A sample is not representative because it looks varied. It is representative only relative to a defined population, decision, selection rule and known limits on what the inspected units can support.
A buyer requests ten ‘random windows’. The factory presents ten accessible units from the same stack: all standard casements, all recently finished and none reworked. The report contains many photographs and a 100% pass rate for the sample, but sliding doors, special glass, changed hardware, packed units and the second production shift were never exposed to selection. The problem is not solved by inspecting more convenient units. The commissioning team must define the population, stratify material risks, choose whether the decision is census, risk-directed review or formal acceptance sampling, and keep the factory from silently substituting selected identities.
1. Define the decision and physical population before choosing a sample
Write the decision first: verify a first-off group, investigate a suspected process, review workmanship across risk groups, or apply a contractually specified lot-acceptance plan. These decisions use different selection logic. A risk audit may deliberately over-select unusual units; a statistical acceptance sample must follow its specified random route and cannot be adjusted after results appear.
Freeze the population register at selection time: order/project, opening marks and unique unit IDs, configuration, quantity, production location/line/shift where relevant, completion and inspection status, rework/NCR history, packed/access state and physical location. Reconcile it to actual production. Excluded, unfinished, already shipped, missing or inaccessible units remain visible in the denominator and report.
Define lot boundaries using homogeneous characteristics required by the governing plan—not merely one purchase order. A lot may need separation by system, process, material/finish batch, production period, shift or other controlled factor. Conversely, splitting an adverse population after inspection to dilute findings is not an acceptable lot definition.
| Population field | Controlled source | Physical reconciliation | Selection consequence |
|---|---|---|---|
| Unit/mark identity | Released schedule and production register | Unique units present/absent | Build complete frame |
| Configuration/risk attributes | Drawings, BOM, glass/hardware/finish schedules | Physical/record match | Stratify or lot definition |
| Process history | Line/shift/batch/first-off/change records | Traceable links | Risk coverage or homogeneity |
| Quality status | Inspection, NCR, rework and replacement records | Statuses verified | Do not hide adverse states |
| Access/completion | Production and packing registers | Inspectable/inaccessible/not complete | Report denominator and limitation |
| Decision scope | Contract quality plan/brief | Authority confirmed | Census/risk/formal sampling route |
2. Build risk strata that follow failure routes
Map characteristics that could create different failure or evidence routes: fixed, hinged, tilt-and-turn, sliding or door operation; dimensions and mass boundaries; handing; system/profile/reinforcement; glass build-up; finish; hardware family; drainage; screens/accessories; special labels or destination evidence; factory/line/shift; subcontract process; first-off, changed, repaired, replacement or NCR-linked status; and packing form.
Create a coverage matrix connecting every material stratum to at least one selection route where the approved plan requires it. Deliberately include high-consequence, boundary, changed and adverse-history units in a risk-directed inspection. Do not call that subset random or use its result to calculate a lot acceptance rate unless a separate valid plan supports the inference.
Control overlap. One selected large tilt-and-turn unit may cover several attributes, but the matrix should show which ones and which remain uncovered. Avoid token diversity: selecting one door does not represent all door systems, sizes, hardware and shifts. State residual gaps explicitly and decide whether to inspect, obtain upstream evidence or retain the unknown.
| Stratum/risk | Population count/source | Why different | Selection route | Selected IDs | Residual gap |
|---|---|---|---|---|---|
| Opening/system family | [register] | Different fabrication/operation | Census / risk / specified sample | [IDs] | [gap] |
| Size/mass boundary | [register] | Geometry/hardware/handling consequence | Deliberate boundary coverage | [IDs] | [gap] |
| Glass/finish/hardware variant | [BOM/schedules] | Material/process/evidence route | [route] | [IDs] | [gap] |
| Changed/reworked/NCR | [change/NCR log] | Elevated recurrence/collateral risk | Deliberate inclusion | [IDs] | [gap] |
| Line/shift/subcontract batch | [production records] | Different process history | [route] | [IDs] | [gap] |
| Packed/inaccessible | [packing register] | Evidence/access limitation | Earlier record/opening under authority | [IDs] | [gap] |
3. Keep risk-directed inspection separate from formal acceptance sampling
Use complete inspection when the governing plan requires every unit or characteristic to be checked. Use risk-directed selection to answer targeted coverage and investigation questions; document that it is purposive and makes no statistical representativeness claim. Use formal acceptance sampling only when the contract or quality plan defines the standard, lot, characteristic, defect classification, inspection level, AQL, sample-size route, random selection, switching rules and authority before selection.
ISO 2859-1:2026 provides AQL-indexed lot-by-lot attribute-sampling schemes. ISO 28590:2017 gives an introduction to the ISO 2859 attribute-sampling series. Neither tells a window buyer which AQL, defect classes or plan to use, and neither turns risk-picked units into a random sample. Variables plans under the ISO 3951 series involve additional assumptions and competent statistical design; do not use them merely because measurements produce numbers.
