Zero-acceptance plans on small samples are a coin flip, and the coin is

A 12 mm crimp lot came back from a contract lab with three leaking seals out of 40 on the same GC run. The supplier’s outgoing record said “passed.” It had passed, technically: the inspector had pulled 13 vials from a 3,000-piece lot, checked them against a zero-defect acceptance number, and shipped. The problem was not that anyone lied. The problem was that the sampling plan accepted a process that was running at roughly 1.5% seal defects, and a zero-acceptance plan on 13 pieces cannot see that. That is the gap this article is about: the sample vial AQL sampling plan and outgoing QC that sits between a finished lot and a shipping label, and why most of the ones I have audited are calibrated to catch nothing.

I have written and rewritten these plans on the manufacturing side for close to fifteen years. The version I trust now looks nothing like the generic AQL table copied into a quality manual. It has a defect classification that separates a cosmetic scuff from a non-sealing crimp, a sample size that scales with lot size the way sample vial defect rates actually behave, and a record that a customer can read in ninety seconds and know what was measured. Below is how I build it, with the numbers.

Zero-acceptance plans on small samples are a coin flip, and the coin is weighted wrong

The default move is to grab ISO 2859-1 general inspection level II, pick a code letter from the lot size, and run a normal single-sampling plan at AQL 1.0 or 2.5. On paper it is defensible. In practice, most people applying it to vials use the wrong column, or they use a zero-acceptance number that was never designed for the lot size they are running.

Here is the arithmetic that matters. A single-sampling plan with acceptance number zero and sample size n will accept a lot with true defect rate p with probability (1 − p)ⁿ. Run it for a real vial defect rate:

• n = 13, p = 1.5%: probability of acceptance = 0.985¹³ = 82%. The plan passes a 1.5% defective lot four times out of five.
• n = 32, p = 1.5%: 0.985³² = 62%. Still passing the majority.
• n = 80, p = 1.5%: 0.985⁸⁰ = 30%. Now the plan has teeth.
• n = 125, p = 1.5%: 15%. This is where I want a critical-defect plan sitting.

If your outgoing plan samples 13 pieces and accepts on zero, you are not screening for a 1.5% seal defect. You are screening for a process that has already fallen apart. That is fine as a gross-error tripwire. It is not a quality gate, and calling it one is how the contract-lab failure above happens.

What AQL level to actually put on the drawing

I split the plan by defect class rather than running one AQL across everything, because a dimensional miss on the crimp skirt and a chipped lot number are not the same risk. My standing recommendation for a 12 mm or 13 mm crimp vial, and for a 9 mm screw-thread vial on a vials with caps assembly line:

Two things about that table. First, critical defects get a special inspection level (S-4) with a tight AQL, not the general level II column, because the sample size has to be big enough that a 1.5% critical rate is caught more often than not. At n = 50 and p = 1.5%, acceptance probability is 0.985⁵⁰ = 47%. That still lets a bad lot through half the time, which is why critical defects also get a 100% functional check on the seal (see below) rather than relying on the AQL sample alone. Second, the major and minor classes share a sample size at level II, so the inspector pulls 200 pieces once and grades each piece against both lists. You do not pull two samples.

Switching rules you have to write down

ISO 2859-1 has normal, tightened, and reduced inspection with switching rules. Most plans I see copy the tables and drop the switching rules. Then the plan never tightens after a rejection, and a drifting process ships for months. Write the rule explicitly: two lots rejected out of five consecutive lots on the same characteristic moves the line to tightened (which for a level II plan roughly doubles the sample size and drops the acceptance number), and it stays tightened until five consecutive lots are accepted. Reduced inspection is a trap for vials and I do not use it. The savings in inspection time are trivial next to the cost of a seal failure at the customer.

Defect classification is where these plans actually fail audits

An auditor does not argue with your AQL number. They argue with the boundary between major and minor, because that boundary is where a lot gets reclassified after the fact. I have watched a supplier defend a shipment by moving “septum not fully seated” from major to minor in the inspection record. The auditor found it. The whole lot became suspect.

