# Synthetic scan alignment study

This is a reproducible export of archived synthetic metrics, not a new experiment or registration run. Run `python generate.py` beside `trials.csv` to regenerate the 27 cell summaries, metadata and PNG/SVG/PDF figure. The CSV is the immutable numerical input. It contains 135 trials: three coordinate-noise levels × three random sample-deletion levels × three scale biases × five repeats.

The calculation aligned a known synthetic bracket with a perturbed version using rigid iterative closest point registration. The reference has 4,728 sample records at 1,184 unique coordinates. Duplicated positions retain their weights. Deletion acts on records, so other records can survive at the same nominal location. Coverage means the fraction of reference records with a nearby aligned record, not physical visible area or a scanner field of view.

The three illustrative gates are pooled bidirectional 95th-percentile distance ≤1 mm, symmetric mean distance ≤0.5 mm and reference sample coverage ≥70% within 0.75 mm. Exactly 89 of 135 archived trials meet all three. Five repeats per condition and deliberately chosen synthetic perturbations cannot predict real scanner reliability. The figure displays all trial outcomes and medians for two metrics; all three gates determine the count.

The symmetric mean weights the two directed means equally. The pooled percentile weights individual records equally, so changing the number of retained samples changes its directional weighting. Original distance quantiles use the NumPy default linear rule. Percentiles cannot be independently reconstructed from summary metrics alone.

No physical scan, printing, mesh generation, material duplication or manufacturing validation took place. A useful real comparison would require an independently measured reference, documented scanner/build conditions, and appropriate inspection of dimensions, surfaces and internal defects. Sources in study.json support measurement methodology, not the synthetic performance numbers.

Checks performed: complete 27-cell factorial; five numbered trials per cell; 135 exact metric records; bounds and reference-count checks; gate recomputation; median comparisons with the archived cell summaries. The generator reproduces export plots and statistics only, not the original point clouds. Exact private provenance is retained outside this public candidate directory.