AeroBIM

Русская версия

Checkpoint: NO_GO. Не прячем. Тихий SKIPPED в IDS-контуре закрыт (AEROBIM-IDS-IFC-VERSION). Осталось: нет корпуса «ПД РФ + заключение экспертизы»; нет подписанного профиля приёмки «Самолёта» (IDS МОГЭ есть, это другое); federated MEP clash NOT_VERIFIED (duplex AABB 654 overlap pairs measured; not clash). Кодом GO не ставится.

Видео 3 мин: artifacts/demo/ — запись 19.08, человек (сейчас placeholder).
Встреча трекера 14.08: docs/demo/TRACKER_MEETING_2026_08_14.md.
План 13–20.08: docs/pilot/KT2_7DAY_PLAN_2026_08_13.md.

Open-source acceptance-criteria assistant for openBIM packages (IFC + IDS + cross-document evidence): helps an expert catch package contradictions before site work.

A schedule on a PDF sheet shows one area; the IFC wall with the same GUID shows another. Each file looks fine alone; the error appears only when they are compared. AeroBIM raises a finding with provenance to the sheet and the GUID, leaves the verdict to the expert, and never authorizes Shared→Published.

Project-package analyze; IFC / IDS / cross-doc; deterministic Shared-gate summary.passed (fail-closed); provenance; structural BCF ZIP; HITL; Docker offline-bundle; CI. Full map — Technical depth below. LOC/test counts — runtime baseline.

Pre-construction package review (expertise / chief engineer / doc QC): model ↔ drawings ↔ requirements. Not a CDE replacement and not a field defect journal. Samolet contour / pilot framing: docs/docs.md · docs/samolet.md.

Checkpoint: NO_GO

Samolet TechLab Task 07 is not ready for customer sign-off. Open blockers: RT-001 (no public «RF PD + expertise conclusion» corpus; AEC-Bench / IFC-Bench / GNI exist and are not that corpus), RT-002 (no Samolet-signed acceptance profile; official MOEXP IDS exist — that is not the same thing), RT-003 (public federated IFC inventory exists; clash NOT_VERIFIED; not MEP delivered) — see audit/reports/CRITICAL_BLOCKERS.md. Claims SSOT: audit/reports/CLAIMS_LOCK_2026_07_17.md · dated eng freeze: audit/reports/CLAIMS_LOCK_2026_07_31.md · verified vs planned: docs/pilot-claim-boundary-2026.md · Aug 2026 eng status: docs/ENGINEERING_STATUS_2026_08.md · Tier-0 docs: docs/TIER0_INDEX.md · verdict ownership: docs/architecture/ADR-001-verdict-ownership-2026.md. Forbidden until evidenced: product accuracy >90%, DWG-ready, MEP delivered, CDE-ready BCF, independent calc correctness.

Engineering readiness improved (2026-07 → 2026-08) without closing customer blockers: LIC-001 Option B; P2-04 / P2-02 honesty; Docker offline track; P0 TechLab eng package WP-01…08 (runtime baseline, Hybrid advisory pre-gate, signature envelope, norm pack v2, package completeness, open-corpora n=7, quality protocol interim 0.60, README/baseline sync) — see docs/ENGINEERING_STATUS_2026_08.md · Red Team rollup docs/quality/RED_TEAM_P0_ROLLUP_2026_08_02.md. Fixture GO ≠ Checkpoint GO.

AeroBIM runs a deterministic Shared-gate style check (ISO 19650 framing: evidence for Shared, not contractual Published authorization). It fuses IFC property/quantity checks, IDS, drawings, and calculation text into a single report with explicit capability honesty, finding provenance, and BCF ZIP export. Independent CDE import and customer accuracy claims remain out of scope until evidenced. Architecture SSOT: docs/architecture/TARGET_HYBRID_ARCHITECTURE_TZ_2026.md.

