Defect classification comes before the AQL number, not after

The outgoing inspection sheet for lot 24-0917 showed 13 cartons of 2 mL clear glass vials with PTFE-lined caps. The AQL sampling plan called for 80 pieces from a 3,200-piece lot at inspection level II, AQL 1.0 for major defects. The inspector found two vials with visible chips at the neck finish. Two out of 80 is 2.5%. The acceptance number for that plan is 3, so the lot shipped. Four weeks later, a customer running USP <621> chromatography on a stability study reported three autosampler crashes in a single sequence, all traced to neck chips that caught the gripper. The sampling plan was correct. The defect classification was wrong. A chipped neck that stops a sequence is not a major defect. It is a critical one, and critical defects get AQL 0. That gap between a valid AQL number and a valid defect class is where most outgoing QC programs fail.

This is what I want to walk through: how to build an AQL sampling plan for lab chromatography consumables that survives contact with a real customer complaint, and what the outgoing inspection record should actually contain so that when a lot is challenged, you can reconstruct the decision. If you buy or manufacture vials, caps, septa, filters, or columns, the sampling plan is the last gate before the product leaves. Get the defect classes right and the AQL levels follow. Get them wrong and you ship a lot that meets a number and fails a method.

Every AQL table in every quality manual starts with the same implicit assumption: you already know which defects are critical, which are major, and which are minor. Most chromatography suppliers inherit a classification scheme from a general-purpose inspection procedure written for machined parts, then bolt an AQL level onto it. That is backwards.

For lab chromatography consumables, I classify defects by what they do to the analytical result or to the instrument, not by how bad they look. Three classes, defined by consequence:

• Critical (AQL 0, zero acceptance): a defect that causes a sequence failure, an instrument fault, or a reportable result that is wrong. Chipped or cracked vial neck finish. A septum that does not seal at the specified torque. A filter housing with a leak path. A column with a void at the inlet that produces a split peak on the first injection. A cap with cross-threading that will not torque to spec. These are not negotiable at any AQL above zero.
• Major (AQL 0.65 to 1.5, depending on product): a defect that degrades performance measurably but does not stop the run. Extractables above the certificate limit. Dimensional out-of-tolerance on a septum thickness that changes the seal geometry. Particulate count above the release limit in a filter. A graduation mark out of position on a volumetric item. These cost the customer time and rework but the method still runs.
• Minor (AQL 2.5 to 4.0): cosmetic or handling defects with no analytical consequence. A scuff on the vial body. A label misaligned by 1 mm. A carton print defect. These matter to the customer’s receiving dock but not to the chromatogram.

The classification has to be written down before the sampling plan is built, and it has to be written in terms of a measurable or observable condition. “Neck chip” is not a classification. “Any visible chip, crack, or crazing within 3 mm of the neck finish, verified under 10x magnification” is a classification. The inspector can apply that. The auditor can audit it.

Why AQL 0 for critical defects is not the same as 100% inspection

People conflate these constantly. AQL 0 with a sampling plan at inspection level II means you accept zero critical defects in the sample. It does not mean you inspect every piece. It means the sampling plan is designed so that a lot with a critical defect rate above the implied quality level has a high probability of rejection. For a lot of 3,200 at level II, the sample size is 125 (per ANSI/ASQ Z1.4, which is the U.S. mirror of ISO 2859-1). At AQL 0, the acceptance number is 0 and the rejection number is 1. That plan gives you roughly a 95% probability of rejecting a lot with a 2.4% critical defect rate. It does not give you certainty. If a critical defect mode is known to occur at rates below that, sampling will not catch it, and you need a 100% functional test on that specific characteristic (a torque check, a pressure decay test, a dimensional gauge) rather than a bigger sample.

That is the decision rule I use: critical defects get a 100% test on the specific characteristic when the test is fast and non-destructive, and AQL 0 sampling when it is not. A neck-finish chip is a visual check, so sampling at AQL 0 is defensible. A cap seal integrity is a torque-and-pressure test, so 100% is feasible on automated lines and I specify it.

Choosing sample size and AQL level for chromatography consumables

The default in most supplier quality agreements is general inspection level II, AQL 1.0 major / 2.5 minor. That is a reasonable starting point for mature, high-volume products with a stable process. It is the wrong plan for a new product, a new line, or a product where the customer’s method is sensitive to a defect mode you have not yet characterized.

Here is how I actually set the plan, product by product:

Two things to notice. First, columns are not sampled in the AQL sense at all. Every analytical column gets a release test (plate count, asymmetry, retention, pressure) and the certificate of analysis carries the actual values. The “lot” for a column is often one or a few units. Applying an AQL sampling plan to columns is a category error. Second, the major AQL for caps and filters is tighter (0.65) than for vials (1.0) because the consequence of a seal failure or a membrane breach is worse than a chipped vial that the inspector catches visually anyway.

