Make the most powerful thing in AEC feel like the simplest.

External design audit, 2026-07-25, read against the README, the CHANGELOG and the live demo shell. 18 findings · 5 principles · 6 interface moves · 4 phases. This file is the durable record; the actionable engineering lands in roadmap.md as the R24 ring.

The platform spans acquisition → turnover on one IFC model. The interface is currently organised the way the engine is built, rather than the way an architect, superintendent or developer actually spends a Tuesday. That gap — not missing capability — is what decides whether any of it gets used.

The consequence is specific. With ~130 modules shipped, the industry baseline says roughly ten are what any one person touches on a given day — and which ten depends entirely on who they are. A catalog with favourites and a filter treats that as a browsing problem. It is a routing problem.

The upside is equally specific: because every record, geometry and cost line is keyed to the same IFC GlobalId, the platform can answer "where did this number come from" in one hop. The interface does not currently cash that in.

The app already knows these roles; it does not yet act on them.

1. One model, one card. A wall, an RFI and a cost code are the same object seen from three sides. One card component renders it everywhere, with the same lifecycle strip. Learn it once.
2. The job, not the module. Nobody opens "Submittal Register" — they answer a thing that is late. Navigation starts from a queue and a command bar; the module is where you land, not where you start.
3. Never hide — explain. A missing button is a support ticket; a disabled button that says needs GC · Project Manager is a lesson.
4. Every number traces. Any figure on screen is one click from its chain: IRR ← NOI ← rent roll ← area ← this slab's GlobalId. This is the entire reason to be IFC-keyed. Make it visible.
5. Fast is a feature. Budgets, not hopes: 100 ms for a click echo, 1 s for a panel; anything longer becomes a named job in a tray you can walk away from. Never a blocking spinner.

1. One bar that does everything. ⌘K searches records, jumps to modules, runs authoring verbs, finds an element by GlobalId — grouped by verb / record / element / report, never a flat list. The existing AI ask box becomes the fallback row rather than a separate feature to find.
2. Replace the catalog with a spine. A persona-scoped rail of ~seven destinations with live ball-in-your-court counts. Nothing is deleted — the rest stay one keystroke away.
3. Home is a queue with a horizon. Work queue left, project health right, one banded verdict on top. Rows actionable inline, without opening the module.
4. The element card, everywhere. Six lifecycle states — designed · checked · priced · scheduled · installed · verified — rendered identically wherever the element is named.
5. Numbers that show their work. Every figure expands into the chain that produced it, tagged model-derived / overridden / market assumption, ending in a clickable element.
6. A field mode, not a small desktop. 56 px targets, 17 px body, 7:1 contrast, capture-first home, permanently visible sync queue. A separate mode with its own rules, not a responsive breakpoint.

Density scale — Field (56 px / 17 px / 7:1) · Default (36 px / 14 px / 4.5:1) · Compact (28 px / 13 px). One switch, per user, persisted. A superintendent and a scheduler should not be given the same row height.

• Colour, restrained on purpose. Blue = interactive & model-derived. Green = solved, healthy. Amber = your attention. Red = blocking. Nothing else is coloured, ever.
• Two faces, no exceptions. A sans for language; a mono for anything a machine produced — IDs, GlobalIds, quantities, currency, dates, statuses. The split is the fastest signal for "this is data, not prose."
• The grid is the brand. A 24 px minor / 192 px major grid behind every workspace at 5–7% opacity — the drawing sheet, the site plan and the spreadsheet at once. Borders, not elevation.

Two findings were partly addressed in v0.3.677 by an independent live audit — worth recording so the ring is not re-litigated:

• Nav density (#01/#02, partial). The Build stage held 13 entries of which 7 were project accounting, and Model & standards held 14. Split into Build / Money and Model & standards / Analyse & check; largest group 14 → 7, nothing removed. That is the shape the spine wants, applied to the existing rail rather than replacing it.
• Controls are labelled. 170 visible controls, 0 unlabelled, 0 console errors — so finding #03 is about permission legibility, not about missing labels.

And one finding the live audit independently reached from the other direction:

• #07 onboarding / #18 marketing → app. The Master Builder panel is already a live 8-step readiness synthesis that names exactly what is missing and links to the tool that fixes it. It is reachable from one destination in one workspace. Promoting it is the cheapest available progress on both findings.

Sources: AGC 2024 technology-adoption survey · Quickbase 2023 gray-work report · Pendo 2019 Feature Adoption Report (615 subscriptions) · Standish Group CHAOS 2002 · Nielsen Norman Group, progressive disclosure · CII IR-153 rework causation.