What Exists Today, What's Missing, and Where We Sit¶

We BIM living the GAP between
DESIGN and our SPREADSHEET
The industry designs buildings. Who compiles them?

How to read this — the honest frame¶

This project is broad. To keep it honest, everything below is sorted into three tiers, and the line between them is never blurred:

Tier 1 — Landed. Works today, use-as-is-where-is, and every claim carries a witness (a §-logged, falsifiable test you can run). This is real commercial value now.
Tier 2 — Wedges. Small hardening or a short build away — the commercial on-ramps. Stated as targets, never as done.
Tier 3 — Frontier. The demonstrated promise — proven in architecture, not yet finished. Shown to be possible, labelled as unfinished.

If you grep the repo, the Tier-1 claims must hold and the Tier-2/3 ones must read as runway. That discipline is the point.

The core problem: geometry is not intent¶

Most BIM tools store geometry as the source of truth. An IFC file carries the building — 51,000+ elements of geometry, relationships, properties — in one monolithic file. Lose the IFC, lose the building. But a 200 MB IFC captures what was drawn, not what was meant. You cannot ask it "give me a building like this but with 4 m ceilings," because the intent was never separated from the output.

Autodesk solved it behind a proprietary wall. Revit's .rvt keeps full spatial fidelity internally, but is editable only in Revit with a shelf-life tied to Autodesk's support cycle [6]. Leave via IFC and "there's always loss of data... all constraints are lost and component parametrics are gone" [7].

The openBIM world has no equivalent. Bonsai/BlenderBIM is IFC-native [8], but IFC is an exchange format, not a compilation target — it does not decompose a building into a reusable BOM recipe, compile from it, and verify the round-trip [9].

The compilation challenge: extract a building's intent from its geometry, express it as a reusable recipe (BOM), recompile the recipe back into spatially correct geometry — and prove the recipe is faithful. Construction is industrialised manufacturing [11], yet architecture is drawn as art; the compiler gives determinism to art. A beam is either at (3200, 0, 2700) or it isn't — no probabilistic guessing [12].

That is what the BIM Intent Compiler does, in two databases — and it is the proven core that makes Tier 1 real.

The proven core (why the rest can be believed)¶

Input DB — an IFC (or OBJ/STL/DAE/GLB) is extracted into a normalised SQLite DB. Geometry hell is resolved here: origin divergence, unit mismatch (the 1000× metres-vs-mm error), axis ambiguity, and GUID identity — documented industry problems [1][2][3][4][5]. Every element becomes a row; every spatial relationship a foreign key. The building is SQL-queryable.
BOM abstraction — 51,000 elements decompose into ~700 BOM lines (73× compression) via formula verbs (TILE, ROUTE, FRAME, CLUSTER). This is the intent: not "12 wall meshes at these coordinates" but "12 of product W-EXT-200, tiled at 2.5 m along the north facade."
Output DB — the BOM + shared library recompiles into spatially placed geometry, each element carrying its original GUID. A 200 MB IFC becomes a ~10 KB semantic definition. The output is disposable — delete it, recompile, the geometry reproduces.
Rosetta Stone — Input vs Output across verification gates (counts, volumes, geometry hashes, spatial digests, GUID provenance, transforms, materials). 21 buildings from 9 authoring tools; 116/157 gates PASS, 4 ALL GREEN; worst-case positional error 0.002 mm. See SPATIAL_COMPILATION_PAPER.md.

The hard problem was always the spatial compilation. Everything downstream — 4D, 5D, cost, ERP — is a projection of the same verified BOM.

Tier 1 — Landed: use-as-is, where-is (witnessed today)¶

These work now, in a browser tab, no install. Each carries a witness you can run.

