SunGold’s cost of quality rose $104 million.Green’s fell $25 million. Same season.

Half of New Zealand’s kiwifruit is on a vessel within five days of harvest. Zespri improved that this season, from seven days in 2025, and wants it faster still. For that half there is no holding decision at all.

The rest is cooled, held, sometimes repacked, and sold across the following months. Somebody chooses which grower lines go into that second group. The industry has a name for the process and a person who runs it: Load Out Priority, managed by inventory managers inside each packhouse. New Zealand Kiwifruit Growers Incorporated defines it for growers in one line — LOP: Load Out Priority, the order in which what fruit is shipped where — and tells growers to ask their packhouse What influence as a grower do I have over my LOP for the season?, because the usual answer is none.

Interviewed packhouse inventory managers describe the eligibility test for holding fruit as four things: the fruit has no market hold, was harvested at good maturity (so no urgency to loadout), has good fruit quality (historically and in the current season) and its condition checking is looking positive. Three of those four are judgements about the fruit. The parenthetical matters — poor keeping quality creates an obligation to ship early.

So this is not a vacuum. It is an experienced human judgement, made line by line, thousands of times a season, and it is different at every one of multiple post-harvest operators, because each runs its own, and we have not found any of them published, nor an industry standard. In 2022 an industry report recommended creating one: There could be an industry wide standardized LOP or storage rating system to define the storage potential of any fruit. We have found no sign it was built.

Follow one line from the orchard. It is picked, and the bins sit to cure — best practice is roughly 24 to 72 hours in a covered, ventilated space, which dries the stem scar and cuts the risk of botrytis later. Then it runs the packing chain, which is where the industry does its sorting. In the industry’s own words, packing is the key control point where the fruit is segregated into market-acceptable product. It is graded for defects, sized, labelled, packed. Quality control staff sample it. It passes.

What the grade cannot catch is damage that has not surfaced. Zespri’s technical manager Frank Bollen put it to growers plainly: Most of the damage is almost impossible to see at the time of packing, but turns into high levels of rots, softs, non-pathogenic fungal growth, and superficial skin rub during storage and shipping. The same briefing recorded fruit dropped onto other fruit or into bins taking five or more days to show damage, so couldn’t be seen during packing.

Five or more days. The grade takes one.

Then the clock starts and never stops. Botrytis, seeded at flowering and entering through the picking wound, surfaces after four to eight weeks in store. Low-temperature breakdown appears after months, and how much of it appears depends on how fast the fruit was cooled and how long it took to reach temperature. Bruising sits under the peel, optically masked. And a single rotting fruit in a tray produces enough ethylene to soften the whole tray.

And the variable most likely to give it away is not in the record that follows the pallet.

The container is not under-instrumented. Its controller logs supply and return air every hour for the whole voyage and holds two years of memory, and Zespri’s guide places a logger on a pallet by the door, deliberately, because that is the warmest air in the box. Temperature is measured, and measured well.

Ethylene is a different matter. The threshold that matters is five to ten parts per billion; at zero degrees sound fruit produces almost none, while one infected fruit does. Onshore the obligation exists — Zespri’s coolstore agreement requires the contractor to monitor ethylene throughout storage. The reefer container controller does not log it. Commercial ethylene monitoring does exist for cold storage and controlled-atmosphere applications; we make no claim about what any vendor offers in transit. What we can say is that the record which crosses the water is a temperature record.

Read the second option closely, because it contains a forecasting problem stated as a commercial one. Sold for as long as it meets the quality expectations required is a question about how long New Zealand fruit holds the shelf. Answer it well and New Zealand keeps the window. Answer it badly in either direction and you either ship fruit that becomes a claim, or hand a month of the selling window to a competitor already inside it.

The length of the New Zealand selling window is now a forecasting output, not a marketing decision.

And the first option — investing in post-harvest innovation — carries a problem of its own. Sizing that investment means knowing how much of the late-season loss was ever avoidable. We have not found that number published. More on it below.

Maturity is tested by an independent laboratory before a block can be picked — dry matter, soluble solids, flesh colour, black seed count — and the industry’s own handbook is explicit that the greater the maturity at harvest the greater the taste and storage potential of the fruit. Pallet temperatures are monitored wirelessly through the chain. A Fruit Monitoring Technician rides every chartered reefer from mid-March to mid-June. And in 2024, 22,582 pallets were physically condition-checked, which is 7.5 percent of everything shipped.

Zespri models risk too, and says so. Its June 2026 grower newsletter credits fruit quality risk modelling among the contributors to this season’s result. Since November 2024 a Strategic Product Inventory project specific to Europe has matched quality, clearance and market-access data right down to individual lines to calculate risks such as Softs, Storage Breakdown Disorder, Skin Disorder.

So the honest gap is not measurement, and it is not modelling. It is what the answer is attached to.

Every one of those readings takes the state of the fruit at a moment, on a sample, downstream of the commitment. The Strategic Product Inventory work runs on fruit already in market and sequences what is already there. We have not found anything that returns a number of the form this line, on this route, has an eighty percent chance of still meeting standard on 14 October — produced on the day the line is packed, while the holding decision is still open.