When to tighten from level II to level III

Level III roughly doubles the sample size. For a 3,200-piece lot, level II gives 125 pieces and level III gives 200. I tighten to level III in four situations:

• A new supplier or a new production line in its first three lots. The process is not yet demonstrated capable, so you sample harder until it is.
• A customer complaint or a returned lot within the last 12 months on the same defect class.
• A product where the customer’s method has a documented sensitivity to the defect (a validated method with a system suitability failure traced to a consumable).
• A lot produced after a process change, material change, or tooling change that has not yet been through a full capability study.

Once three consecutive lots pass at level III with no critical and no major defects, I drop back to level II. That is a documented switching rule, not a judgment call at the loading dock.

What a realistic outgoing inspection record actually contains

Most inspection records I audit are a checkbox form. “Visual: pass. Dimension: pass. Label: pass.” That record cannot answer the only question that matters when a lot is challenged: what exactly was measured, on what sample, by whom, against what limit, and what did the numbers look like. Here is what a record that survives an audit and a customer complaint actually holds.

The header block

• Lot number and quantity: the actual lot, not the work order. 3,200 pieces, lot 24-0917.
• Product and part number: with the revision of the drawing and the specification the lot was inspected against.
• Sampling plan reference: ANSI/ASQ Z1.4 or ISO 2859-1, inspection level, AQL for each defect class, sample size, acceptance and rejection numbers. Write the numbers, not the standard’s name alone.
• Inspector and date: named, with a training record reference. An unnamed “QC” stamp is not traceability.
• Gauges and equipment used: gauge ID, calibration due date. A micrometer out of calibration invalidates every dimension on the sheet.

The results block

For each characteristic, the record shows the measured value, the specification limit, and pass/fail. Not just pass/fail. If the characteristic is a neck-finish dimension, record the actual micrometer reading. If it is a torque, record the actual torque in inch-pounds. The reason is simple: when a customer reports a failure, the first thing I do is pull the outgoing record and look at the distribution of measured values, not the pass/fail. A lot where every torque reading was 14.8 in-lb against a 15 to 18 in-lb limit passed, but it passed at the edge, and that tells me the capping head was drifting and the next lot is at risk.

For attributes (visual defects), the record shows the defect description, the count, and the location in the sample. “Two vials with neck chips, positions 14 and 51 of 80” is auditable. “Visual: 2 defects” is not.

The disposition block

Accept, reject, or accept with deviation. If deviation, the record names the deviation number, the approver, and the justification. A lot accepted with a deviation on a major defect is a decision someone made, and the record should show who and why. That is the difference between a quality system and a filing cabinet.

Field case: the septum that passed torque and failed the seal

A customer running headspace GC on a 7697A autosampler reported intermittent “vial not sealed” errors, roughly one in 400 vials, across three lots. The outgoing records for all three lots showed torque at 16.2 to 17.1 in-lb against a 15 to 18 in-lb spec, all passing. The septa were 20 mm silicone/PTFE, 3.0 mm thick nominal.

The measurement that identified the problem was not torque. It was septum thickness distribution. We pulled 50 septa from each lot and measured thickness at the center and at the perimeter with a 0.001-inch resolution drop gauge. Center thickness was 2.94 to 3.02 mm, within tolerance. Perimeter thickness on one lot ran 2.78 to 2.85 mm, about 0.15 mm under nominal at the edge. The cap torqued to spec because the torque was measured on the cap, but the seal is made at the septum perimeter against the vial neck. A 0.15 mm thin edge meant the cap bottomed out before the septum compressed enough to seal.

The fix was twofold. First, we added a perimeter thickness check to the outgoing plan as a major characteristic, sample size 20 per lot, limit 2.90 to 3.10 mm. Second, we moved the torque specification to a torque-to-seal test: cap the vial, pressurize to 10 psi, and check for leak-down over 30 seconds. That test catches the thin-edge septum that torque alone misses. The defect classification changed from major to critical because a vial that will not seal stops the sequence. The AQL for that characteristic went to 0, and the test went to 100% on the automated capping line.

Field case: the filter lot that passed particulate and failed the method

A lot of 0.22 µm PTFE syringe filters passed outgoing particulate count (release limit 10 particles/mL above 10 µm) and passed bubble point. The customer, running a method with a 210 nm detection wavelength, saw a baseline drift of 0.8 mAU over a 40-minute run on the first 5 mL through the filter, then it settled. Their method blank had a specification of 0.2 mAU drift.

