How PointCheck Full Analysis Works

We benchmark the shape search on a held-out test set of 793 reference photos spanning the full type library (July 2026):

Caveats: the test set uses reference-quality photos under corpus conditions; field photos with rulers, clutter, or unusual angles may perform below these numbers. Many named types are near-identical regional variants — the shape-family number reflects how often the search lands in the right morphological neighborhood even when the exact name differs. Authenticity accuracy (measured Aug 2026, n=168 balanced benchmark): ~89% overall agreement at the primary decision boundary, with zero modern reproductions reaching the top confidence tier. Every report shows its evidence on both sides so you can weigh it yourself.

Type identification runs against thousands of reference images of over 1,000 documented point types. Authenticity assessment is benchmarked against a balanced dataset of artifacts with verified provenance — museum specimens, survey finds, and professionally authenticated pieces — against documented modern reproductions from skilled knappers.

Authenticity probabilities are mapped to human-readable verdict labels. The boundaries are derived from our authenticity benchmark (168 artifacts: 84 with verified provenance, 84 documented modern reproductions) rather than chosen arbitrarily — on that benchmark, zero modern reproductions reach the “High confidence authentic” tier, about 2% of genuine artifacts read “Likely modern,” and overall agreement with ground truth is ~89% at the primary decision boundary. The system is deliberately tuned so that mistakes lean toward caution about authenticity rather than false reassurance.