# Design principles learned from the supplied NYC geometry

This is a working reference for future Rhino, Grasshopper, and Blender projects. Read [the findings](FINDINGS.md) and [measurement limitations](METHODOLOGY.md) before treating a number as a design constraint. Observed shapes are evidence of geometry; they do not reveal the designer's intent, building use, structural system, or zoning rationale.

## 1. Start with the urban hierarchy, not a field of towers

**Measured:** the files separate building geometry from lot lines, roadbeds, sidewalks, parks/open space, rail/subway features, and district context. The actual paths vary between files.

**Proposed design rule:** generate or preserve street/block structure first, subdivide plots where the brief needs it, then place building envelopes and roof zones. Keep public space as an explicit output. The square-grid city built earlier is a useful control example, not a learned representation of NYC.

Do not equate a roof polygon with a building, a building-footprint curve with a tax parcel, or a district-wide bounding box with buildable land. Repeated context layers must not inflate regional density estimates.

Layer labels are also fallible: the QN01 follow-up found upright, facade-like curves in its nominal footprint layer, and QN11 has Breps in that role. Check geometric type and orientation after semantic classification. See [diagnostic evidence](DIAGNOSTICS.md).

## 2. Treat scale, location, and terrain as separate parameters

**Measured:** the collection mixes declared feet and millimeters; source settings and large coordinate magnitudes are preserved in each report.

**Proposed design rule:** normalize each file into an agreed physical unit system before geometric comparisons. Use one reversible source-to-local transform for a crop and record its inverse. Preserve terrain elevation independently from modeled height. A base at elevation 30 m with a roof at 80 m gives a 50 m height, not an 80 m building.

Never move districts independently to arbitrary origins and then attempt to combine them without their transforms. In Blender, store the geographic origin separately and keep local mesh coordinates near zero. In Rhino, confirm units before importing rather than changing the unit label without scaling geometry.

If the application's import operation already converts units, verify a known dimension and do not apply the same scale a second time.

## 3. Measure footprints in their own orientation

**Measured example:** in `NYC_3DModel_MN05.3dm`, footprint GUID `e51db0a1-cbe3-47f3-8597-b58282977f0c` has area about 7,943.28 m². It fills only 0.447 of its world-axis bounding rectangle, but about 0.915 of its longest-edge-aligned rectangle. Its aligned dimensions are about 148.98 × 58.27 m. These are calculated geometric extents, not surveyed frontage/depth.

**Inference:** a low world-axis fill ratio can be caused mostly by rotation. It is not sufficient evidence for a courtyard, a porous building, or a large setback.

**Proposed design rule:** use block or footprint orientation to define local planes before measuring aspect ratio, setbacks, or facade directions. A longest-edge alignment is a cheap baseline; use a true minimum-area rectangle or street frontage association if orientation affects acceptance criteria.

## 4. Learn distributions locally, and preserve relationships

**Measured:** footprint area, perimeter, complexity proxies, and associated roof heights vary by file. The district reports preserve their distributions instead of supplying one citywide constant.

**Proposed design rule:** choose a context or geometric family before sampling. Preserve relationships between area, aspect ratio, local orientation, and height rather than drawing each independently. Use robust quantiles for typical massing and inspect exceptional examples separately.

Small/elongated, broad/compact, and large/irregular are defensible geometric descriptions. Calling those buildings homes, warehouses, or offices requires outside evidence. Longest-edge-aligned fill is still not parcel coverage, and footprint area is not gross floor area.

## 5. Use skyline structure as a hypothesis, not a universal radial formula

**Measured:** this study reports height proxies only where roof patches pass a conservative spatial association. Coverage and exclusions are explicit.

**Proposed design rule:** use the existing radial-attractor Grasshopper city as a procedural baseline. Replace its single falloff with selected context measurements only after reviewing actual examples. Allow multiple clusters, low districts, isolated landmarks, and height discontinuities when the brief supports them.

Do not fit a rule to a selected tallest object and call it a district law. Missing or incorrectly matched roofs bias height statistics. Matched entries alone are not a representative sample of every building.

## 6. Represent a building as an envelope with roof zones

**Measured:** the source building envelope is distributed across footprint, rooftop, and facade curves/surfaces. A rooftop surface can be one small patch; the matching evidence records how many patches contribute to a footprint proxy.

**Proposed design rule:** retain a hierarchy such as `building candidate -> base outline -> massing zones -> roof zones -> optional facade detail`. Support changes in plan and elevation rather than forcing every outline into a single extrusion with one decorative cap.

Roof fragmentation is not itself a setback count: adjacent coplanar patches, triangulation, and authoring seams can create multiple surfaces without separate architectural tiers. Derive tiers from connected geometry and compatible elevations, then inspect sections. Do not invent the original Grasshopper or CAD construction history.