Capability	What it does	Witness
IFC handoff	Drop IFC/OBJ/STL/DAE/GLB/glTF/3DS/FBX → queryable DB → view, classify, export back to IFC. Geometry hell resolved at import.	Rosetta gates; `import.js` round-trip
4D Time Machine	Construction-sequence playback from BOM depth; stacked S-curve folded from real orders (Σ == PlannedAmt).	`W-SHOP-SCURVE`
5D cost (editable)	BOQ + cost rollup with editable per-jurisdiction rate templates; Variation Order Excel (FIDIC Clause 12).	`4D5DAnalysis.md`; VO demo [10]
BIM↔ERP — to the cent	A BIM-pushed building folds into a real procurement/project order and ERP documents, reproducing iDempiere/Odoo output at `maxDiff=0c`. No other tool connects BIM to ERP over one signed log.	`W-PROJ-FOLD`, `W-GW-HOSP-FOLD`, `W-FOLD-COMPLETE`
Budget vs Actual (EVM)	Planned vs Committed at project + phase + task grain; CV/SV/CPI/SPI in BigDecimal; cost overrun surfaced on the 4D S-curve.	`W-GW-HOSP-COSTVAR`; `proj_control.js`
What-If (cost)	Speculative VO branch — revised = original + approved + pending — kept separate from the official ledger, reversible.	`W-FIN-BLUE-SPEC` (5/5)
Dashboard / analytics	Generic multi-view over any data model: donut grid, "By-X" group-by chips that fill the grid, pivot lens, scrubbable timeline filmstrip, CSV/SVG/PNG export. Field-driven, not hardcoded per table.	`W-DASHBOARD`; `pivot_lens.html`
POS sale loop	Ring → complete → backflush BOM → hold/recall → deliver-later → register, all over the signed op-log; the signed orderline doubles as the buyer's receipt artifact. WAN bench to 10,000 stations with idempotent retry + email-backup recovery.	`W-POS`; `poc_pos_wan_scale.js` (B1–B7)
iDempiere DB extraction	Connect a live iDempiere PostgreSQL → raw, non-inventive PG→SQLite extraction (`--list-clients`, `--masters`).	`migrate_agent.js`; `ERP_RAW_MIGRATION.md`

The standout no one else has: building → procurement order, in one browser, tied to the cent. The Dashboard is the answer to "where's your Odoo kanban / SAP analytics" — and because it's AD-field-driven, it's one dashboard for every data model, not a bespoke screen per report.

Tier 2 — Wedges: small hardening, the commercial on-ramps¶

Real markets, short runway. Stated as targets — not yet shipped.

POS → Malaysian e-invoicing + personal accounting (the long tail). The LHDN/MyInvois mandate forces every business onto e-invoicing, and the government's own central service has buckled under server-side load — a structural opening for a serverless, local-first POS that each merchant runs themselves. The sale loop is landed; the wedge is hardening it against real-world bugs and adding the compliance surface. The signed orderline is already a receipt artifact — evolving it into a MyInvois-format submission is the build. (Not yet in code; this is the target.)
Touch-kitchen + self-order tabs + QR payment. A self-order surface (URL-fetched remote ordering, QR payment display, payment-status fold returned by email) sits naturally on the same op-log. (Not yet in code; the POS loop it rides is.)
iDempiere DB health-check report — "the diagnosis; the diet is optional." A read-only analysis pass over the already-working extraction: scan a user's DB for dirty data, orphans, GL imbalances, and a migration-gap score. Sold as a health check + migration plan — standalone value even if they never migrate. Low lift (the extraction plumbing exists), high value (migration paralysis is real). (Analysis pass not yet built; extraction is.)
What-If (schedule ripple). Finish-to-start cascade on the timeline. Engine done; browser drive pending.

Tier 3 — Frontier: the demonstrated promise (the dragon's head)¶

DAGeVu modeller — a browser-native BIM authoring tool whose endgame is to sever the Revit-license tether. The hard part is shipped and witnessed: an occt-wasm B-rep kernel as a pure ops → mesh fold, the signed op-log as the feature tree (scrub, undo, tamper- evident), and IFC4 export that round-trips (IfcWall + profile + IfcOpeningElement, re-imports exact). A user can today author a few walls, a door, an opening, and MEP runs, and export usable IFC — without Revit.

