Changing a Test Method: Reclassify Borderline Parts Before Releasing the Change

Measurement purpose

Evaluate acceptance-decision changes caused by replacing a ceramic-circuit measurement method.

Specimens and conditions

Paired parts

Identified stable specimens including relevant boundary cases and ordinary range coverage

Requirement

One defined characteristic and explicit old/new method and decision-rule versions

Equipment and records required

• Comparison methods: Documented fixtures, settings and routine reading procedures
• Disagreement resolution: Independent recalculation or suitable reference measurement where needed

Method sequence

1. Design

   Separate method changes from rule changes and select the challenge population

   Record: Comparison plan

2. Pairing

   Record both values and decisions for each specimen

   Record: Paired classification table

3. Resolution

   Investigate reversals and validate the transition boundary

   Record: Method-change approval scope

Decision and uncertainty

Approve the changed method only after relevant decision reversals are understood and meet the agreed acceptance-risk requirement.

Boundary enrichment, dependent repeats and an imperfect old method limit interpretation of simple agreement percentages.

The measurement owner verifies method behaviour; the customer or designated quality authority approves decision consequences.

Traceable outputs

Separate a method change from a decision-rule change

Identify what is changing: instrument, fixture, contact method, software calculation, measurement condition or acceptance rule. Changing both the measured value and the rule at once makes a classification shift difficult to interpret. Record the old and proposed methods and, where practical, compare them under one common rule before evaluating any separately authorized rule change.

For a printed resistor, a different current or stabilization time can change the physical measurement state. For a dimension, a new datum construction can change the quantity being reported. The comparison must begin with the same defined characteristic. Two fields with the same label are not necessarily measurements of the same thing.

Include parts near the actual decision boundary

A collection of comfortably conforming parts is unlikely to reveal classification risk. Select specimens near the relevant upper and lower limits, together with ordinary parts and known out-of-range conditions where available. Preserve how these specimens were selected. A deliberately boundary-enriched set is useful for method validation but does not estimate the routine production reject rate.

Keep specimen identity and condition stable through the paired measurements. If a destructive or conditioning test changes the part, an appropriate comparison design is required instead of pretending that the second reading sees the same state. Include repeated placements or sessions when they represent ordinary measurement use, while retaining their relationship to each physical part.

Build a paired classification table

For each specimen, retain both measured values and the decision each method would produce under the stated rule. Count both directions of disagreement. A new method that passes old-method rejects presents a different review question from one that rejects old-method passes, but neither direction proves which method is correct.

In an illustrative hundred-part comparison, eighty parts pass both methods, ten fail both, six pass only the old method and four pass only the new method. Agreement is ninety percent. The ten disagreements remain the focus; reporting only ninety percent agreement hides their location and consequence.

Separate numerical shift from rule implementation

Recalculate selected decisions independently from the retained raw values. Check units, rounding, inequality direction, inclusive endpoints and the applicable uncertainty treatment. If the values agree but decisions differ, the problem may be in the rule implementation rather than the instrument. A software comparison that omits an equality condition can affect only a few boundary cases and remain invisible in averages.

If the measured values differ, examine the discrepancy against proximity to the limit, part geometry, contact position and measurement order. A small consistent offset can create many reversals when production lies close to a limit. A wide random spread can create reversals in both directions even when the mean offset is almost zero.

Investigate disagreement without declaring the old method correct

The previous method is a baseline, not automatically a reference truth. Use an independently justified measurement or technical investigation to resolve important discrepancies. Confirm specimen state, method capability and the requirement definition. Repeating the old method until it returns its earlier value is not independent confirmation.

Retain an unresolved category when the evidence cannot support a definitive disposition. Do not label every newly accepted part a false acceptance unless its nonconforming status has actually been established. Similarly, an additional rejection is not necessarily improved quality; it can be a measurement error or an unnecessarily restrictive rule.

Examine decision stability under routine repetition

Measure selected boundary parts under the normal removal, reseating and session changes expected in use. Record whether the same method alternates between pass and fail. That instability helps explain inter-method disagreement and may require a more suitable measurement or an agreed treatment near the boundary.

Repeated decisions are not new independent specimens. Keep the unit hierarchy in the analysis and avoid inflating the comparison by counting every repeat as another part. If the routine result uses an average of several readings, test that actual procedure. A validation based on extensive laboratory averaging does not establish the behaviour of a single rapid production reading.

Control the first production decisions after approval

Define the effective lot or date for the new method, identify affected reports and preserve the method version with each result. A changed instrument program carrying the old name can make later disputes impossible to reconstruct. Keep any approved correction and its validation linked to the version that uses it.

Where the change review calls for an overlap period, specify its purpose and duration in advance. Use it to challenge the accepted boundary behaviour, not to choose whichever result allows shipment. Disagreement handling should follow the approved escalation rule. The decision owner must be identifiable when the methods produce different classifications on a live lot.

Report numerical and classification conclusions separately

The final report includes paired values, agreement counts, disagreement direction, proximity to limits, repeatability observations and resolved causes. State the specimen-selection strategy so an enriched boundary study is not presented as an overall yield estimate. Describe exactly which characteristic and method configuration are approved.

A method may be suitable for one resistance range or geometry but not another. ChipSimple can review the drawing-defined verification interface by project; method transfer requires evidence appropriate to the intended decision. A matching mean is useful, but it cannot substitute for proof that the new method makes acceptable part-level decisions.