What works: project-package analyze; IFC/IDS/cross-doc; summary.passed Shared-gate (ADR-001); pilot/production fail-closed profiles; ACL 404; SSRF outbound guard; provenance stamp/persist; BCF 2.1/3.0 structural ZIP; HITL review-events; evidence bundle CLI (python -m aerobim.tools.export_evidence_bundle); annotation claimed-GUID presence confirm (P2-04); core PDF via pypdfium2/pdfminer (LIC-001 Option B); HybridRouteGate advisory pre-gate (WP-02); detached signature envelope audit (WP-03 ENG_PARTIAL); norm pack v2 eligibility (WP-04); package completeness inventory (WP-05); open-corpora profiles (WP-06 regression/timing only); quality measurement protocol (WP-07 Wilson planner; interim 0.60); L1 open-bench IFC-Bench smoke + AECV live counting on Yandex Qwen (claim_level=open_bench_only, ≠ RT-001 — evidence); Docker offline-bundle smoke; pytest / vitest counts SSOT via runtime baseline (frontend.tests_passed when recorded; else see baseline/CI).

Experimental: OpenCDE BCF API push; BCF 3.0 consumer path; optional clash/OCR extras; IFC KG advisory scaffold; MEP federated ENG_FIXTURE graph + AABB broadphase (capability stays NOT_VERIFIED); advisory LlmExtractionPort (regex/Kimi/Qwen eval harness — never verdict).

Available (eng): PackageOutcome on summary.outcome (pass / pass_with_warnings / review_required / blocked / failed); run manifest + reproducibility hash; stage timeout budgets; Hybrid AI routing (classify/policy/guard/audit + WP-02 HybridRouteGate advisory pre-gate on Analyze + kimi smoke PUBLIC egress gate) — domain-pure, verdict-neutral (OFF==ON), never sets summary.passed; masking ≠ anonymity; Checkpoint NO_GO.

Planned: Stage-3 finding field expansion; profiling-driven performance wave; customer-gated RT-001/002/003 evidence.

Needs customer: RF PD+expertise corpus · Samolet-signed acceptance profile · measured federated MEP on public IFC (CRITICAL_BLOCKERS, DATASETS). Official MOEXP IDS are already in-repo. Public federated models exist; we have not published a measured run.

Not claimed: product accuracy >90%; customer ≤30 min SLA; native DWG; MEP system clash delivered; independent calc correctness; CDE-ready BCF; bare-metal offline without Docker; AABB/connects = verified geometric clash. See capability-claim-matrix · PROJECT_STATUS_AUDIT · ENGINEERING_STATUS_2026_08 · pilot-protocol · benchmark-evidence.

Statuses below are repository / fixture capabilities unless marked otherwise. Optional extras and fail-closed policies govern whether a green summary.passed is honest.

All three releases pass through the same IfcOpenShellValidator and IfcTesterIdsValidator adapters. Pset/property name divergence between releases is surfaced as a ValidationIssue rather than a silent skip. IFC2x3, IFC4, and IFC4x3 fixture files live in samples/ifc/. See docs/ifc-compatibility-matrix.md for the formal compatibility matrix and per-feature degradation rules.

Canonical taxonomy: docs/architecture/BCF_EVIDENCE_LADDER_T0_T4_2026_07.md.

Allowed: structural ZIP AVAILABLE. Forbidden until T2: “BCF ready for CDE”, “CDE interoperable”.

Iteration B.1 has started with a compatibility-first storage foundation:

• ObjectStore domain port for binary artifacts (put/get/delete/presign);
• LocalObjectStore for current local/runtime flows;
• S3ObjectStore for S3/MinIO-compatible buckets via optional enterprise extras;
• PostgresAuditStore foundation that adds a Postgres report-summary index while keeping full payload round-tripping on the existing JSON/object path;
• AEROBIM_REPORT_TTL_DAYS retention knob for persisted report payloads.

