From CSV output to customer acceptance

Production vision is not finished when the tool finds the feature. It is finished when the customer's quality engineer, your applications engineer, and the operator on the line can all reconstruct what happened — which part, which pin, what was measured, against what nominal, under what tolerance, and what the system decided. Most acceptance disputes I have worked through were not measurement disputes. They were evidence disputes.

The minimum viable evidence package

• Barcode-linked CSV with measured, nominal, tolerance, and deviation fields per feature — one row per pin, one file per inspection, named so a human can find it six months later.
• Pass/fail flags separated from raw values. When the acceptance criteria change (and they will), historical data can be re-scored instead of re-measured.
• Defect descriptions, not just codes. "Pin R3-C14 radial deviation 138 µm vs 120 µm tolerance" ends an argument that "FAIL 0x0E" starts.
• Overlay bitmaps showing the ROI, the found feature, and the tolerance zone — the single most persuasive artifact in any customer review.
• Runtime diagnostics so instability (lighting drift, calibration mismatch, clipping) is visible in the data before it becomes visible in the yield.

Then make the data do work

A folder of CSVs is evidence. It is not yet intelligence. The next layer aggregates pin-level output into questions engineers actually ask:

• Is deviation drifting over time or between platforms? (drift timelines, source correlation)
• Is failure spatially clustered — a corner, a row, a fixture side? (heatmaps, pin maps with tolerance rings)
• Are we repeatable enough to trust the tolerance? (repeatability review, coplanarity and seating-plane analysis)
• Which pins are at risk before they fail? (per-pin risk tables)

This is exactly the gap the TopX Inspection Analyzer was built for: it watches inspection CSV folders offline, renders those views, and exports shareable interactive reports that mirror what the customer needs to sign off.

The loop closes at the RMA

When a returned board comes back with a field failure, the same barcode-linked trace lets you walk from the RMA to the original inspection record, the overlay, and the measurement history of that exact pin. That is when inspection stops being a gate and starts being production learning.

← Previous: weak-feature recovery · All notes · Next: 3D deployment →