Do not convert sample findings into invented confidence claims, ‘quality percentages’ or predictions for the whole order. The observed sample rate is a description of selected inspected units. Any formal inference depends on a valid design, executed selection, measurement system and model assumptions. Keep the claim bounded when those conditions are absent.
| Route | Use | Required controls | Permitted conclusion | Prohibited shortcut |
|---|---|---|---|---|
| Complete inspection | Every defined unit/characteristic | Complete population, access, methods and traceability | Results for inspected full scope | Silence on inaccessible units |
| Risk-directed | Coverage of high-risk/boundary/adverse cases | Risk strata and purposive identities | Findings for selected scope and risk questions | Statistical lot acceptance claim |
| Formal attribute sampling | Specified lot acceptance decision | Named standard/plan, random route and decision rules | Plan-defined decision | Invented AQL/sample size after results |
| Investigation expansion | Contain suspected process/population | Causal trace and authorised expansion rule | Bounded containment evidence | Calling it random sampling |
| Upstream record review | Concealed/process evidence | Identity and scope links | Evidence for referenced population only | Treating records as physical inspection |
4. Protect selection integrity from register to physical unit
The commissioning or authorised selection party should generate selected IDs from the frozen population under the specified method. Preserve seed/tool/version or draw record where a formal random route requires reproducibility. Do not ask the factory to present ‘representative examples’ and then describe them as inspector-selected.
Trace every selected ID to its physical unit before inspection. Record missing, duplicate, substituted, inaccessible, unfinished or already packed selections. A substitute is a new selection event requiring the approved method and reason; it must not silently replace the difficult unit. Factory preselection, cleaning, adjustment or rework after selection should remain visible.
Maintain identity through photographs, results, NCRs, reinspection and report. Each result needs selected unit/opening mark, package or component where relevant, characteristic, criterion, method and evidence reference. Aggregated tables should be derivable from unit-level records, not manually reconstructed from unlabelled photographs.
| Event | Required evidence | Exception state | Control response |
|---|---|---|---|
| Population freeze | Timestamped reconciled register and eligibility | Unknown/missing identities | Hold selection or qualify population |
| Selection | Method, selector and exact IDs | Factory-selected/convenience units | Relabel route truthfully or redraw |
| Physical location | ID-to-unit match and condition | Unavailable/substituted/duplicate | Record and apply replacement rule |
| Inspection | Unit-characteristic-method-result link | Identity lost | Result invalid/indeterminate |
| Rework/reinspection | Original finding and same/replacement unit link | Clean photo overwrites failure | Preserve both states |
| Report | Population, selected IDs and exceptions reconcile | Denominator drift | Correct before decision |
5. Expand findings by traceable cause—not an improvised doubling rule
When a finding appears, identify the requirement, physical identity, process/material/revision links and potentially affected population. An authorised plan may require rejection, tightened inspection, full screening, investigation or another response. Do not invent ‘inspect twice as many’ or stop after the next few pass unless the governing plan defines that rule.
Separate local repair from population containment. Reinspect the corrected unit for the original and collateral characteristics, but also trace other units sharing the suspected program, fixture, material batch, shift, installer, change or packing event. Narrow the held population only with evidence.
Report the frozen population, exclusions, selection design, risk strata, selected IDs, substitutions/access failures, criteria/methods, unit results, findings, expansion, dispositions and residual gaps. A passed sample does not prove uninspected goods, concealed work, shipment readiness or destination conformity.
FAQ
How many windows should be inspected?
There is no universal number. The authorised route depends on the decision, lot/population, risk strata, governing plan and whether inspection is complete, purposive or formal acceptance sampling.
Does choosing many different window types make a sample random?
No. Diversity can improve risk coverage, but random selection requires a defined eligible population and controlled chance-based selection method.
Can an inspector pick the easiest accessible units?
That introduces selection bias and hides access limitations. Select from the frozen population and record every unavailable or substituted identity.
What AQL should a buyer use?
This guide supplies none. Defect classes, AQL, inspection level, lot and decision plan require contractual, technical and statistical authority before inspection.
Can the pass percentage of inspected units be called the factory quality rate?
No. It describes only the selected inspected scope and may be strongly biased. A broader claim needs a valid statistical design and assumptions.
Should sample size automatically double after a defect?
Not unless the governing approved plan says so. Use its switching/expansion rules or trace the connected risk population under authorised containment.
Can reworked units be excluded from the population?
Not silently. Their status is risk-relevant. Define eligibility and coverage before selection and report rework/NCR populations explicitly.
Continue the project review
Connect this guide with the proposed window and door systems, then review how Cavorlen approaches manufacturing and quality control. When your schedule is ready, send it through the project review form.
Official sources used
ISO — ISO 2859-1:2026 sampling procedures for inspection by attributes ↗ISO — ISO 28590:2017 introduction to the ISO 2859 attribute-sampling series ↗ISO — ISO 3951-1:2022 sampling procedures for inspection by variables ↗NIST/SEMATECH — e-Handbook of Statistical Methods: acceptance sampling ↗ISO — ISO/IEC 17020:2012 requirements for inspection bodies ↗Related buyer guides
This article is general inspection-selection and evidence-control education. It is not a statistical sampling plan, product standard, conformity assessment, certification, inspection-body accreditation decision, acceptance/rejection rule or shipment release. It supplies no AQL, inspection level, sample-size code, sample size, acceptance/rejection number, defect class, tolerance, confidence statement, randomisation algorithm, switching rule or inspection frequency and does not claim a selected sample proves uninspected goods or any Cavorlen quality, population, process or practice. The contract quality plan, governing standards, defined lot and decision, actual risk evidence, and named authorised quality/statistical/technical parties must establish and apply the real selection and acceptance route before results are known.