How to build the world’s first autonomous packhouse

· the plant today

BLIND TO BIOLOGY?

Here’s the list of jobs on a fresh-cut line. Third column is what each job actually reads. It’s almost always one of two things: how the product looks, or a chemical stand-in. The biology never gets read while you can still do something about it. The blanks are the point.

One row reads real biology. It’s also the row that does the least. A finished-product test pulls a composite of a few pounds out of a lot running to tens of thousands. Less than one percent. And it’s hunting an organism that sits in clusters, not spread evenly, so missing it is the default outcome. Then it spends a day or two growing the sample before it says anything.

So: under one percent, two days late, against a product that leaves the building in six hours.

The plant is not lightly instrumented. On the one question that matters it isn’t instrumented at all.

· the two failures

TWO FACES OF THE SAME PROBLEM

Two things go wrong in produce. Product rots, and once in a while it makes people sick. These sit in different departments with different budgets. They’re the same failure.

Rot: cut greens ship in pack-date order, so a lot that was warm at harvest and already carrying load sits behind a lot that would have kept another week. Shrinkage.
Safety: you dose to a chlorine number, ship in six hours, hear from a lab two days later. Shipment’s gone by then. Recall.

Same blindness, twice.

· FOOD IS A BIOREACTOR

FOOD IS ALIVE

A bag of shredded romaine is a few thousand wounded pieces of living tissue, still running in a way. Respiring, senescing, losing water, leaking sugar out of every cut edge into a film of water that is full of bacteria and other microorganisms.

Several clocks, all ticking. Each has a state and a rate. The rate is set mostly by temperature. Cold slows every clock, warm speeds them all up together. Cutting speeds them up permanently, which is why a whole head keeps for weeks and the same head shredded keeps for days. Drag the temperature:

predicted remaining life

{{ life }} days

respiration · burning sugars{{ r0 }}

senescence · yellowing, browning{{ r1 }}

transpiration · wilt{{ r2 }}

microbial load · soft rot & safety{{ r3 }}

product temp {{ tval }}

{{ take }}

One number moves and all four clocks climb. The microbial one climbs fastest.

What spoils bagged greens is mostly Pseudomonas, which turns cut edges soft and ends the product. What causes people sick is E. coli O157:H7, Listeria, Cyclospora. Different organisms. Same water.

· A new WAY OF ORGANIZING

Sorter vs. router

Today the building reads looks, ejects defects, ships in pack-date order. The version I want reads the water, estimates remaining life per batch, sends each batch where its clock allows. Short-life to the near DC, or foodservice that turns in two days. Long-life to the far account. Unsafe nowhere. Same greens, same building, different question.

In order. Each step stands alone. The safety work in steps 2 and 7 pays for the routing work in 3 through 5.

1 Buy the boring 80%, then delete

Most of a packhouse is already automated and isn’t new. Fresh-cut lines are a mature market. Turatti, Sormac and FTNON build the coring, trimming, washing and drying lines. Urschel makes the cutting equipment most of the industry runs. TOMRA and Key Technology make the optical sorters. Integrators bolt these into working plants right now.

Drizzle has now built the sense organ. Add AMRs, forklifts, vacuum cooling and ORP dosing and you have everything.

If you design around per-batch measurement from scratch, several jobs don’t get automated but stop existing altogether. Why? Because they were only ever there to compensate for not knowing.

Automation moves a person off a task. Deletion means the task was an artifact of not knowing the underlying biology that is central to decision-making in this new type of packhouse. Most of the labour on a fresh-cut line is the second kind, which is the argument for building rather than retrofitting.

2 Put an eye on the water

· CAN BIO BE MEASURED?

Census, not sample

The obvious objection to reading the water: the water is shared. Every leaf floats through the same flume, so a reading belongs to the tank, not to any one batch. 

Compare what the two measurements cover. A finished-product test reads a few pounds out of tens of thousands. Under one percent, hunting something that sits in clusters. Miss the cluster and the lot passes quality and safety checks.

The flume touches every leaf in the lot. Whatever is on the product has one place to go, and that’s the water.

Sampling the product is a lottery ticket. Sampling the water is a census. And the plant already built the machine, it just uses it as somewhere to pour chlorine.

· BALLPARKS

PILOT PARAMETERS

3 Read it, don’t look at it

With the sensor in, “process” shifts from “eject the defects” to “measure & route”. The same instrument reads two things out of the same water, because spoilage and safety live in the same habitat. Total and Pseudomonas load on one channel gives you how long the product has. Pathogen targets on another decide whether it goes anywhere. One architecture, multiplexed.

Reading state is bounded engineering. Turning state into a per-batch countdown, “this one has eleven days,” is the frontier and the models will get there with enough labelled biology in the plant.