The fix is a defect classification sheet with a physical test attached to each class, not a verbal description. For a crimp vial the critical class has to be defined by a measurement, not an opinion:

• Seal integrity: cap the vial to the specified torque (I use 1.5 to 2.0 in-lb for a 9 mm screw cap, hand-crimp to a skirt diameter of 7.9 to 8.1 mm for a 11 mm crimp), then pressurize to 1.5 bar for 30 seconds and look for a pressure drop or a bubble at the septum. A vial that fails this is critical, full stop.
• Septum puncture count: pull a sample and puncture each septum five times with a 22-gauge needle at the same point, then run the seal test again. A septum that leaks after five punctures on a lot specified for autosampler use is a major defect, because the customer will hit it on the fifth injection.
• Neck and thread dimensions: measure neck OD and thread pitch with a go/no-go gauge or a pin gauge. A 13 mm neck OD outside 13.0 ± 0.1 mm is major. It will fail to seat in a crimper or a screw cap.
• Particulate: inspect the cavity under 10x magnification against a black background. Any visible particle is major. This one gets skipped constantly and it is the defect that shows up as a ghost peak in someone else’s chromatogram.

If you want the reference for how these consumables are specified upstream, the Agilent chromatography consumables documentation and the Thermo Fisher sample handling pages both publish dimensional and material specs you can benchmark against. Do not invent tolerances when the instrument maker has already published what the autosampler will accept.

The classification sheet is a one-page document

Keep it to one page per product family. Columns: characteristic, method (gauge, torque, pressure, visual), class (critical/major/minor), and the accept/reject number. If the inspector has to interpret, the plan is broken. I have seen a two-page classification sheet get applied three different ways by three shifts, and the outgoing records looked identical because nobody wrote down which method they used.

What a realistic outgoing inspection record contains

This is the part customers ask for and suppliers fake. A real record from a lot of 20,000 12 mm crimp vials, sampled at level II, looks like this. It fits on one page and it is auditable:

• Lot and traceability: lot number, molding date, resin lot, septum lot, cap lot, and the machine and cavity range that produced it.
• Plan reference: the document number of the sampling plan, the AQL level per defect class, the inspection level, and the sample size actually drawn (n = 200 for major/minor, n = 50 for critical).
• Measured values, not just pass/fail: neck OD mean and range, crimp skirt diameter, cap torque mean and range, seal pressure-hold result, puncture-count result. A record that says “dimensional: pass” is not a record.
• Defect tally by class: number found in each class, the acceptance number, and the disposition (accept, reject, sort, rework).
• Instrument and gauge IDs: the torque wrench calibration due date, the pressure rig ID, the gauge set. Auditors pull this first.
• Inspector, date, and shift, plus the signature of whoever released the lot.

One number I insist on seeing: the observed defect rate per class, even when the lot passes. If a lot passes at 2 major defects in 200 (1.0%), that is a pass under a 0.65 AQL plan with Ac = 3, but it is a warning. Two consecutive lots at 2 or 3 majors is a process drifting toward the acceptance number, and I want the line to know that before the third lot rejects. The record should carry that trend, not just the disposition.

Field case: the crimp that passed the AQL and failed the autosampler

Symptom: a customer running a 9 mm screw-thread vial on a split/splitless inlet reported intermittent septum coring and a rising baseline after about 40 injections on a tray. Measurement: we pulled the returned vials and measured cap torque. The mean was 1.1 in-lb, below the 1.5 in-lb floor, and the septa had deformed under the lower compression so the needle punched rather than pierced. The outgoing record showed “torque: pass” because the inspector had checked torque on the capper, not on the sampled vials. Fix: moved torque measurement into the outgoing plan as a measured value on the AQL sample (n = 200, mean and range recorded), set the acceptance window to 1.5 to 2.0 in-lb, and added a five-puncture seal test to the critical class. The coring stopped on the next lot.

Sampling frequency and where in the lot you pull from

Two decisions here, and both get made badly. First, frequency. A single AQL sample at end of lot is the minimum. On a line running more than one shift, I sample per shift, because the capper torque drifts with temperature and with tool wear and a single end-of-lot sample averages that drift away. On a 20,000-piece lot produced over two shifts, that is two samples of 200, not one.

Second, and more important: where in the lot you pull. The classic mistake is pulling the top of the carton. Vials at the top of a tote have taken different handling than the ones at the bottom, and the crimp skirt burr that only appears on pieces that ran against the chute wall will not be in the top layer. I pull stratified: from at least five positions across the lot (first-off, mid-run, last-off, plus two from different cartons), and I record which position each defect came from. If all your majors come from one carton, you have a handling problem, not a molding problem, and the fix is different.