## 7. Ground each candidate before deriving height

**Measured:** a footprint's base elevation is retained, and a matched roof height is calculated relative to that base. Whole-file Z spans include terrain, infrastructure, and sometimes city context.

**Proposed design rule:** terrain fit and local base choice precede extrusion. For a sloped base, explicitly choose whether the massing sits on a stepped foundation, follows terrain, or uses a reference plane. Do not flatten all bases merely to make a diagram convenient.

In a diorama, deliberate flattening is acceptable as a new design choice if it is recorded as a departure from source geometry.

## 8. Keep negative space intentional

**Measured:** base linework distinguishes streets, lot boundaries, parks, pavement edges, and other context. Those curves are not automatically a clean connected planar subdivision.

**Proposed design rule:** preserve the difference between streets, sidewalks, planted space, and leftover gaps. A uniform building-footprint ratio produces a uniform street-like gap; it does not reproduce a road network or a neighborhood's public realm.

Before deriving widths, pair compatible boundaries and measure local cross-sections. Before deriving area coverage, assemble correct polygons, holes, and district boundaries. Open linework cannot be treated as closed usable land.

## 9. Separate massing evidence from invented facade detail

**Measured:** this audit distinguishes geometry types and layer roles. It does not establish window grids, floor levels, facade materials, occupancy, or facade assemblies from massing surface names.

**Proposed design rule:** preserve source massing as one layer of truth. Any generated glazing, mullions, floor bands, balconies, or materials should be a clearly labeled visualization layer driven by an explicit brief. The earlier city's floor bands are an aesthetic parameter, not data recovered from NYC.

For a close-up, obtain appropriate reference geometry or photographs before claiming architectural fidelity. Geometry alone cannot prove a building's use or construction method.

## 10. Match detail to the output scale

**Proposed design rule:** use neighborhood outlines and simplified envelopes for a city-scale overview; retain characteristic roof transitions for district views; add verified architectural details only for close studies. Keep recognizable exceptions as separate assets when warranted.

Store a loss budget for simplification: footprint area change, maximum boundary deviation, silhouette change, and preserved height. A successful decimation or export is not sufficient evidence that its important features survived. Source model detail is itself uneven; see [publisher provenance](SOURCES.md).

## 11. Distinguish open geometry from invalid geometry

**Sampled:** the health reports separately record `IsValid`, closed-solid status, and naked-edge counts. A valid individual facade patch can be open by design. Some sampled objects genuinely fail Rhino's validity checks; their GUIDs and diagnostic logs are retained.

**Proposed design rule:** for massing visualization, retain valid open envelope patches when appropriate. For fabrication or solid Boolean work, isolate a candidate building, establish correspondence, join at a justified tolerance, and check orientation, closure, intersections, and volume. Do not automatically fill every naked edge or increase tolerance until a join succeeds.

Keep the original geometry alongside any repair, and compare measured dimensions before and after. Metadata errors, geometry errors, and intentional surface decomposition are three different issues.

Inspect the scale of each failure. The Queens parks/airports diagnostics include small physical edge/trim deviations against much smaller stored tolerances; ratios near 304.8 suggest, but do not establish, conversion-related bookkeeping. Conversion must preserve tolerance semantics as well as shape. A native validity failure, its visible impact, and its suitability for the next operation require separate judgments.

## 12. Treat identifiers as provenance, not semantic guarantees

**Measured:** source object GUIDs and sampled user strings are retained; there is no verified collection-wide building-ID linkage in this study. Context GUID signatures can show repeated identifiers without proving geometric equivalence of every object.

**Proposed design rule:** preserve source filename, original GUID, layer path, unit conversion, crop transform, and association method in derived assets. Never call a geometric match a source-record join. Do not assume different GUIDs imply different buildings or identical GUIDs prove identical geometry across revisions.

## 13. Validate the generator against examples, not its own appearance

**Proposed acceptance criteria:**

1. A known source crop retains its orientation, physical scale, and measured envelope after import and coordinate normalization.
2. A representative footprint reconstructs within a stated area/boundary tolerance; a simple rotated rectangle is a known-answer test.
3. Representative high and low envelopes retain base-relative heights and visible roof transitions.
4. Parametric inputs affect the intended output, preserve branch correspondence, and do not accumulate baked duplicates.
5. Native and exported deliverables are reopened and compared. Geometry validity, topology requirements, and visual fidelity are evaluated separately.

No new NYC-derived procedural city has been implemented or validated by this documentation study. These rules are a grounded brief for that next task, not a claim that the previous toy city now reproduces NYC.
