fix: close audited CAD delivery gate gaps
This commit is contained in:
@@ -25,14 +25,14 @@ Before modeling, write `scope_manifest.json` from the user request and source in
|
||||
"mode": "campus",
|
||||
"requested": ["1#", "2#", "3#"],
|
||||
"buildings": [
|
||||
{"id": "1#", "status": "PASS", "artifacts": ["outputs/1/model.glb"]},
|
||||
{"id": "1#", "status": "PASS", "artifacts": ["../../outputs/1/model.glb"]},
|
||||
{"id": "2#", "status": "BLOCKED", "reason": "unresolved XREF", "artifacts": []}
|
||||
],
|
||||
"site": {"status": "PASS", "artifacts": ["outputs/site_top.png", "outputs/site_manifest.json"]}
|
||||
"site": {"status": "PASS", "artifacts": ["../../outputs/site_top.png", "../../outputs/campus.glb"]}
|
||||
}
|
||||
```
|
||||
|
||||
Every `requested` ID must have exactly one ledger entry. `PASS` requires at least one existing non-empty artifact. `BLOCKED` requires a concrete reason and makes the overall campus result `BLOCKED`, not complete. The campus site must likewise be `PASS` with artifacts. Feed the ledger to `cad_showcase.py run --scope-manifest ...` or `verify_delivery.py --scope-manifest ...`.
|
||||
Every `requested` ID must have exactly one ledger entry and there may be no unrequested extras. `PASS` requires at least one verifiable model artifact: a GLB with mesh primitives, a Blender file with a valid header, or a schema-valid `building_spec.json`; an arbitrary non-empty text/JSON file does not satisfy scope. Campus site evidence must include a valid model or readable image. Relative paths resolve from the scope manifest's directory. `BLOCKED` requires a concrete reason and makes the overall campus result `BLOCKED`, not complete. Feed the ledger to `cad_showcase.py run --scope-manifest ...` or `verify_delivery.py --scope-manifest ...`.
|
||||
|
||||
## Placement Evidence
|
||||
|
||||
@@ -64,6 +64,8 @@ Use `assemble_campus.py` with per-building `rotation_deg` where needed. Require:
|
||||
- a top-view overlay or render showing each model inside its site outline;
|
||||
- numeric envelope dimensions and neighbour gaps in the notes/evidence.
|
||||
|
||||
Do not classify a small gap from distance alone. Corroborate a 0.15-0.30m attached-building/expansion-joint interpretation with the neighbour roof shown on the subject building's roof plan and matching floor elevations. A negative gap is always a placement error. A multi-metre hypothesis that pushes either building through the red line is also wrong.
|
||||
|
||||
`--allow-overlap` is for a documented shared structure or verified interlocking geometry, not a shortcut around placement bugs.
|
||||
|
||||
## Site Context
|
||||
|
||||
@@ -28,7 +28,7 @@ Tool discovery is environment variable, PATH, then platform defaults:
|
||||
- Blender: `BLENDER` -> PATH -> macOS/Windows install defaults.
|
||||
- ODA: `ODA_CONVERTER` -> PATH -> macOS/Windows install defaults.
|
||||
|
||||
ODA requires existing absolute input/output directories. It may show GUI errors while returning no useful stderr; an exit 0 with no output is still a failure. The script stages ASCII file names and verifies produced files.
|
||||
ODA requires existing absolute input/output directories. On macOS the command shape is `ODAFileConverter <abs_in_dir> <abs_out_dir> ACAD2004 DXF 0 1 "*.dwg"`; create both directories first. Bad arguments are often reported only through a GUI dialog, with empty stderr, and the process may return exit 0 while producing nothing. Treat “exit 0 + no DXF” as failure, check the screen for a blocked dialog, and keep the script's ASCII staging plus produced-file verification.
|
||||
|
||||
If native converters are unavailable, use the tested WASM bridge:
|
||||
|
||||
|
||||
@@ -185,6 +185,14 @@ Use `origin_strategy: "center"` for large CAD coordinates. It mirrors the A208 p
|
||||
|
||||
Use `clip` rectangles for irregular sheet layouts. Use `split_x` plus `side` when a sheet cleanly places plans left/right; this is less likely to accidentally cut off geometry than a tight clip rectangle.
|
||||
|
||||
Arm cross-evidence checks whenever the drawings contain the evidence:
|
||||
|
||||
- Copy per-floor areas from the 经济技术指标表/area labels into `declared_areas_m2`; use `declared_area_tolerance` only for a justified source convention. A model area off by roughly 100% usually means two plans or two wings were swallowed into one clip, not balcony noise.
|
||||
- Copy the above-ground count from 总图 labels or elevation/section ladders into `expected_storeys_above_ground`. When one typical plan represents a range, instantiate those floors and record that they are source-supported repetitions.
|
||||
- Pass every frame-detector config through `--frames`; list deliberately unused frames in `discarded_frames` with reasons.
|
||||
|
||||
These checks report `NOT_APPLICABLE` when no independent evidence exists. That is not proof of correctness; document what was searched before leaving them unarmed.