Honest distance to the mountain top: ~35% — but the 35% that's done is the part that takes others decades. The gaps are UX/performance, not physics: dimension-driven parametric edit, a regen cache (so a 50-wall model stays instant), multi-select + properties panel, native wall/door/window tools, and snap-to-geometry. The read: weeks, not years, because kernel fidelity, signed history, and IFC round-trip are already proven.

Witnesses: W-BONSAI-* (bonsai_signed_live.js, bonsai_ifc_live.js, bonsai_sweep_live.js, bonsai_fillet_live.js, bonsai_move_live.js). See ModellerKernelFold.md.

The landscape — nobody else compiles, nobody else connects¶

Tier 1 — Incumbents (geometry authoring)¶

Tool	Role
Autodesk Revit	Full BIM authoring. Industry standard.
ArchiCAD (Graphisoft)	Architectural BIM. Strong in EU/Asia.
Tekla Structures (Trimble)	Steel/concrete detailing, fabrication-grade.

They create IFC. They model geometry. They do not decompose it into a BOM recipe, compile from intent, or verify the round-trip.

Tier 2 — Visual newcomers¶

Tool	What it does
Snaptrude	Browser sketch-to-BIM
TestFit	AI generative site planning
Arkio	VR/AR collaborative design

Design exploration. No BOM, no compilation, no verification.

Tier 3 — Open source (IFC-native)¶

Tool	What it does
Bonsai/BlenderBIM	IFC-native authoring inside Blender
IfcOpenShell	IFC parsing/generation library
ThatOpen (IFC.js)	Web IFC viewer/editor

They parse and display IFC. They do not abstract intent, compile from recipes, or prove round-trip fidelity — and none connect BIM to a transaction ERP over one signed log.

Moats¶

Spatial compilation is solved — and hard to replicate. Intent extraction, recompile, and a 0.002 mm round-trip across 21 buildings from 9 tools. Years of domain work.
BIM↔ERP over one signed op-log — unique. Building → procurement order, ERP documents reproduced to the cent. Requires rare BIM and manufacturing-ERP knowledge in one head.
One generic dashboard for every data model. AD-field-driven group-by/pivot/timeline — not a bespoke report per table.
Serverless / local-first by construction. Each browser is its own server; the only shared resource is a stateless signature gatekeeper. Scales to 10,000 POS stations with no central database to overload — the exact failure mode that sank the national e-invoicing rollout.
Domain-agnostic pipeline. Houses, terminals, bridges, rail (93% BOM compression) — one pipeline, a YAML mapping per domain. See INFRA_DESIGNER_SRS.md.
Op-log = git-for-data. Every state is a deterministic, reversible fold of a signed log — what makes What-If branches, audit, and crash-replay fall out for free.

The asymmetry: adding a GUI to a compilation foundation takes weeks. Adding spatial compilation — or a to-the-cent ERP fold — to a GUI-first tool takes years.

Who uses this, and what they'd pay for¶

Role	Workflow	Value	Tier
Quantity Surveyor	Drop IFC → BOQ + Variation Order Excel	Automated takeoff, no Navisworks	1
Contractor (tender)	Import architect's IFC → classify → costed BOM	Quantities tied to verified geometry	1
Project Manager	Push building → ERP project; track Planned vs Committed + EVM	Budget/actual + cost What-If in one place	1
Developer / Investor	Share a URL → browse the model, no install	Instant stakeholder view	1
SME merchant (Malaysia)	Run a local POS that does e-invoicing + accounting	Mandate compliance without a server to crash	2 (target)
ERP owner (migration)	Run a DB health-check → dirty-data + migration-gap report	Knows what they're sitting on before committing	2 (target)
Architect / small practice	Author basic geometry in-browser, export IFC	A path off per-seat license fees	3 (frontier)

Get involved¶

The project is open source (MIT) and actively developed. Roadmap: ACTION_ROADMAP.md. For the journey from the IfcOpenShell Federation branch (Oct 2025) to today, see PROJECT_CHRONOLOGY.md.