Current behaviour is intentionally safe-by-default:

• without enterprise extras, AeroBIM keeps working with local storage;
• when AEROBIM_DB_URL and enterprise dependencies are available, report summaries are indexed in Postgres;
• IFC source binaries and persisted drawing previews are stored behind the ObjectStore abstraction, so S3/MinIO rollout no longer requires HTTP contract changes.

# Clone
git clone https://github.com/KonkovDV/AeroBIM.git
cd AeroBIM/backend

# Create virtual environment
python -m venv .venv
source .venv/bin/activate  # Linux/macOS
# .venv\Scripts\activate   # Windows

# Install (pdf-agpl = overlay PNG via PyMuPDF; raster = RapidOCR degrade path)
pip install -e ".[dev,raster,pdf-agpl]"

# KT#2 vertical slice (~10 min from clone): fail-loud; writes artifacts/vertical-slice-demo/
python -m aerobim.tools.run_demo_vertical_slice
# Open artifacts/vertical-slice-demo/report.html — fragment, finding, evidence, non-PASS verdict.

# Optional extras
# pip install -e ".[clash]"    # enable geometry clash detection
# pip install -e ".[docling]"  # enable non-text document extraction
# pip install -e ".[enterprise]"  # enable S3/Postgres enterprise storage adapters

# Run tests
pytest tests -q

# Extraction quality gate (Russian AEC corpus)
python -m aerobim.tools.evaluate_extraction --min-macro-f1 0.70
python -m aerobim.tools.evaluate_detection_precision \
  --labels ../samples/benchmarks/detection-precision/labels-synthetic.json \
  --detections ../samples/benchmarks/detection-precision/detections-synthetic.json \
  --min-precision 0.6 --min-recall 0.6 --min-f1 0.6

# Seed one deterministic runtime smoke report
python -m aerobim.tools.seed_smoke_report

# Or run the full live review smoke chain in one command
python -m aerobim.tools.run_live_review_smoke

# Or run the baseline throughput rail against the representative benchmark pack
python -m aerobim.tools.benchmark_project_package --iterations 1 --warmup-iterations 0

# Or run the second fire-compliance benchmark profile explicitly
python -m aerobim.tools.benchmark_project_package --pack ../samples/benchmarks/project-package-fire-compliance.json --iterations 1 --warmup-iterations 0

# Or run the stress multisource benchmark profile explicitly
python -m aerobim.tools.benchmark_project_package --pack ../samples/benchmarks/project-package-stress-multisource.json --iterations 1 --warmup-iterations 0

# Start server
python -m aerobim.main
# → http://127.0.0.1:8080/health

Before pushing to main, run the same baseline checks used by CI:

cd AeroBIM/backend
python -m ruff format --check src tests
python -m ruff check src tests
python -m mypy src
pytest tests -q

If ruff format --check reports files to reformat, run:

python -m ruff format src tests

Verified capabilities are backed by tests, API contracts, or persisted report artifacts. Planned / missing contours (DWG, human CV, MEP system clash, calculation correctness, customer accuracy) are explicit on GET /v1/system/capabilities and in the Claims Lock.

cd backend
python -m aerobim.tools.benchmark_project_package --iterations 1 --warmup-iterations 0
python -m aerobim.tools.measure_package_sla --corpus-kind fixture
python -m aerobim.tools.verify_bcf_structural_handoff
python -m aerobim.tools.run_ablation_study
python -m aerobim.tools.generate_benchmark_report --output-dir ../docs/evidence
python -m aerobim.tools.export_runtime_baseline
python -m aerobim.tools.export_evidence_bundle \
  --pack ../samples/benchmarks/project-package-techlab-demo.json \
  --output ../artifacts/evidence-bundle/techlab-demo

