Evidence buyers should request from factory evaluation: part 3
The records a buyer requests from a factory evaluation decide what the visit can prove. Foundation documents first, then process-specific logs, each tested by tracing random entries to supporting proof. Single samples give bounded observations rather than verdicts, and corrective actions count only when chained from finding to verified closure. This third part of the evidence series turns the document request from a formality into a working test of the supplier's system.
Key takeaways
- Demand the six foundation documents first, then build outward to process-specific records.
- Test every record by tracing random entries to their supporting proof two hops deep.
- Read single-sample findings as bounded observations and look for patterns across wider pulls.
- Structure corrective-action evidence as a numbered chain from finding to verified closure.
- Judge documentation against quality management principles that link policy to retained evidence. [2]
Factory evaluations produce paper, and buyers must know which paper matters. A thick report full of certificates can still hide a plant that keeps no working records, while a thin file of well-chosen logs can prove a disciplined operation. This third part of the evidence series lists the records buyers should request, explains how to test them, and shows how sampling logic shapes what any record pull actually proves.
The guiding principle is traceability. Every claim the factory makes about its system should resolve into a record, and every record should resolve into observable practice on the floor. A training plan means little until attendance sheets name the operators; attendance sheets mean little until the operators met can explain the procedure. Buyers who follow these chains need no special auditing skill, only persistence.
Quality management thinking underpins the whole exercise. The vocabulary and principles of documented quality systems, from policy down to retained evidence, are codified in the ISO 9000 family of references on quality management foundations. [2] An evaluation tests how much of that structure the factory actually lives by, and the evidence buyers request is the test material.
The six documents that anchor every request
Begin with legitimacy and system bones: the current business license showing approved scope, the quality manual or equivalent document set, the most recent internal audit report with its findings, the management review minutes that disposition those findings, the measuring-equipment list with calibration status, and the corrective-action log. These six reveal whether a managed system exists at all, before anyone discusses product-specific controls.
The quality management framing behind this request is well established: documented policies, assigned responsibilities, and retained evidence of operation form the backbone described in references on quality management foundations. [2] Factories that live by such a structure produce these six documents in minutes; factories that perform compliance only before visits produce apologies. The speed of retrieval is itself evidence.
Where buyers want the request handled by an independent party, the document list is commonly built into UTS supplier verification UTS supplier verification assignments, so the auditor arrives with the same checklist the buyer approved. That alignment prevents the familiar failure where the buyer expected training records and the auditor spent the day photographing certificates instead.
Trace records two hops deep
An index proves nothing until its entries resolve into practice. Take the training plan, pick three names at random, and ask for their attendance sheets; then find one of those operators on the floor and ask them to explain the procedure they were trained on. Take the calibration list, pick two gauges at random, and check the physical labels; then ask for the last calibration certificate behind one label. Each chain has two hops, and genuine systems survive all of them.
Watch for the failure modes that tracing exposes. Attendance sheets signed in one pen on one date suggest a batch-signing session rather than real training. Calibration labels dated after the certificate's expiry suggest stickers without testing. Rejection logs with no linked supplier complaints or segregated stock suggest rejections recorded to fill a quota. None of these require laboratory analysis to spot; they require only the discipline to follow the chain.
Sampling logic governs how much tracing is enough. A handful of random traces bounds the risk that the records are decorative, and the mathematics of inferring population quality from samples is the subject matter of acceptance sampling references. [1] Buyers need not compute operating curves; they need only understand that five clean traces beat fifty untested index pages, and that findings from small pulls should be read as signals for wider checking rather than final verdicts.
Evidence request matrix
| Record group | What to ask for | How to test it |
|---|---|---|
| Legitimacy | License, approved scope, lease proof | Match scope to actual products on floor |
| System documents | Manual, procedures, revision history | Check a recent change propagated to stations |
| People | Training plan, attendance, competence checks | Interview a trained operator on content |
| Equipment | Gauge list, calibration labels, certificates | Match two labels to certificates on the spot |
| Materials | Supplier list, receiving checks, rejections | Trace a rejection to its disposition |
| Production | Work instructions, parameter logs, change notes | Compare a log line to machine settings |
| Corrective action | Finding log, root-cause notes, closure proof | Re-check one closed item for persistence |
Read single samples with proportion
Suppose the auditor pulls twenty training entries and three lack signatures. That observation describes those twenty entries, not the entire workforce, and the correct response is a bounded statement plus wider checking: pull more entries, interview the affected operators, and determine whether the gap is a dated lapse or an ongoing habit. Over-claiming from small pulls is the most common way buyers misread evaluation evidence.
Patterns across pulls carry the real meaning. The same unsigned-training gap in three departments over six months signals a system that does not enforce its own procedure, while a single cluster dated to one supervisor's leave month signals a lapse with a correctable cause. The inferential reasoning that connects sample results to population judgments is developed in standard treatments of acceptance sampling. [1] Buyers who internalize that reasoning ask for the next pull instead of issuing the final verdict.
Write findings with their boundaries attached: which records, which dates, which scope. Bounded findings are actionable because the factory knows exactly what to fix; sweeping verdicts invite argument because nobody knows what evidence would reverse them. The best evaluation reports read as a set of testable statements, each with its proof attached.
Demand chained corrective-action evidence
A corrective action is a chain with five links: the numbered finding, the agreed correction, the named owner, the evidence that the correction happened, and the verification that it worked. Buyers should request all five for every significant finding and reject chains with missing links. A new procedure with no training records behind it is an intention; training with no follow-up check is activity without proof.
Quality management principles explain why the chain must be complete: corrections that do not address causes recur, and recurrence is itself auditable in the next cycle's finding log. Foundational references on quality management describe this closed-loop expectation. [2] Repeat evaluations should therefore re-sample previously closed items, because fixes that held prove the system learns while fixes that evaporated prove the opposite.
Set a simple rule for closure disputes: the factory proposes the evidence, the buyer or auditor judges whether it demonstrates the correction, and the judgment is recorded against the finding number. This keeps discussion concrete and prevents the slow drift where findings stay open for months while both sides exchange general assurances. Closed means evidenced, and evidenced means traceable.
Request few documents, test each one deeply, and write every conclusion with its boundaries attached. That discipline turns evaluation paper from decoration into proof.
Sources
- Wikipedia: Acceptance sampling —— Overview of acceptance sampling including AQL concepts used in product inspection.
- Wikipedia: ISO 9000 —— Overview of the ISO 9000 family of quality management standards.
Frequently asked questions
Which six documents should every buyer request first?
At minimum the current business license, the quality manual or its equivalent, the last internal audit and management review records, calibration lists with labels, training records for key posts, and the corrective-action log. Missing any of these suggests the system exists on paper only. Everything else in this article builds on that foundation.
How can buyers test whether records are genuine?
Pick entries at random and trace each one to its supporting proof: a training line to an attendance sheet, a calibration line to a labeled gauge, a rejection line to a segregated lot or a supplier complaint. Fabricated indexes usually collapse at the second hop because the supporting layer was never built. Real systems survive random tracing.
What should buyers conclude from a single failed record sample?
Record the finding as a bounded observation with its date range and affected scope, then check whether the same gap appears in other samples. One dated cluster often traces to a correctable lapse, while the same gap across months signals a systemic absence. Broader inference from samples follows the standard statistical reasoning of acceptance sampling.
How should follow-up visits verify closed corrective actions?
Ask for the finding number, the agreed correction, the owner, the evidence of completion, and the verification result, in that order. Anything shorter than that chain is a promise rather than a closure. Repeat evaluations should re-check a sample of closed items to confirm the fixes held.