If you work with IFC models, run an ERP you're afraid to migrate, or just want verified spatial compilation — try it, break it, tell us what's missing. Contributions welcome: product catalogs, jurisdiction rules, format importers, test buildings.

Cross-references: SPATIAL_COMPILATION_PAPER.md — academic paper (0.002 mm proof), MigrateComparisonPaper.md — ERP fold, to the cent, BOMBasedCompilation.md — compilation pipeline spec, DATA_MODEL.md — 4-database schema, TestArchitecture.md — Rosetta Stone gates and traceability, ModellerKernelFold.md — modeller as signed-log fold, PROJECT_CHRONOLOGY.md — dated history + commit ledger, ACTION_ROADMAP.md — project roadmap

References¶

[1] Muller, M.F. et al. "On BIM Interoperability via the IFC Standard: An Assessment from the Structural Engineering and Design Viewpoint." Applied Sciences 11(23), 2021. — Documents geometry loss and property loss across IFC exchanges between Revit, ArchiCAD, Tekla, and others. doi:10.3390/app112311430

[2] Pazlar, T. & Turk, Z. "Interoperability in practice: Geometric data exchange using the IFC standard." ITcon 13, 2008. — Early benchmark showing "distortion or loss of information related to the geometry of the elements" and "incorrect connection between elements" across five IFC-certified tools. ResearchGate

[3] Diakite, A. & Zlatanova, S. "About the Geo-referencing of BIM models." TU Delft, 2018. — Analysis of coordinate system divergence in IFC georeferencing, origin offset problems, and IfcMapConversion limitations. PDF

[4] BIMcollab. "Coordinating IFC Models with World Coordinate System information." — Documents how models without IfcMapConversion "will be shown somewhere far away from the already loaded model." BIMcollab Help

[5] Autodesk. "Revit 2024: Enhancements to IFC Geometric Fidelity." 2023. — Autodesk's own acknowledgement that IFC geometric fidelity required improvement, with fixes for "complex families (parametric railings, helical stairs) which may generate fragmented geometries." Autodesk Blog

[6] CAD Interop. "Revit File Formats: BIM Interoperability, IFC Conversion." — Notes that .rvt files are "editable only in Revit" with "a shelf-life of 3 years (the lifespan of Autodesk support)." CAD Interop

[7] Moult, D. "How to create better IFC files with Revit." thinkmoult.com. — Documents that "even when you manage to export your geometry through IFC, there's always loss of data" and "importing that into Revit makes it utterly useless." thinkmoult

[8] Bonsai BIM. "Beautiful, detailed, and data-rich OpenBIM." — Bonsai is IFC-native: "you're not creating geometry that gets converted to IFC later. You're working directly in IFC." bonsaibim.org

[9] OSArch Community. "How to import IFC with large coordinates?" — Documents floating-point precision limits with georeferenced files requiring local origin offsets, and that "horizontal construction where distances frequently exceed 1km presents challenges." OSArch

[10] Oon, R.D. "BIM OOTB — Browser Variation Order from IFC Import." 2026. — Demonstrates what becomes possible when BIM data lives in a queryable DB rather than a file: geometry stored as hash-keyed BLOBs (identical meshes instanced, not duplicated), revision diff as SQL EXCEPT on GUID sets, cost impact as GROUP BY on diff × rates template. IFC import, 4D/5D variance, and costed Variation Order Excel — entirely in the browser, no server. YouTube

[11] Autodesk. "Industrialized Construction." — "Applies the discipline and systematized fabrication process of manufacturing to the design and build process... as consistent and replicable as widgets rolling off a factory assembly line." Autodesk Emerging Tech

[12] Olanrewaju, O.I. et al. "Quantifying the influence of BIM adoption." Automation in Construction 161, 2024. — Notes "a significant gap between research and industry practice" and that the industry "still lacks its own quantification methodology for BIM benefits." ScienceDirect