Three things are open at once, and all three will close. The industry is writing a longer-term fruit quality strategy with the Industry Advisory Council’s Fruit Quality Steering Committee, and that includes setting shared quality targets. Cost of quality is named in the Five-Year Outlook as a key priority — managing our costs, with a particular focus on managing the cost of quality. And the Annual Report describes planning and implementing a new framework for soft defect analysis to remove variability and inconsistencies in measurement across regions.

That last one matters most, and it is not about us.

If a soft fruit gets counted one way in Rotterdam and another way in Shanghai, then comparing the two markets tells you as much about the counting as it does about the fruit. Zespri knows that and is sorting it out. So this work has to come after theirs, not instead of it — and we would rather it did.

But counting the same way everywhere does not explain why the numbers differ. Once the counts mean the same thing, somebody still has to say what is driving them, and a target set on a number that has not been taken apart is a guess with a decimal point on it.

One caution is worth saying plainly. If you compare markets without first allowing for which fruit went where, you are not really comparing markets. Say Europe tends to get fruit from different regions, picked on different days, held longer before it ships. Then Europe softens more is partly just a statement about what Europe was sent. Telling the journey apart from the fruit can only be done line by line.

Two incumbents, and both are better than they look.

The first is Load Out Priority — the inventory manager’s judgement, applied line by line, using maturity, orchard history, current-season quality and condition-check results. It is real expertise and it is doing genuine work. Its limits are the limits of any expert judgement made thousands of times: it is unwritten, it differs at every packhouse, it produces eligible-or-not rather than odds, and no two managers weigh the four criteria the same way.

The second is the fruit loss curve. Time Payments compensate growers for fruit sold late — coolstorage, condition checking, repacking, fruit loss, foregone taste. In 2024 the average rates ran from $1.65 a tray for Class 1 Green to $2.94 for Organic Gold3, across roughly 220 million trays. Zespri’s own Grower Payments booklet describes how the rate is set: from actual industry fruit loss records averaged over a set number of seasons, accumulated and volume-weighted by week, creating the predictive fruit loss curve for the coming season. And then, to every category, a premium is added — the storage incentive — to cover the risk of fruit value or the fruit loss curves used in the calculation of the time rates being different to the actual results at the end of the season.

Read the second half again. Zespri has written into its own payment system an explicit acknowledgement that the average loss curve does not describe any particular season’s fruit, and has priced that uncertainty as a flat margin spread across everybody.

It is the same failure as ordering stock for every shop against one national size curve. The store-level curves are real and they differ. The average erases them, and then someone adds a safety margin to cover the erasure. The machinery to reward putting the right fruit in the early slot already exists and is already funded. It has no per-line signal to act on.

New Zealand has been working on this since at least 2003, and the work is good. Massey doctoral theses in 2003, 2014 and 2017; published papers in 2013 and 2022; a Plant & Food Research post-harvest programme funded by Zespri. We have not found any of it in commercial use.

That is not because the industry ignores research. Look at what was adopted. Curing fruit for 24 to 72 hours before cooling — standard practice. Keeping vine canopies open so botrytis cannot build up in dead leaves — adopted, and it largely solved a storage rot problem that was costing the industry around ten million dollars a year. Minimum dry matter and brix before picking is permitted — independently laboratory-tested and enforced on every block.

Every one of those is the same instruction for everybody. Every one of them worked.

Now look at what did not get adopted. This line will last until October; that one will not. That is not a rule. It is a different answer for every line, every time, and there is no way to write it into a standard, train a workforce on it, or audit compliance with it. It has no institutional home, so it stayed in the journals.

And there is a consequence that is easy to miss. Every rule the industry adopted removed the part of the loss that responds to a uniform fix. What is left is, by construction, the part that does not. So the share of remaining loss that needs a per-line answer rises every time the industry succeeds — even as total loss falls. This is worth more now than it was in 1990, not because anyone got worse, but because they got good.

And the research had three handicaps we would not have

It had to sample. Between 27 and 108 grower lines, over one or two seasons, using an instrument bought for the study. You cannot put a spectrometer on a national crop, so the studies were small by necessity.

It measured a stand-in for the outcome. Every study predicted laboratory penetrometer firmness after controlled storage — in the most recent case, with ethylene held below 5 nL/L for the whole experiment, which is to say with one of the biggest real-world causes of failure deliberately switched off. We have not found a model fitted to a commercial outcome.

And the season moved underneath it. A segregation rule calibrated in one New Zealand season correctly categorised 60 percent of lines that year and 53 percent the next, against a 33 percent chance baseline. Underneath that, the share of genuinely poor-storing lines swung from 35 percent to 53 percent between two consecutive seasons. That is a season effect large enough to swamp a line effect in any single year.

