Verification — 1 November 2026

Verified locally

Apple Silicon, Python 3.13.18, PyTorch 2.15.0.
Metal GPU tensor operations and backward pass work outside the app's filesystem sandbox.
10 CPU tests passed: Gaussian PLY round trip; mesh/point PLY rejection; coordinate bounds; deterministic cell selection; cache hash and overwrite protections; aligned geometry target import; missing-teacher masking; quaternion sign invariance; query-chunk attention equivalence; object-level split integrity; and measurable loss reduction during training.
2 additional Metal tests passed: default 16³ student forward/backward; and the official Idol architecture's tiny randomly initialized denoiser/sampler, octree and Gaussian decoder. The tiny teacher uses the pinned upstream SDK with the real-valued RoPE / chunked attention adapter.

End-to-end synthetic smoke test

12 procedural ellipsoid image/splat pairs. No teacher-generated samples.
10 training objects, 2 held-out validation objects; seed 42.
Student: grid 8, width 64, 306,321 parameters, FP32.
100 AdamW updates on MPS, batch size 1, learning rate 0.001.
Held-out loss: 1.35705 at step 1 → 0.55299 at step 100.
Held-out occupancy IoU at step 100: 0.48258.
Timed training loop including periodic validation/checkpointing: 6.383 seconds. This tiny 8³ fixture test is not a runtime estimate for teacher inference or serious training.
Reloaded the saved checkpoint and generated a 40-Gaussian PLY for held-out fixture 007.
Inspected the four-view diagnostic PNG. It is a coarse, sparse synthetic result, not a realistic reconstruction.

Not verified / not completed

No pretrained Idol weights or gated encoder weights were downloaded.
No complete real-image Idol inference on MPS; tiny component success does not establish full-size memory feasibility.
No InstantMesh or Hi3D inference on Mac. Their outputs can only be imported into the student dataset.
No training on real teacher outputs, real-image evaluation, teacher-quality matching, 1-bit weights, or binary kernels.
No multiview rendering loss, automatic cross-teacher alignment, or teacher disagreement filtering.

At inspection time, only about 7.4 GB of disk was available before installing dependencies. The teacher downloader uses a conservative 16 GiB free-space guard for new downloads. The user still needs access to DINOv3 and FLUX.2-dev encoder repositories and actual object images/teacher data.

The files in examples/synthetic-smoke are included for reproducibility, not as pretrained production assets.