# P2-04 wall-guid presence demo (fixture GO pin)
python -m aerobim.tools.export_evidence_bundle \
  --pack ../samples/benchmarks/project-package-wall-guid-demo.json \
  --output ../artifacts/evidence-bundle/checkpoint2-wall-guid

| GET | /v1/system/capabilities | Static honesty surface (DWG/CV/MEP/calculation claim boundary) | | GET | /health | Readiness probe | | POST | /v1/validate/ifc | Validate IFC against requirements + IDS | | POST | /v1/analyze/project-package | Multimodal validation (spec + calc + drawing + IDS + IFC) | | POST | /v1/analyze/project-package/reinforcement-digest | OpenRebar provenance digest (сверка labels; not correctness verification) | | POST | /v1/analyze/project-package/submit | Accept a same-process background analysis job for larger packages | | GET | /v1/analyze/project-package/jobs/{job_id} | Poll async project-package job status | | POST | /v1/uploads | Multipart document ingest; returns storage-relative path for analyze | | GET | /v1/reports | List persisted reports with optional project, discipline, and passed filters | | GET | /v1/reports/{id} | Get report by ID | | POST | /v1/reports/{id}/review-events | Append HITL review telemetry (does not affect pass/fail) | | GET | /v1/reports/{id}/review-events | List review events for a report | | GET | /v1/reports/{id}/review-kpi | Aggregate triage/acceptance KPIs | | GET | /v1/reports/{id}/source/ifc | Download the report-scoped IFC source for browser viewing | | GET | /v1/reports/{id}/drawing-assets/{asset_id}/preview | Download a report-scoped drawing preview for 2D evidence overlays | | GET | /v1/reports/{id}/export/json | Download JSON export | | GET | /v1/reports/{id}/export/html | Download HTML export | | GET | /v1/reports/{id}/export/bcf | Download BCF 2.1 ZIP by default; use ?version=3 for BCF 3.0 |

POST /v1/analyze/project-package also supports optional OpenRebar provenance fields:

• reinforcement_report_path: path (inside AEROBIM_STORAGE_DIR) to an OpenRebar canonical *.result.json report;
• reinforcement_source_digest: expected SHA-256 digest for report provenance fingerprint checks.
• reinforcement_waste_warning_threshold_percent: optional waste threshold (percent) for coordination warnings.
• reinforcement_provenance_mode: advisory (default) or enforced to escalate OpenRebar provenance warnings into blocking errors.

Use /v1/analyze/project-package/reinforcement-digest to generate reinforcement_source_digest directly from a stored OpenRebar report before calling project-package analysis.

For offline or CI shell workflows, use:

python -m aerobim.tools.openrebar_provenance_digest <path-to-openrebar-result.json>

When provided, AeroBIM adds cross-document warnings if:

• OpenRebar report contract ID is unexpected;
• OpenRebar optimizer indicates fallback master solver usage;
• OpenRebar master-problem strategy does not indicate a HiGHS-backed path;
• project context mismatches (project_name vs metadata.projectCode);
• supplied provenance digest does not match report fingerprint.
• reported summary.totalWastePercent exceeds the configured warning threshold.

Five-layer Clean Architecture with strict inward dependency direction:

core/          DI container, tokens, config (no project imports)
domain/        Immutable models, Protocol ports, logging contract
application/   Use case orchestration (requirement fusion, cross-doc detection)
infrastructure/ Adapters: IfcOpenShell, IfcTester, Docling, IfcClash, BCF, filesystem
presentation/  FastAPI HTTP API, correlation middleware

Infrastructure now also includes an artifact ObjectStore seam plus an optional Postgres summary-index adapter for Iteration B.1.

48 domain Protocol ports → 72 infrastructure adapter modules → 63 DI tokens — wired in bootstrap_container(). Counts are live inventory regenerated into docs/evidence/runtime-baseline-latest.json (architecture_inventory) and checked in CI via export_runtime_baseline --check-readme against README EN/RU; do not hand-edit older 20/30/28 or stale 47/71/60 figures. Report payloads include an explicit capabilities object (ok / skipped / failed) so optional engines (clash, IDS, unit scale, raster, schema) cannot silently look like a clean PASS. Any FAILED capability forces summary.passed=false.