The measurement that found it was extractables by UV absorbance at 210 nm, not particulate count. We flushed 10 mL of HPLC-grade water through 12 filters from the lot and read absorbance at 210, 254, and 280 nm. The lot read 0.31 AU at 210 nm on the first 2 mL, against a limit of 0.10 AU. The cause was a change in the membrane rinse solvent at the supplier: they had switched from a 50% isopropanol rinse to a water rinse to reduce cost, and the water rinse did not remove the residual membrane manufacturing aid. Particulate count was unaffected. UV absorbance was not.

The fix was to add a UV extractables test at 210 nm to the outgoing plan for every lot of that filter, with a first-flush limit of 0.10 AU and a total-flush limit of 0.05 AU. That test is now a major characteristic at AQL 0.65, sample size 12 per lot, and it has caught two more supplier rinse changes since. The lesson: the release test has to match the detection wavelength the customer uses. A particulate count at 10 µm tells you nothing about a chromophore that absorbs at 210 nm. For a deeper look at how membrane chemistry and rinse affect this, the PTFE membrane filter reference is worth reading alongside the release spec.

Building the sampling plan into the supplier agreement

The sampling plan is only enforceable if it is in the quality agreement, not just in your internal procedure. I put four things in every supplier agreement for chromatography consumables:

• The defect classification table, by name and definition. Attach it as an appendix. A supplier cannot meet a classification they have not seen.
• The AQL levels and inspection levels by product and characteristic. State the standard (ANSI/ASQ Z1.4 or ISO 2859-1) and the specific level and AQL, not “AQL as agreed.”
• The record format. Require actual measured values, not pass/fail, for every variable characteristic. Require the gauge ID and calibration date.
• The switching rule. Level III for the first three lots, level II after three clean lots, back to level III on any complaint. Write the rule so both parties apply it the same way.

I also require the supplier to send the outgoing record with the shipment, not on request. A record that arrives three weeks after the lot is a record you cannot use to make a receiving decision. For suppliers who push back, the argument is simple: the record is the evidence that the sampling plan was executed, and without it the plan is a promise. If you are sourcing across a network, the same discipline applies to how you specify the consumable itself; the lab chromatography product reference and the HPLC filter specifications are the kind of documents that should sit next to the sampling plan in the supplier file. Where method-specific release criteria are involved, the U.S. EPA SW-846 test methods and the manufacturer technical resources from Agilent, Thermo Fisher, and Restek are the reference points I cite when a customer asks why a particular characteristic is on the release plan.

Troubleshooting checklist: when the outgoing record does not match the complaint

When a customer reports a failure on a lot that passed outgoing inspection, work through this in order. The first three items resolve most cases.

• Was the defect class correct? Pull the classification table and the complaint. If the failure mode is a sequence stop or a wrong result, the defect should have been critical. If it was classified major, the plan let it through by design.
• Was the sample representative? Check the sample size against the lot size and the inspection level. A 32-piece sample on a 3,200-piece lot is level I, not level II. Someone may have used the wrong table.
• Were the measured values at the edge? Pull the actual readings. A lot that passed at 0.95 of the limit is a process drift signal, not a sampling failure. Escalate to the supplier’s process, not the sampling plan.
• Was the characteristic measured the same way the customer measures it? The septum case: torque versus leak-down. The filter case: particulate versus UV absorbance. If the test method differs, the plan is measuring a proxy, and the proxy may not track the failure.
• Was the gauge in calibration, and was the inspector trained on this product? Check the calibration due date and the training record. An out-of-calibration micrometer invalidates a dimension, and an untrained inspector invalidates a visual call.
• Was the lot mixed or reworked? Check the lot genealogy. A reworked lot that was re-inspected at a reduced sample size is a common source of escape. Reworked lots go back to level III, always.

If all six check out and the complaint is still real, the sampling plan is not the problem. The process is. That is the point where the conversation moves from QC to the supplier’s capability study, and the sampling plan becomes the evidence that the escape was statistical, not negligent.

The decision rule

Classify every defect by its consequence to the method or the instrument before you pick an AQL number, put critical defects at AQL 0 with a 100% functional test wherever the test is fast and non-destructive, set the major AQL at 0.65 for seal and membrane products and 1.0 for glass and general consumables, and require the outgoing record to carry actual measured values, gauge IDs, and the named inspector. A sampling plan that meets a number but misclassifies a chipped neck is not a quality system. It is a form that says pass.

References & further reading

Related resources from our lab consumables network:

• Lab Chromatography at labchromatography.com
• Hplc Filter at hplcfilters.com
• Ptfe Membrane Filter at ptfemembranefilter.com
• Cod Vial at codvial.com

External references and standards:

• Agilent chromatography consumables and columns
• Thermo Fisher chromatography and sample handling
• U.S. EPA SW-846 test methods for waste and materials
• Restek technical resources for chromatography