One method, end to end

A bullet land, a breech face, a toolmark — each is a 3-D height field z(x, y), measured as an X3P topography scan. Verity never looks at a photograph; it works on the geometry itself, so lighting and color are irrelevant. Drag to rotate the real scan.

The overall form (the curve of the barrel) and broad wavinessare shared by many sources — they hide the signal. An ISO 16610 band-pass removes them and keeps the fine roughness band that the specific tool left behind.

The striae run in one direction. Verity orients them, collapses along them, and crops the high, common groove shoulders — leaving a 1-D striation signature, the object we actually compare. (Impressed marks keep a 2-D areal map instead; the rest of the pipeline is identical.)

Slide windows of one signature against the other. Genuine matches make many windows lock onto one common shift; coincidences scatter. The windows that agree are the congruent matching regions — a generalization of the consecutive-matching-striae and CMC ideas to any mark. Their count is the score.

Here many regions link up with near-parallel connectors — a consistent, real correspondence.

A score is not evidence until it is calibrated. Verity compares this score against a named reference of known same-source and different-source pairs, and maps it to a likelihood ratio — with a characterized cost (Cllr) and a bound on how strong a claim the data can support. The decision stays a transparent, monotone transform of the score.

The output is a number an examiner and a court can audit: the likelihood ratio, its verbal weight, the exact regions that drove it, and an explicit statement of scope. Because the calibration is a bounded, monotone map, the report is interpretable regardless of how the score was computed — the firewall against the black box.

And it ships a reproducible recipe — every step, parameter, and input hash — under a content handle, so the result can be verified by re-running it. See the glass-box API →