All settings are read from environment variables (see backend/.env.example):

AEROBIM_ALLOW_ANONYMOUS_DEV AEROBIM_API_BEARER_TOKEN AEROBIM_API_TENANT_ID AEROBIM_APP_NAME AEROBIM_BCF_API_BASE_URL AEROBIM_BCF_API_PROJECT_ID AEROBIM_BCF_API_TOKEN AEROBIM_BCF_API_VERSION AEROBIM_BSI_API_TOKEN AEROBIM_BSI_VALIDATION_URL AEROBIM_CLASH_AFFECTS_PASS AEROBIM_CORS_ORIGINS AEROBIM_CROSS_DOC_SEVERITY AEROBIM_DB_URL AEROBIM_DEBUG AEROBIM_ENV AEROBIM_GATES_ATTESTED AEROBIM_HOST AEROBIM_HTTP_RATE_LIMIT_PER_MINUTE AEROBIM_HYBRID_PROVIDER_CONFIG AEROBIM_IFC_PARSE_CACHE_DIR AEROBIM_KIMI_API_BASE_URL AEROBIM_KIMI_API_KEY AEROBIM_KIMI_CACHE_DIR AEROBIM_KIMI_CACHE_NAMESPACE AEROBIM_KIMI_CACHE_PROJECT AEROBIM_KIMI_MODEL AEROBIM_KIMI_REASONING_EFFORT AEROBIM_LLM_429_RETRIES AEROBIM_LLM_ADVISORY_ENABLED AEROBIM_LLM_ADVISORY_MAX_ISSUES AEROBIM_LLM_ALLOWED_HOSTS AEROBIM_LLM_API_KEY AEROBIM_LLM_AUTH_SCHEME AEROBIM_LLM_BASE_URL AEROBIM_LLM_BUDGET_LEDGER AEROBIM_LLM_BUDGET_TZ AEROBIM_LLM_DATA_LOGGING_ENABLED AEROBIM_LLM_FOLDER_ID AEROBIM_LLM_LOCAL_ENABLED AEROBIM_LLM_MAX_COMPLETION_TOKENS AEROBIM_LLM_MAX_CONCURRENT AEROBIM_LLM_MAX_TOKENS_PER_CALL AEROBIM_LLM_MAX_TOKENS_PER_DAY AEROBIM_LLM_MAX_TOKENS_PER_RUN AEROBIM_LLM_MODEL AEROBIM_LLM_MODEL_REVISION AEROBIM_LLM_MODEL_SHA256 AEROBIM_LLM_PROVIDER AEROBIM_LLM_RESPONSE_FORMAT_MODE AEROBIM_LLM_SEND_SEED AEROBIM_LLM_TIMEOUT_SECONDS AEROBIM_MAX_IFC_BYTES AEROBIM_MEP_AABB_FILTER AEROBIM_MEP_FEDERATED_SCOPE_PATH AEROBIM_MEP_SCOPE_MEMO_REF AEROBIM_NORM_RULE_PACK AEROBIM_OIDC_AUDIENCE AEROBIM_OIDC_ISSUER AEROBIM_OIDC_JWKS_EXTRA_HOSTS AEROBIM_OIDC_JWKS_URL AEROBIM_OIDC_ROLES_CLAIM AEROBIM_OIDC_TENANT_CLAIM AEROBIM_PDF_BACKEND AEROBIM_PORT AEROBIM_PRIORITY_PROFILE AEROBIM_REDIS_URL AEROBIM_REMARK_LOCALE AEROBIM_REPORT_TTL_DAYS AEROBIM_REQUIRE_CLASH AEROBIM_REQUIRE_MEP_SYSTEM_CLASH AEROBIM_S3_ACCESS_KEY_ID AEROBIM_S3_BUCKET AEROBIM_S3_ENDPOINT_URL AEROBIM_S3_PREFIX AEROBIM_S3_REGION AEROBIM_S3_SECRET_ACCESS_KEY AEROBIM_SIGNOFF_PROFILE AEROBIM_STORAGE_DIR

