One calibration target. Observed reconstructions.

The Helios Gaussian view is initialized from the 16,274 measured S3D target points and optimized with 71 corrected RGB1 side-orbit images; A000 is withheld for validation. Its photogrammetry option is a separate 72-view known-pose RGB1 silhouette visual hull with an image-derived texture atlas. It is experimental, not dense MVS, and does not fuse S3D geometry.

Each phone has a matched 5,000-step OpenSplat appearance result using 90 RGB training views and one held-out view, plus its final canonical textured Apple Object Capture mesh. The splats use the final mesh surface as a geometric seed but do not ingest the auxiliary-depth channel. The CAD metrics remain based on the measured Helios point cloud and the phone meshes—not the appearance splats.

The CAD uses 100,000 triangle-area-uniform samples with NumPy PCG64 seed 20260811. Each phone mesh uses 100,000 samples with seed 20260812. Before sampling, every phone is isolated with the same complete-target rectangular prism at a 5 mm CAD margin, retaining every disconnected surface inside it and applying no component filter. Distances use exact Euclidean nearest-neighbour lookup.

Helios contributes all 16,274 retained canonical measured points rather than an interpolated mesh surface. That representation difference is explicit so the measured-point and triangle-surface outputs are not treated as identical geometry products.

The untouched source sets contain 98, 100, and 101 HEICs for iPhone 12, 15, and 17 Pro Max respectively. Every frame verifies its model in make/model metadata, matches its package-manifest hash, and contains auxiliary depth, confidence, calibration, wide-to-depth transform, and point-cloud metadata.

All three phones use the same Apple PhotogrammetrySession settings: reduced detail, high feature sensitivity, sequential sample ordering, object masking off, bounding box ignored, and model + poses + bounds requested. The full HEIC containers—not JPEG conversions—are supplied to Object Capture.

The audit classifies the embedded channel as sensor-unclassified auxiliary depth because neither ImageIO nor the package identifies its sensor modality. Object Capture does not expose whether its final mesh used that channel, so this page does not relabel it as verified LiDAR or imply causal depth assistance.

The scored geometry is an Apple Object Capture photogrammetry mesh—not a raw LiDAR point cloud, raw ARKit scene mesh, or Gaussian splat.

All three phones use the same direct four-AprilTag multiview metric-rigid recovery. Maximum per-tag reprojection RMSE is 4.091 px for iPhone 12, 4.937 px for iPhone 15, and 5.320 px for iPhone 17; all pass the explicit 8 px publication gate. Rigid target-fit RMSE/p95 is 0.255/0.339 mm, 0.604/0.813 mm, and 1.528/1.856 mm respectively.

The solve accepts only target IDs 12, 13, 26, and 27. Other AprilTags in the room are irrelevant and never enter the correspondence set. The 10 px upstream filter is a separate whole-marker-frame, all-corner observation gate.

Native Object Capture metres remain fixed at 0.001 native units per CAD millimetre. No free scale is applied. The scorer performs no ICP, registration, similarity rescaling, or reflection. Helios retains its independently validated fixed-scale proper-rigid target registration.

The reference is the FULL_BORE_v1 corrected scan-visible CAD revision for PF-T31-00001059-C694BB02. Despite the source filename’s “as-built” wording, it is nominal corrected CAD—not a CMM measurement of this printed specimen.

Width and depth measure coherent physical platform side walls from each full raw reconstruction in the exact z=0–4 mm band, with fixed native scale and no CAD-centred normal-axis crop. Height is the local apex-cap centre above the AprilTag top-surface plane, not total envelope height.

Parameter, stratified-deletion, and half-sector sensitivity are reported as estimator confidence on one reconstruction—not independent-capture precision. Full bidirectional surface errors remain visible so skirts and extra shells are not hidden.

The finished PCD has no per-angle provenance. Removing points would test point decimation, not capture count. Each balanced image subset must be reconstructed independently in Helios.

The acceptance plan allows median worsening ≤0.5 mm, p95 worsening ≤1.0 mm, CAD-coverage drop ≤3 percentage points at 2 mm, additional height-bias magnitude ≤1.0 mm, and additional base width/depth bias ≤0.5 mm relative to 60 views, with all five fruits and four leaf modules supported.