Ten Customers From the Same Production Batch Reported Tunnelling — At What Point Should I Suspect a Batch or Wick Problem Rather Than Customer Usage?

Ten is not a number yet. Make it a rate.

The messages arrive in a heap, all saying the same thing — burned straight down the middle, wax left round the edge — and the instinct is to act. Act on what? A count on its own is not evidence; the denominator is where the information lives. Not "how many did I make" but how many of that batch have reached a customer and been lit: units in a carton cannot complain, and units dispatched three days ago have probably not been burned properly yet. The honest denominator is units of that batch delivered long enough ago to have had a real burn — for most small makers, what shipped up to a fortnight before the first complaint landed.

Work the middle case, because it is where most small Indian makers sit: a batch of 300 candles, 240 delivered, ten reports. That is 10 ÷ 240 = 4.17%, or 4.2 complaints per hundred delivered units — the first thing on this page that is a fact rather than a feeling.

Time matters as much as volume. Ten spread over six months, one at a time from different runs, is a standing feature of the product and needs a product answer. Ten arriving inside nine days, all naming units dispatched in the same fortnight, is an event — and events have findable causes, because something was different in a short window you can inspect. Put every complaint on a timeline with two dates: the day the unit shipped and the day the customer wrote. If the ship dates bunch, you have a window. If only the complaint dates bunch, something changed in the world rather than in the wax.

A rate means nothing without a baseline

On its own, 4.2% is a decimal. Against a house average of 4% it is an ordinary batch; against 0.3% it is fourteen times your normal and the strongest evidence you will get that something went wrong in one identifiable place. The number does not change; its meaning changes entirely. And the uncomfortable part, plainly: if you have never counted, you have no baseline and this investigation cannot be completed. The clustering test is still worth running, but you will not know whether 4.2% is abnormal, so you will decide about your stock on judgement rather than evidence. Starting the count is itself the first corrective action, and it is free.

Arithmetic rather than statistics, to calibrate expectation: if your true background rate were 1%, then in 240 delivered units you would expect two or three complaints, not ten. Ten is roughly four times that. Not proof of a defect and no cause named, but a reason to stop and look — and a far stronger basis for action than "ten people complained".

The clustering test, as a procedure

This is the part that actually distinguishes a batch or wick problem from customer usage, without either party being blamed in advance. List everything you know about the ten complaining units, then list the same things for the units that did not complain. The cause, if it is on your bench, is in the difference. On the 300-candle example that is ten complaints against 230 delivered units that have not complained — a control group costing nothing. All 230 share your instructions, customers, season and channel with the ten; what they need not share is what went into them.

You need not characterise all 230 — sample twenty non-complaining units at random and check the same columns; a wick lot common to all ten complaints and present in two of twenty controls is discrimination. Then ask the customer for what you cannot see: a photograph of the jar as it is now, how long each session ran, whether the candle was moved between rooms, whether a fan was running, whether the first burn was short. Ask neutrally, because you are gathering data, not building a defence, and their answers are half your evidence.

Batch traceability: the minimum viable record

Blunt rather than gentle: a batch number printed on a label is not traceability, it is a label. Traceability is the batch number connecting to a record, and most small makers discover the difference at the worst moment — when ten messages arrive. The minimum viable record is one side of A4 per pour session, filled in at the bench rather than reconstructed a week later. Not software: a sheet, a pen and a marker on the base of the jar.

Retained samples, three costs, and a decision that is yours

You have a rate, a comparison and a lead. What you do not have is an observation. Every photograph you have been sent shows an end state, and end states are ambiguous. The retained sample is the one piece of evidence where you control the usage, which makes it the one place usage and product can be separated. If you kept none, this arm cannot be run, and starting to keep them is the first corrective action however this ends.

On the wick, we will not be helpful in the way you may want: CSI publishes no wick sizing chart, no jar-diameter recommendation and no burn-performance figure for any wick in the range. Eco Candle Wicks are listed as Thin (C1) and Thick (C2) and that is the specification in full. No lookup resolves "is my wick right for this jar", so only your own burn record will. And against our own interest: do not take any supplier's sizing claim without its test conditions, ours included — a recommendation with no stated wax, load, fill weight, vessel and burn pattern is a number without a method.

Now the money, because a decision about remaining stock deserves numbers rather than nerves. Cost one unit at CSI live pricing — a 200 g candle at 8% in Luxury Soy Wax Chunks at the 10 kg tier, Sandalwood at the 1 kg tier, an Amber Shiny Glass Jar with Black Lid, one Eco Candle Wick Thin (C1) and one wick sticker. Where a jar publishes no fill capacity — most do not, and "200 ml" is a volume, not a 200 g fill — pour one and weigh it.

The test costs 2.96% of the re-pour — ₹2,186.46 against ₹73,923.00 — and it is the only one of the three that produces information. The hold buys time, and it is not a loss unless the stock proves fine and you have missed a selling window. Test before you decide.

Then write down what you found, including the null result — if the grid produced no discriminating variable, the cause is not in the columns you recorded, which tells you which column to add next time. Log the rate, the grid, the burn records and the conclusion against the batch number, and the next cluster starts from a baseline rather than from scratch. It compounds with volume: the fault costing ₹73,923 at 300 units costs ten times that at 1,000 orders a month, which is where the arithmetic of pricing and margin stops forgiving unexplained losses.

Frequently asked questions