Network and Vertical Shafts 16-seed sweeps: 5,171,796 passage-interior capsule samples each, 0 solid;
no-passage controls fire 48/48; masked-gap ripple controls 48/48. Only seed 2 fails (known interface 2-3 shell
wall_missing, outside the interior). Owner hashes/plan/layout identical to HEAD.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
2026-11-01 13:29:58 +01:00

4.9 KiB
Raw Blame History

Linux147 — passage audits on the Core cave planner

Result: The L145 ripple-distance audit and L146 passage-interior air gate pass across all 16 seeds in both requested matrices. The two matrices each measured 5,171,796 capsule samples inside passage interiors with zero solid samples. Their no-passage controls detected solid in 48/48 passage interiors, and their masked-gap ripple controls passed 48/48 times. No generation code changed.

The only non-green result is seed 2's known interface 2–3 shell/seam failure. Vertical Shafts remains 15/16, matching L139. The Tunnel Network comparison is 15/16 against L135's earlier 16/16; seed 2 is already known to fail after the L139 shaft change. It is not a failure introduced by either new audit gate.

Audit changes

  • StreamBenchmarkPipeline.h now requires zero solid capsule samples inside each passage interior. It runs the same interior classifier over the passage-stripped field and requires positive solid detections for every passage.
  • The ripple fitter keeps each valid wall sample's actual route distance instead of compressing distances across missing probes. Its synthetic 5 m masked-gap control measures a 0.0581 m sine ripple and rejects a smooth cubic at 0.0000 m.
  • No Core planner, evaluator, Unreal Source/, or generation code changed. git diff --check passed.

Per-seed verdicts

Tunnel Network compares with the L135 16/16 composed sweep; Vertical Shafts compares with the L139 15/16 sweep. P means pass and F means fail.

Seed Tunnel Network baseline → L147 Vertical Shafts baseline → L147 Verdict change
0 P → P P → P None
1 P → P P → P None
2 P → F F → F Tunnel comparison: L135 predates the L139 shaft change. Current seed 2 has the existing interface 2–3 shell/seam failure; the new interior and ripple gates pass. Vertical Shafts remains the same known failure.
3 P → P P → P None
4 P → P P → P None
5 P → P P → P None
6 P → P P → P None
7 P → P P → P None
8 P → P P → P None
9 P → P P → P None
10 P → P P → P None
11 P → P P → P None
12 P → P P → P None
13 P → P P → P None
14 P → P P → P None
15 P → P P → P None

There were no newly passing seeds and no verdict changes caused by the added gates. For seed 2, all three passages reported interior_air=PASS; its interface 2–3 shell had eight wall_missing samples at route stations 489.627–490.321 m. The samples were outside the passage interior. The first was at (-21.678, -1.714, -888.614) m, with composed and native density +2.341949, expected wall classification, and wall-shell margin −0.297164 m. The passage and seam summaries fail as they did in L139. No interior-solid root cause trace was needed because the measured interior-solid count is zero.

Measurements and checks

  • Clang 19 standalone build (-O2, -ffp-contract=off, four jobs) passed. L147 source stamp: c6b8a100355e271ddc40eb5973569d2ff4e2ba8d9c628c72f8e0af2493f0609f.
  • Tunnel Network: seeds 0–15, one composed run per seed, 430.715 s summed process time (13.542–51.457 s per seed); 15/16 passed. Logs and resolved layouts are under Saved/Linux/l147/seeds/tunnel-network/.
  • Vertical Shafts: seeds 0–15, one composed run per seed, 392.185 s summed process time (13.069–39.769 s per seed); 15/16 passed. Logs and resolved layouts are under Saved/Linux/l147/seeds/vertical-shafts/.
  • Across each matrix: 5,171,796 passage-interior capsule samples, 0 solid; 48/48 no-passage controls detected interior solid; 48/48 masked-gap ripple controls passed. Passage walkability, ripple, and all other existing gates passed for seeds other than the known seed-2 shell/seam failure.
  • Fast tier passed in 357.3 s: Saved/Linux/l147/fast/fast-summary.json. The first invocation stopped before running checks because the verifier's expected binary path was empty; it passed after using the completed L147 build there.
  • Owner identity comparison against the HEAD-stamped L140 binary passed: all pipeline hash lines, plan hash 538ee11456eb0dab, resolved-layout identity resolved-layout-v3-fnv64:d310515572977ebf, and resolved-layout SHA-256 50a0342902b6b897c5746fd86fac559244151894155cab6db0822229bfb877e4 matched. Details: Saved/Linux/l147/identity/summary.json.
  • git diff --check passed. No L80 timing, Unreal build, Windows build, or cross-compiler determinism run was requested or performed.

Runners: RunL147SeedSweep.py and RunL147OwnerIdentity.py. No Unreal Source/ files changed.

Remaining risk

The seed-2 interface 2–3 shell/seam failure remains unresolved. It does not put solid samples inside the passage interior and is unchanged from the known L139 Vertical Shafts result. This round made no generation fix.