Your sampling plan assumes bad luck. The battery fails as a batch.

Here is an uncomfortable observation. CATL – the largest battery maker on earth – has said the word batch out loud, publicly warning of batch battery failures while China's industry launches nearly three new EV models a day. I have spent two decades in automotive and aerospace quality. I have yet to meet an acceptance plan that survives the word "batch" intact. Not because the plans are badly written. Because they were written for a different animal.

Your AQL tables and end-of-line testers were built for scattered, independent, immediate defects: a stray burr, a missing clip, a mis-torqued bolt. Batch failure is correlated, latent and lot-shaped. Three words, and each one removes a load-bearing wall from the mathematics. When defects travel as families, unit sampling is theatre.

the mathematics your sampling plan quietly assumes

Pull n parts from a lot of N. Accept if the defect count stays at or below c. Every operating-characteristic curve behind that decision rests on three quiet assumptions: the defect probability is identical for every unit in the lot, one unit's fate tells you nothing about the next, and the defect is visible at the moment of test.

Batch failure breaks all three. When one cell's defect raises the odds for its siblings – shared electrolyte, shared formation cycle, shared coating drum – the defect has clustered. Genuinely random draws would still be sound; the hypergeometric does not care where defects sit. But nobody at goods-in draws randomly from a sealed pallet, and the sample steps straight over the cluster. The end-of-line tester breaks the third: it meets the battery at hour zero of a chemistry that fails on a clock. Lithium plating and contamination-driven internal shorts are month-fourteen problems. The bench cannot see time.

Sampling is a bet that defects are strangers to one another. Batch failure is the proof that they arrived as a family.

why batteries fail as families, not individuals

Walk the mechanisms. Electrode coating thickness drifting across one shift. A weld-energy window missed on one fixture. A formation protocol run at the wrong temperature. An incoming raw-material lot sitting at the bottom of its spec – passing piece by piece, failing as a population. Each mechanism stamps parts with the same birthday and the same failure mode, and because packs wire cells into electrical families, the same consequence multiplied.

We have seen this film before. Takata's propellant degraded by production lot – latent, lot-shaped, found in the field years after the dock accepted it. Automotive quality did not lack the lesson. Batteries raise the stakes; they do not change the logic.

release the lot, not the sample

I built a greenfield QA/QC department for a 900-plus-employee plant from a blank sheet of paper, incoming acceptance and lot control included. Every instinct in the building said inspect harder – bigger samples, tighter AQL. I went the other way. Fewer samples, more lot evidence. Every delivery carried a record: supplier lot, production window, first-piece approval, SPC signals from the supplier's own process wherever we could get them. Sampling became a spot-check on lot integrity, not a substitute for it.

The QRQC room taught me the euro version. A suspect lot of welded brackets – hairline cracks from one drifted shift at a supplier – 6,000 pieces, flagged at goods-in. We sorted at the dock for €14,000 and two days of buffer. The same lot escaping to the field was the other ledger: containment at the OEM, sorting on their line, premium freight both directions, warranty reserves, chargebacks, an 8D team parked in a hotel for weeks. That bill crossed €400,000 before anyone said "root cause". One lot. Twenty-eight times the cost.

The 70% defect-cost reduction we delivered at SNOP did not come from inspecting harder. It came from reading lot-level patterns early – Pareto by lot signature rather than defect mode – so families were caught at the dock instead of in the field. The lot-acceptance discipline I absorbed across two decades in automotive and aerospace – AS9100/EN 9100, VDA 6.3, first-article inspection, positive release per lot – never trusted unit sampling alone. Automotive inherited its AQL tables from MIL-STD-105 and forgot the fine print. The aerospace instinct is the right one for batteries.

launch velocity will not slow down

Nearly three new EVs a day. Exports running past last year's totals. Every model a new pack, every pack a new process, every process a fresh crop of first articles. No launch team has ever volunteered to go slower. So the release logic has to be lot-shaped before the ramp, not after the recall. Release the lot, on lot evidence, every time the process changes. Pay at the dock, where the maths is merely inconvenient. Or pay in the field, where it compounds and the customer keeps score.