# C02 independent re-fit and follow-up audits (evidence)

Prepared 2026-10-17 by a Quik-side Claude session. DRAFT: not yet committed or pushed. Not part of the sealed TEST-ONLY package. Nothing here is an Unreal result.

## Inputs (all read from main; hashes SHA-256)

| File | Path on main | SHA-256 |
|---|---|---|
| C02 GLB | production/handoffs/quik/2026-10-16/crown-landmark/SHELL-C02/MODEL/PI-BLD-0001-V01_CrownLandmark_SHELL_C02.glb | b26adda0d266531a54cc6650b458faa92474e946b17a3ab32ca284ca6d5c9b9f |
| C02 OBJ | ...SHELL-C02/MODEL/PI-BLD-0001-V01_CrownLandmark_SHELL_C02.obj | bf8eb1c2c643d97596eee94f5519146966bdc37d0bbf7443151eed114538e1bb |
| INT R02 layout | production/assets/buildings/PI-BLD-0001-V01/INT/R02/06_Handoff/EditableLayout/INTERIOR_LAYOUT.json | aa14f38b150b966670397a3a12cd9bf054dafb387645a34149ebb5ed2a2942ff |
| Published C02 fit CSV | production/handoffs/quik/2026-10-16/crown-landmark/FIT-COVERAGE/C02/C02_FIT_COVERAGE.csv | 48500e9898a5712b2d92b67457e487c5b8312a110382cbeacbabe6deed2b5b4a |

The layout hash is identical at 4ce28d8 and d63324d. Hash raw git blobs or check out with `core.autocrlf=false`; a default Git for Windows checkout converts text files to CRLF and breaks every manifest hash.

## Re-fit command and versions

Python 3.11.9, numpy 2.4.6, shapely 2.2.0 (venv). Blender 5.2.2 LTS for the import check only.

```
python refit_c02.py <C02.glb> <INTERIOR_LAYOUT.json> <C02_FIT_COVERAGE.csv> refit_results.csv
```

Method: GLB converted to the source frame (x, -z, y), checked against the OBJ (source y = -GLB z). Podium_wall_1/_2 sliced at each room's floor level; Tower_curtainwall_skin at the room's floor and clear-height top; worst margin from the room rectangle to the slice outline. The governing-enclosure assignment (L00-L02 tier-1 wall, L03 tier-2 wall, L25/L26 skin) is the published interpretation and was not independently derived. Walls are single open surfaces (no thickness).

Result (`refit_results.csv`, rerun 2026-10-17): 75 FITS, 5 INSIDE_BELOW_ALLOWANCE, 1 OUTSIDE_WALL_AND_DECK; 0 status disagreements; wall margins equal the published CSV to rounding on all 81 rooms (max difference 0.000 m).

## Follow-up audits (audits/)

Four read-only subagent audits (general-purpose agents; model unknown) plus coordinator checks. Each folder has notes.md, findings.csv/json and its script. Only the Rib_R1 / L03-RM001 result was independently re-checked by the coordinator (`helper_rib_r1_l03_rm001.py`); the door, core and collision results are agent-reported measurements and have not been reproduced by a second pass.

- A_doors: 102 doors vs rooms, room graph, passage-width sweep (capsule unknown).
- B_vertical: cores RM090-RM094 across levels, shell and rib clearance, assumption-labelled stair sweeps.
- C_ribs: rib/mullion intrusions; only L03-RM001 by Rib_R1.
- D_collision: export integrity, floors, openings, interpenetration.
- blender_import.json + blender_import_check.py: Blender 5.2.2 headless import of the TEST-ONLY package GLBs (not Unreal).

Scratch note: helper_*.py were run from a scratch folder and still contain absolute scratch paths; refit_c02.py is the parameterised one.