aerobim/
├── backend/                 # Python FastAPI backend (see generated baseline below)
│   ├── src/aerobim/         # Source: core → domain → application → infrastructure → presentation
│   ├── tests/               # Backend test suite (see generated baseline below)
│   └── pyproject.toml
├── clients/revit-plugin/    # Thin authoring-side client boundary (planned)
├── docs/                    # TechLab jury docs only (see docs/README.md)
├── frontend/                # Browser review shell
├── audit/                   # Claims lock, blockers, citeable honesty fixtures
├── samples/                 # IFC, IDS, drawing, spec fixtures
├── .github/workflows/       # CI pipeline (lint, typecheck, test, benchmark-smoke) + manual release-readiness gates
└── LICENSE                  # MIT

tests_passed: backend=2163, frontend=48; commit 3489cad44697; see docs/evidence/runtime-baseline-latest.json · src ~61743 LOC; tests ~42440 LOC; extraction macro_f1=0.86 (fixture corpus; not product accuracy)

Public GitHub is the TechLab jury pack only: code + TZ / claims / architecture + curated eng status / Red Team summaries under docs/quality/. Operator runbooks, phase RT dumps, and archive stay in .local/ (not published).

Use scripts/git_commit.ps1 or the VS Code task AeroBIM: commit. Keep authorship honest: Co-authored-by: trailers are allowed when an assistant materially contributed. Optional hooks (pass-through, no trailer stripping):

git config core.hooksPath .githooks

Suggested repo About — .github/repository-metadata.md.

• Contributing
• Security Policy
• Citation Metadata
• Support
• Maintainers
• Release Policy

Use the manual GitHub Actions workflow .github/workflows/release-readiness.yml when preparing a release candidate.

It runs benchmark rails by default and can optionally run the full live review smoke harness with browser artifacts. The live-smoke path now installs Playwright and Chromium inside the workflow job so browser capture is reproducible in CI. Main CI benchmark-smoke runs now also emit a compact benchmark summary table in workflow output and artifacts. When needed, require_live_smoke_gate=true enforces live-smoke execution as a mandatory policy gate for that release-readiness run. CI benchmark-smoke now also runs advisory threshold evaluation from samples/benchmarks/benchmark-thresholds.json and publishes the threshold summary alongside benchmark artifacts. Release-readiness benchmark rails now support benchmark_threshold_mode (advisory or enforced) plus explicit threshold profile path selection.

• Python 3.12+, FastAPI, Uvicorn
• IfcOpenShell / IfcTester / IfcClash (buildingSMART toolchain)
• web-ifc + Three.js for browser-side IFC review
• pypdfium2 + pdfminer.six for core PDF (LIC-001 Option B); optional PyMuPDF only via pdf-agpl
• RapidOCR only when .[raster] is installed (EI OCR-aware signals PARTIAL)
• Docling (optional, document parsing)
• 5-layer Clean Architecture, constructor DI, Protocol ports

MIT for AeroBIM-authored code. Third-party components keep their own licenses:

• pypdfium2 / pdfminer.six / Pillow — production PDF path (permissive; see inventory)
• PyMuPDF — dual AGPL-3.0 / Artifex commercial; optional pdf-agpl only (absent from runtime lock / Docker after LIC-001 Option B)
• IfcOpenShell / IfcTester — LGPL-3.0-or-later
• web-ifc — MPL-2.0

Machine-readable inventory: audit/dependency_license_inventory.json · policy: docs/license-policy-2026.md.
Not a court opinion. Do not claim “entire product is MIT” without third-party disclosure.