# Diagnostic findings that change how these models should be used

These are follow-up checks against actual geometry and native diagnostic logs, not just observations of filenames or layer labels. Original source files were not repaired or changed.

## QN01: the nominal footprint layer contains upright outlines

**Complete count:** `NYC_3DModel_QN01.3dm` contains 209,798 nonhorizontal curves in the layer classified as building footprints. The source declares feet. The main extractor excludes these from its horizontal-footprint metrics; it retains 24,685 eligible footprint curves in that file.

**Independent follow-up sample:** a separate deterministic reservoir of 64 nonhorizontal curves was checked using both XY shoelace area and native planar 3D area. All 64 had projected XY area below 1% of their native planar area. These are facade-like upright outlines, not useful horizontal footprints. This finding supports the geometric classification; it does not recover why the source objects were assigned that layer or establish that every excluded curve has the same purpose.

One concrete example is GUID `1ad8c78b-9ba2-4f49-a436-29bc86c30cd4`: its closed polyline uses two distinct XY positions at two elevations, spans about 3.866 m vertically, has zero XY area, and has about 14.359 m² of native planar area. That is direct evidence of an upright rectangle rather than a terrain-following plan outline.

[Follow-up measurements and original vertices](evidence/QN01-mixed-footprint-layer-check.json).

**Reusable rule:** normalize layer names, then check geometry type, closure, plane/orientation, and dimensional meaning. A layer called `Building_FootPrint` is not a sufficient extraction contract. Flattening these outlines into XY would create degenerate footprints and inflate counts.

## QN11: the footprint role also contains Breps

The `Buildings::Building_FootPrint` layer in `NYC_3DModel_QN11.3dm` contains 39,073 PolylineCurves and 14 Breps. The curve-metrics pipeline does not convert those Breps into additional footprint curves. The full layer census preserves them. This is why role totals, eligible curves, and verified building counts must remain separate.

## Queens parks/airports: tiny deviations fail very small stored tolerances

Rhino reports invalid sampled Breps in `NYC_3DModel_QN_Parks_LGA_JFK.3dm`, whose declared unit is millimeters. Several diagnostics concern the mismatch between a trim endpoint and its corresponding 3D edge.

For example, GUID `4389ebe6-49c7-4fea-a407-59ad113a86ca` reports a deviation of about 0.15773 source units against a stored edge tolerance of 0.000517487. At this file's declared scale, that is about 0.158 mm of deviation: a real native validity failure, but a different practical issue from a visibly missing wall or a gross survey-position error.

**Measured numerical pattern:** in 14 retained diagnostic logs across the footprint, roof, and facade surface samples, deviation divided by stored edge tolerance ranges from approximately 304.798876 to 304.800568, with median 304.800091. The 304.8 feet-to-millimeters factor is a strong conversion-related clue.

**Inference, not established history:** these ratios are consistent with conversion-related tolerance bookkeeping. They do not prove which software step caused the problem or that every invalid surface has the same cause. The original conversion pipeline was not reconstructed, and no source tolerance was altered.

[Ratios and traceable GUIDs](evidence/QN-parks-tolerance-ratio-check.json). The original messages remain in the file-specific report and raw JSON.

**Reusable rule:** when geometry has been rescaled, inspect tolerance-related geometry metadata as well as displayed units and dimensions. Evaluate both physical deviation and the requirements of the next operation. Do not claim that an invalid patch is a closed solid, and do not equate every native validity failure with unusable visualization geometry. Repair copied candidates only when the target deliverable requires it, preserving measured shape and documenting the operation.

## Context linework: coincident consecutive points

Other sampled failures include polylines with coincident adjacent vertices. Examples are `cda46d52-687d-4d04-9747-f6751895598a` in BK01's water linework and `5f99fccd-e7e7-4c52-aac0-29b2cb3a4c6d` in BK04's open-space linework. Native `IsValidWithLog` identifies the offending vertex pair in each file report.

**Reusable rule:** diagnose the exact defect before choosing a repair. Removing duplicate consecutive vertices may be appropriate for a derived copy, but it is not a general solution for edge/trim mismatches, self-intersections, missing surfaces, or ambiguous building associations. No cleanup was applied in this study.

## Selection bias in the statistical reference

Eligibility filters remove geometry that cannot safely enter the particular planar-footprint calculation. Their exclusions are evidence, not disposable noise. QN01 contributes most of the nonhorizontal exclusions; that cannot be generalized into a citywide statement that a quarter of true building footprints are sloped.

The district parameter reference and plots describe eligible entries. Height distributions are further conditional on incomplete geometric association. Use representative source crops and candidate-level checks before treating any range as a procedural building-height prior.