Whatever model you use starts coarse and sharpens on the record the plant feeds back to it. 

4 Route on the clock

Now “out” changes. Instead of pack-date order, every batch leaves with a "passport": clocks, estimated remaining life, where that life lets it go, priority. Short life to the DC two hours away, or foodservice that turns stock in a day. Long life to the far account. The plant ships product and an instruction set.

Five decisions off based on that passport.

1. Routing. High-load batches go to short-shelf channels: the near DC, foodservice, anything that turns in two days. Not onto a truck crossing three states.
2. Ship order. Move the short-life batches first and hold the clean ones. Pack-date order ignores which batch is actually about to go.
3. Customer tiering. Your longest-life batches go to the accounts that charge back hardest. Fewer rejections at the dock.
4. Water management. Break and change the flume when the load says it is spent, not when a timer says so.
5. Hold. When the pathogen channel says a batch’s remaining safe life is zero, its destination is nowhere.

Food safery verdicts can land ninety minutes into a six-hour window, so you hold a pallet in the cooler instead of chasing it through 27 states.

5 PACKHOUSE IS the quarterback

A quarterback needs two things at once. Information, and choices still open.

Ranches have choices and no information. Retailers have information and no choices. They learn about a bad lot as shrink, at the end, attributed to nobody.

The packhouse plant holds both at once. It’s the only place where every unit converges, stops, gets identified, and still has somewhere else it could go. That window is about six hours for lettuce.

Regulation are already putting the plant in that seat. Under FDA’s proposed Food Traceability Rule, leafy greens sit near the top of the Food Traceability List, and packing is where the traceability lot code gets assigned. Everything downstream inherits identity from that. Compliance was pushed to July 2028, but the commercial deadline passed a year ago, because major retailers started requiring lot-level data in 2025.

So the per-batch record is getting built anyway. This adds biology to that record.

Four things follow, roughly in the order they would actually happen.

1. Sell shelf life instead of lettuce. Today a case ships with a date derived from pack date, which is a guess dressed as a fact. With a measured per-batch estimate you can commit to days-of-life on arrival and stand behind it. Shrink is the buyer’s problem right now, and a supplier who can move that risk onto their own balance sheet, priced, is not selling the same product as the plant next door. You are the only party who can underwrite that contract, because you are the only one that has that information.
2. Turn dispatch into an optimisation. You have batches with estimated lives and orders with known transit times and turn rates. Today the rule is pack date. With remaining life it is an assignment problem, and assignment problems have solutions worth several points of margin. No new equipment, easy money.
3. Call the audible upstream. This is the real power and the least obvious. You are the only entity that measures every grower’s output on the same instrument under identical conditions. The grower sees only their own product. The retailer sees only failures, late, attributed to nobody. So you become the scorer, and the score is actionable agronomy: this block, harvested after eleven, carries a heavier load and delivers four fewer days. Over a season that can change harvest hour, cooling protocol, crew practice, water source. Over several seasons it can change what gets planted where. The plant stops taking whatever the field sends and starts specifying it. This is also part of the model thesis.
4. Survive the incident that takes out everyone else. Yuma resolved to a region. July 2026 resolved to a region and a season. In both cases the remedy was applied at the coarsest available resolution, because when you cannot resolve you have to over-include, and over-including is the responsible choice. A plant with separated batches and per-batch biological records can show which batches were clean. It keeps selling while its competitors are pulled from shelves. That reorganises a category in a season, which is a different order of thing from a cost saving.

And the loop compounds. Every batch you route and then watch is a labelled example. Measured this, sent it there, here’s what happened. Nobody else closes that loop, because nobody else sees both the biology at pack and the outcome at retail. Prediction improves with volume, so the lead widens.

The big fresh-cut processors are already vertically integrated, so for them this is an internal upgrade, not a seizure of anything.

(Grower scoring creates adverse selection. Score growers, pay them less, and they’ll resist the measurement or send weaker lots to whoever isn’t looking. Probably you pay a premium for measured-good product instead of penalising the rest.)

6 Two clocks

Dose sanitizer will still have a hard time working on a 90 minute run-time.

So the real-time kill stays on the fast proxy, ORP dosing in seconds, and the assay runs slow, certifying each batch and gating dispatch. In HACCP language the dosing loop is the preventive control and the assay is verification. 

7 WOULD THIS REMOVE ALL HUMANS?

1. FSMA wants a qualified individual owning the food-safety plan. It doesn’t put them on the floor, and one can cover several sites.

2. Machines break, which is a service contract. 

3. Manual swabbing in the facility itself. A fresh-cut plant is wet, cold, and full of the niches Listeria likes. Crevices, etc. This is probably solvable via robotics.

· the economics

WHAT IS AN AUTONOMOUS PACKHOUSE'S EDGE?