Scanned/raster input — the 3A76 lessons (ratified by Aristide 2026-08-27)

kerfmaster pin list · Preflight — how files come in

rulingA decision by Aristide or Jordan. True because it was decided; it can be superseded, but it cannot be stale.

Content last changed 2026-08-27 — computed from the item itself, not typed.

Measured on 3A76-LAT-1X (builds 141 scan trace vs 142 customer DWG):

• Leader/arrowhead ink fuses with the outline. An arrowhead is a solid 6-10 px triangle; hair-shaving (r=2 opening) cannot remove it, so the trace grew three beak/cusp defects exactly where callout leaders touched the part edge (Jordan's markup found all three).
• Callout-to-feature association is a judgment call. The .316 and .283 callouts were the width and height of ONE rounded D-shaped boss hole; the trace read .316 as a circle diameter and hunted for a .283 web that never existed.
• Scan scale cannot be trusted, even proven from callouts. The trace came out 1.0-1.5% undersized ANISOTROPICALLY (1.54% along one axis, 1.03% across, axis ~67°) — no single scale factor can fix a stretched scan. Measured from true hole spacing in the customer DWG.
• The audit trap: the trace was 'verified' against a centerline rebuilt by the same pipeline — systematic errors appeared on both sides and measured as agreement (max 0.006"). One marked-up screenshot from Jordan beat the numeric audit. Rule: verify against the raw source or an independent reference, never a re-run of yourself.
• Always ask for native CAD first. The customer's DWG ran through the unmodified standard pipeline to PASS at 0.00 thou; the best possible trace was superseded the moment it arrived.