|
||||
|
||||
Keep `expand_blocks` false for the first pass. If the audit shows important geometry in `INSERT`, set `expand_blocks` to `true` or to a list of block-name patterns, then rerun extraction/audit/overlays. When virtual block entities live on layer `0`, `block_insert_layer_fallback: true` lets the parent INSERT layer participate in category matching.
|
||||
|
||||
## 5. Spec Extraction
|
||||
@@ -203,6 +211,7 @@ Review `building_spec.json` before rendering:
|
||||
- Segment counts are not dominated by dimensions/hatches/annotations.
|
||||
- Footprint polygons exist when slabs/roofs are meant to follow CAD floor plates.
|
||||
- `default_disabled_categories` does not hide `slab` for presentation/digital-twin views.
|
||||
- Generated cap edges do not protrude as slab fins. The historical `cap_buffer=1.3m` value fits colonnade/overhang cases; use about `0.1m` for ordinary facade-aligned floors, then confirm front/side views.
|
||||
|
||||
## 6. Agent Audit and Overlay Review
|
||||
|
||||
@@ -238,6 +247,7 @@ Agent review rules:
|
||||
- If frequent blocks appear in the audit, inspect block names and decide whether to recursively explode block geometry.
|
||||
- If hatches carry wall/slab/column information, do not rely only on line extraction.
|
||||
- Continue only when missing geometry is acceptable for the requested output level.
|
||||
- A floor can be internally self-consistent yet occupy the wrong repeated slot. If wall-profile correlation finds a strong non-zero shift, treat it as the 13号 F3 “错槽” pattern and correct the frame offset rather than trusting envelope centers.
|
||||
|
||||
## 7. Blender Generation
|
||||
|
||||
|
||||
@@ -51,6 +51,8 @@ node scripts/web_smoke_test.mjs \
|
||||
|
||||
Install Playwright/Chromium first when needed. The script checks real 1440×900/390×844 viewport dimensions, desktop/mobile control visibility, console/page/request failures, and idle render behavior. Feed the report to final verification with `--browser-report`. Supplying `--web-dir` makes the report mandatory: no report is `BLOCKED`, never PASS.
|
||||
|
||||
The verifier validates the report contract, not just its top-level `status`: all required named checks must exist and be PASS, viewport details must match exactly, and console/page/request arrays must be present and empty. A hand-written `{status: PASS}` report is invalid.
|
||||
|
||||
When an execution tool already has a `run_in_background` option, launch `python3 -m http.server ...` as the foreground command of that background process. Do not append shell `&`; the child server can die with the wrapper. Before presenting the viewer, probe the URL and keep that server process alive. Present the formal `web/index.html` URL when Web was requested; `viewer.html` is a separate offline fallback, not a substitute.
|
||||
|
||||
## Strict Evidence Manifest
|
||||
@@ -61,6 +63,8 @@ Run `verify_showcase.py --strict` for spec previews and GLB budgets, then:
|
||||
python3 scripts/verify_delivery.py \
|
||||
--source-report outputs/<project>/cad_read_report.json \
|
||||
--spec outputs/<project>/building_spec.json \
|
||||
--audit outputs/<project>/extraction_audit.json --require-audit \
|
||||
--scope-manifest work/<project>/scope_manifest.json --require-scope-manifest \
|
||||
--variants-manifest outputs/<project>/delivery/variants_manifest.json \
|
||||
--web-dir outputs/<project>/web \
|
||||
--offline-viewer outputs/<project>/viewer.html \
|
||||
@@ -71,4 +75,4 @@ python3 scripts/verify_delivery.py \
|
||||
--evidence-out outputs/<project>/evidence_manifest.json
|
||||
```
|
||||
|
||||
The verifier checks schema versions, audit gates, optional campus scope, GLB headers/mesh structure/naming/budgets, render dimensions/nonblank variance, Web assets/placeholders/features/variants, offline payload, offline `file://` browser load, and served-Web browser report. Only overall `PASS` is complete. `WARN`/`SKIP` produce overall `BLOCKED` unless the exact check has an explicit `--waiver 'CHECK=REASON'`; `FAIL` cannot be waived.
|
||||
The verifier checks schema versions, audit gates, scope artifacts, GLB headers/mesh structure/naming/budgets, render dimensions/foreground occupancy/noise, Web assets/placeholders/features/variants, offline payload, offline `file://` browser load, and served-Web browser report. Only overall `PASS` is complete. Applicable `WARN`/`SKIP` produce overall `BLOCKED` unless the exact check has an explicit `--waiver 'CHECK=REASON'`; `NOT_APPLICABLE` is recorded but non-blocking, and `FAIL` cannot be waived.
|
||||
|
||||
@@ -9,7 +9,7 @@
|
||||
5. Honesty rules
|
||||
|
||||
Prioritized playbook for pushing a working showcase model toward digital-twin
|
||||
presentation quality. Apply AFTER the geometry passes the One-Shot Definition
|
||||
presentation quality. Apply AFTER the geometry passes the Hard Definition
|
||||
of Done — polish never excuses a wrong model. Each item names the code to
|
||||
touch. Ordered by value-per-effort.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user