The result is a consistent asymmetry across every attempt: the models identify good fruit at around 78 to 80 percent, and miss roughly half the bad. Bad fruit is the only fruit that costs money. That, in one sentence, is why the work sits on a shelf, and the authors say so themselves — the level of accuracy achieved was not adequate for online sorting purposes.

Note what they blame. Many pre-harvest factors contribute to variation in fruit quality at harvest and during coolstorage, resulting in the difficulty in segregating fruit for their storage potential. Pre-harvest heterogeneity, and not being able to read the signal precisely enough. Not one of them says the outcome is decided after packing.

Estimate at the finest unit where signal-to-noise supports estimation. Then check that a decision is actually made at that unit. If the two do not coincide, you have a research project or a feature, not a product.

The published evidence draws that line precisely. When Massey researchers segregated 27 commercial ‘Hayward’ grower lines at harvest using full visible and near-infrared spectra, classification of the individual fruit that would go soft reached about 54 percent sensitivity in calibration and 32 percent on a new season. But segregating between grower lines cut soft fruit by 30.7 percent, from 25.1 percent to 17.4 percent.

And the size of the difference between lines is not marginal. Across those 27 commercial lines, soft fruit ranged from about 2 percent in the best line to about 64 percent in the worst. A thirty-fold spread, from fruit that all passed the same grade.

The signal survived aggregation to the line and did not survive at the fruit. A grower line is a real physical object — one orchard, one harvest date, one packrun, one shared handling and cooling history — which is why noise averages inside it. It is also the unit at which fruit is actually held and sequenced.

But the line’s average is the wrong summary of it.

Two results say the same thing from opposite directions. A 1999 New Zealand study found that when firmness was determined prior to shipping, no device discriminated between orchards for subsequent softening — and then that softening was predicted by the high between fruit variability in the firmness of a 20 fruit sample. The spread, not the mean. And work on variance propagation in post-harvest systems showed that fruit-to-fruit variation inside a batch is not noise but a distribution that moves forward predictably, reaching an adjusted R² of 0.96 in tomato from nothing but the spread of colour at harvest.

So the question to ask of a line is not will it fail. It is what fraction of it will fail, and by when. Anything that summarises a line by its average dry matter has thrown the predictive part away before it starts.

Ross Ashby put the general form of this in 1956: every good regulator of a system must be a model of that system. The regulator has to match the variety of what it regulates — no coarser, and no finer.

We went looking for a published answer specifically — variance components, random effects, intraclass correlation — and checked the four New Zealand theses that would carry it. We did not find one, for kiwifruit or for any other fruit export chain we searched. We may have missed it, and if the work exists inside Zespri we would rather build on it than repeat it.

That is the number the whole question turns on. How much of the final outcome is attributable to the grower line, how much to the season, how much to the route and the handling, and how much is simply irreducible.

Producing it is not a model. It is a measurement, and it is the honest first piece of work. It also settles whether the second piece — a per-line forecast conditioned on the route, and an optimiser that assigns lines to the coolstore, controlled-atmosphere and sailing capacity that actually exists — is worth building at all.

Telling the fruit apart from its journey

Lines are not sent to markets at random. Somebody chooses, and they choose on what they know — fruit expected to hold well probably gets the long route, or gets held longer. That tangles up the effect of the line with the effect of the journey, and makes one look bigger than it is at the other’s expense. It is the same problem doctors have when a treatment was not handed out at random: the sicker patients got the stronger drug, so the drug looks worse than it is.

What pulls the two apart is overlap. The same line split across several destinations, and the same route carrying many lines. Where there is plenty of overlap the two effects can be told apart; where there is little, they cannot, and the honest answer there is a wide range rather than a single number. Finding out how much overlap actually exists is part of the first piece of work — and it is one of two reasons the study needs every line in its slice rather than a sample of them.

The other obstacle, stated honestly

The outcome has to reach back to the line, and the chain is not clean. Fruit mixes on the packline and again at repack. New Zealand research published in 2007 established that grower-to-pack traceability through a packhouse is probabilistic, not deterministic, and built the measures for it — precision of traceability, packs per bin for tracking, bins per pack for tracing. That work named the two applications it would enable: feedback to growers on measured quality and feedback to the market on predicted quality. We have found no sign it was finished.

Nor does the commercial outcome come back down the same thread. A quality claim, in Zespri’s own Pool Policy Manual, is credit issued to customer following assessment and acceptance of a formal quality claim, valued for offshore losses at ‘Arrived Sound Market Value’ applicable to the Zespri variety at the week of arrival, and for onshore insured losses at forecast FOBS value at the time of loss. It is then allocated by way of the material master code to the actual pool of sale — which is how a claim reaches a pool, not how it reaches a grower. Variety. Week. Consignment. Pool. It is a finance record, and it was never built to point at a packed line.

Attributing one claim to one line is probably impossible. It is also not necessary.

Where enough lines appear across enough overlapping deliveries, line effects can be recovered without ever knowing which tray failed. That is what needs the mathematics, and it is precisely why the work needs every line within its scope rather than a sample of them. A study with 27 lines needs a clean join or it has nothing. A study with thousands does not.