The sampling plan should state the stratification rule explicitly. Mine reads: “Draw from a minimum of five positions spanning the production run; record position code on each piece inspected.” That one sentence has caught more real problems than any AQL table I have ever used.

The functional checks the AQL sample cannot replace

No AQL plan on a sample of 50 to 200 pieces will reliably catch a 1% seal defect, as the arithmetic above shows. So the critical class needs a 100% or near-100% functional gate in addition to the statistical sample. For vials I run two:

• Seal leak test on a defined subsample: pressurize to 1.5 bar, hold 30 seconds, accept on zero leaks. I run this on every piece in the critical sample (n = 50) plus a random 1-in-50 pull from the full lot for lots going to regulated customers. That second pull is the one that catches the slow leak the 50-piece sample missed.
• Puncture test on the septum lot: five punctures at 22 gauge, then re-seal test. Run per septum lot, not per vial lot, because the septum is the variable and it arrives in its own lot.

These are not statistical sampling. They are functional gates, and they belong in the outgoing record as measured results, not as an AQL disposition. Mixing the two in the record is how the “passed” label gets applied to a lot that was never actually screened for the failure mode the customer cares about. The Restek technical resources on septa and inlet hardware are a good reference for what puncture and bleed behavior to expect from a given septum material before you set the acceptance criteria.

Field case: the particulate that the AQL plan graded as minor

Symptom: a pharma customer running LC-MS reported sporadic ghost peaks at a consistent retention time across three lots. Measurement: we inspected the returned vials under 10x against a black background and found visible particles in roughly 0.8% of the cavity, all in one cavity range from one molding machine. The outgoing record had graded particulate as minor at AQL 2.5, and the sample of 200 had found zero, so the lot passed. Fix: reclassified visible particulate as major, added a 10x black-background inspection to the major class, and added a cavity-code trace to the record so a single-cavity problem can be isolated. The ghost peaks traced to a mold release residue on that cavity, not to the resin.

Writing the plan so a customer can audit it in one sitting

The plan document itself is short. Mine runs to about four pages and has five sections: scope and product families, defect classification sheet (one page per family), the sampling tables with the switching rules, the functional-gate procedures, and the record template. That is it. The U.S. EPA SW-846 container requirements are worth reading once for the language on what a container must do, because it sharpens how you phrase the critical class, but the plan is your document, not the method’s.

If you are buying rather than making, the ask is specific: request the sampling plan document number, the defect classification sheet, and two recent outgoing records for lots of similar size. Read the records for measured values, not pass/fail. If the torque is not a number with a range, the plan is decorative. The Aijiren technical documentation is one example of a supplier publishing the dimensional and material basis for these specs; use it as a benchmark when you compare suppliers, and compare the chromatography products range each one actually controls in-house versus buys in.

Decision checklist before you sign off a plan

• Critical class has a physical test (pressure hold, puncture count, go/no-go gauge), not a description.
• Sample size for the critical class is large enough that a 1.5% defect rate is caught more than half the time (n ≥ 50 as a floor, and pair it with a functional gate).
• Major and minor classes share one sample at level II; you are not pulling twice.
• Switching rules are written: two rejects in five lots moves to tightened, five clean lots returns to normal, reduced inspection is disabled.
• Stratified pull from at least five positions across the lot, position recorded per piece.
• Record carries measured values (mean and range) for neck OD, crimp skirt or cap torque, seal result, puncture result.
• Gauge and instrument IDs with calibration due dates are on the record.
• Observed defect rate per class is recorded even on passing lots, and trended across lots.

The decision rule I would put in front of any procurement manager is this: a sample vial AQL plan is only as good as its critical-class definition and the functional gate behind it. If the critical class is defined by a measurement (1.5 bar hold, five punctures at 22 gauge, 13.0 ± 0.1 mm neck OD) and backed by a 100% or near-100% seal and puncture gate on a defined subsample, the AQL number on the drawing is doing real work. If the critical class is a verbal description and the only gate is a 13-piece zero-acceptance sample, the plan is a formality, and the 1.5% leak rate will find its way to a customer’s GC run eventually. Set the critical AQL at 0.065, pull 50 for it, run the seal test on all 50, and keep the record in measured numbers. That is the plan I would defend in an audit, and it is the one I would want to receive if I were the one buying.

References & further reading

Related resources from our lab consumables network:

• Sample Vial at samplevialprice.com
• Chromatography Products at chromatographyselling.com
• Aijiren at aijirentech.com
• Vials With Caps at vialswithcaps.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