Fr0zkaandClaude Opus 5.5
753cc9c259
BuildLinux.sh hashes a sorted manifest of compiled inputs into VOXELFORGE_SOURCE_STAMP, printed first by every executable; Standalone/Tools/verify/verify.py checks it before and after each tier and records binary/input hashes and the resolved recipe, style map and layout. Fast (163 s) runs on the canonical owner world only and asserts FI Style 2 / SW Style 1; medium (8 min) and full (26 min) add the style matrix, GCC, determinism and cost. Controls: stale source, mislabelled style-0 owner, all-black capture, L30/L31, L35/L36 and the L38 step are all rejected. No geometry change; owner hashes identical to L37. Research R6: leak table, generator contract and a 10-step metre-native migration plan. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016dhxM43dnWvUu2B5Ph1iG1
55 lines
4.8 KiB
Markdown
55 lines
4.8 KiB
Markdown
# Linux39 verification report
|
||
|
||
## Tier results
|
||
|
||
| Tier | Result | Wall time | Coverage |
|
||
|---|---:|---:|---|
|
||
| Fast | PASS | 161.0 s | Canonical owner only; resolved FI Style 2 / SW Style 1, passage ripple and coverage, adjacent-profile step, W1/W12 hashes, LOD0 rings and controls. |
|
||
| Medium | PASS | 472.9 s | Clang style overrides, transitions, L13/L9 proofs, LOD0–8 rings, owner shape/crease/fragments and captures. |
|
||
| Full | PASS | 1573.7 s (26m 14s) | Revalidated medium evidence, GCC supplement, 30 determinism cells, and five interleaved cost pairs per recipe. |
|
||
|
||
Fast is below the five-minute limit. The five-pair median cost change was +4.12% for Floating Islands (17.053 s L36, 17.756 s L39) and +0.94% for SurfaceWorld (1.800 s L36, 1.817 s L39). All commands used the CPU 0–11 / 6 GiB limits; the full tier had a 29-minute timeout.
|
||
|
||
## Source stamp and owner identity
|
||
|
||
`BuildLinux.sh` hashes a sorted manifest covering Core inputs, standalone pipeline sources and generated wrappers, build files, and the source-stamp helper. It includes the Unreal adapter when that adapter is enabled for compilation. The manifest SHA-256 is compiled as `VOXELFORGE_SOURCE_STAMP`; headless and capture executables print it as their first runtime line. The runner compares the expected stamp before and after execution, probes every available binary at both ends, and stops on a missing or mismatched stamp. Summaries record the stamp, binary SHA-256, input JSON SHA-256, resolved recipe, style and layout ID.
|
||
|
||
Current stamp: `644cd7ed510034b1855f249cc1dba338da2336c4071069792663c88810b59b90`. The canonical owner JSON SHA-256 is `7b095f9bcc6b90133d5a463f99f0b571133b2c549ce1ecf2d64fe9175e41e183`, unchanged from L37. FI and SW each produced 12 matching W1/W12 hashes, and both signatures match the saved L37 controls. No Core geometry source changed; C++ edits add build stamps and resolved-layout evidence.
|
||
|
||
The instrumented L36 cost baseline was exported from commit `8be1c4d` and carries its own recorded stamp: `28c74aac57643360cb563e07bae278e2e14eceda0b63a3a50cf5c9876003e765`.
|
||
|
||
## Negative controls
|
||
|
||
| Control | Evidence |
|
||
|---|---|
|
||
| Core edit after build | Runner refused the stale binary; original source bytes were restored. |
|
||
| Style-0 fixture named `owner_world.json` | Resolved snapshot exposed FI Style 0; the owner assertion requiring Style 2 rejected it. |
|
||
| All-black JPEG | Rejected at stddev 0 and coverage 0 (floors: 2.0 and 0.001). |
|
||
| L30 terrace rings | Rejected: 3.125 m flat run followed by a 0.243 m step. |
|
||
| L31 flattened floor | Rejected: amplitude range 0.000252 m. |
|
||
| L35 wall ribs / L36 floor washboard | Detected amplitudes 0.0918 m / 0.0943 m against 0.035 m / 0.015 m limits. |
|
||
| L38 owner step | Detected 0.2443 m against the 0.100 m adjacent-profile limit. Fitted period 7.875 m is diagnostic only. |
|
||
|
||
## Cleanup and remaining diagnostic
|
||
|
||
The duplicate L13/L9 runner was removed. The crease/fragment audit now decodes each owner mesh once for fragment, near-crease and overview scans. Style-override artifacts use `style-override-*`; `owner` runs use the canonical input with asserted resolved styles. Capture quality decodes pixels and enforces variance and coverage floors, including on JPEGs.
|
||
|
||
The current FI owner has six small fragments. That count remains diagnostic because a strict zero-fragment gate fails on the L37 owner; a baseline-backed fragment threshold remains open. The merged audit still reports the count alongside near and overview crease results.
|
||
|
||
## Failed attempts
|
||
|
||
- `full-final` stopped after determinism with a missing `VerificationRunner.bounded_run` method reference; switched to the shared bounded runner.
|
||
- `full-final2` could not find the historical temp tree's `Tools/verify` directory; created it before copying the stamp helper and computed the stamp from the temp repository root.
|
||
- `full-final3` was interrupted during an expected long FI W1 cell after the log went quiet; rerun allowed the cell to finish.
|
||
- `full-final4` passed all 30 determinism cells, then exposed an expected-name typo for the L36 binary (`...-stamped-clang` vs `...-stamped`). The corrected full run passed the baseline build and all cost pairs.
|
||
|
||
How-to instructions are in [`Standalone/Tools/verify/README.md`](../../../Tools/verify/README.md). Logs and JSON summaries are in `Saved/Linux/l39/`.
|
||
|
||
## Orchestrator review (Claude)
|
||
|
||
Accepted. I re-ran `verify.py fast` myself on the final tree: PASS in 162.7 s, the stamp matched, owner hashes match
|
||
L37 for both recipes at W1/W12, and the resolved style map asserted {FI: 2, SW: 1}. Note for readers: the
|
||
`style=[2]` on SurfaceWorld rows is the style of the single FI→SW passage, which belongs to the upper (FI)
|
||
strate. The owner world has no SW-owned passage, so "SW Style 1" never shapes a tunnel today. Earlier rounds'
|
||
"owner-sw Style 1" captures show the lower end of the Style 2 passage.
|