A validated allergen clean is a different class of event from a colour change or a size change, because it carries a verification step and a hold for the result, and a plant recording one average changeover time cannot see it. The planning object is a transition matrix between allergen profiles rather than between SKUs, which on a fourteen product line usually collapses to five profiles and twenty ordered transitions. Ascending allergen load through a cycle puts one validated clean at the end, which then fixes the cycle length and the minimum campaign for the free-from range. Carry-over travels through rework and shared upstream equipment, neither of which appears in the sequence the scheduler is optimising.
It is Thursday and a customer has pulled forward an order for a mustard dressing on the sauces line. The scheduler agrees, because the customer is large and the line has the hours. What the schedule does not show is that the pull inserts the mustard run between two free-from products, so the week now contains three validated cleans instead of one. Each is five hours and twenty minutes of wet clean plus a hold for the swab result. The line loses most of a shift, the Friday orders slip, and the weekly review records it as a capacity problem.
Nobody made a mistake in the sense the plant would recognise. The schedule was built against due dates and a single average changeover figure of fifty five minutes, and that figure is the mean of two populations with nothing in common.
What a validated clean actually costs
Transitions on a food line fall into three classes, and they differ by an order of magnitude.
A push-through or dry clean applies where the allergen profile does not change or gets wider. Running a plain base before a base with added milk means the milk is being introduced rather than removed, so the transition needs a product flush and a line clear. Twenty to forty minutes on most wet fill lines.
A wet clean without allergen verification applies where the profile is unchanged and the concern is quality: colour, flavour carry, viscosity. An hour to ninety minutes, depending on how much of the line has to be broken down.
A full validated clean applies when an allergen leaves the line. The equipment is stripped, cleaned to a validated procedure, inspected, and then verified before the next product runs. Four to seven hours is the normal band on a line with pumps, plate heat exchangers and a filler, and the tail of that distribution is long because the verification can fail.
The verification is where schedules go wrong, because it is a wait rather than work. Visual inspection is immediate. ATP swabs return in a few minutes but measure organic residue rather than the allergenic protein, so they confirm that cleaning happened without confirming that the allergen is gone. An allergen-specific lateral flow strip gives a result in about ten minutes and is the usual production floor tool. A laboratory ELISA is more sensitive and takes hours, which means it belongs in periodic validation rather than in every transition.
The Codex Alimentarius Commission's Code of Practice on Food Allergen Management for Food Business Operators, adopted in 2020, sets the expectation plainly: cleaning between allergen and non-allergen production is validated once and verified every time. The validation fixes the procedure and the time. The verification adds a hold with a small but real chance of a re-clean behind it.
A plant that records line stop to line start on a single changeover field has no way to distinguish the twenty five minute case from the six hour case, and the schedule inherits that blindness.
Build the transition matrix before arguing about the sequence
The instinct is to build a matrix over products. On a fourteen SKU line that is 182 ordered pairs, which is enough work that the project dies.
The matrix that matters is over allergen profiles. Group the SKUs by the set of regulated allergens each contains, and the fourteen usually collapse to four or five groups: free-from, milk, milk and soy, mustard and celery, sesame. Five profiles give twenty ordered transitions, which one quality manager and one scheduler can populate in an afternoon.
Two properties of that matrix drive everything downstream. It is asymmetric, because adding an allergen is a flush and removing one is a validated clean. It is also sparse in cost terms, with most cells cheap and a few carrying nearly all the hours. On a typical week with nine changeovers, two or three cells account for seventy to eighty per cent of the changeover time.
Sequence-dependent setup scheduling has a long literature and the survey to start from is Allahverdi, Ng, Cheng and Kovalyov in the European Journal of Operational Research in 2008, which classifies the problem forms and their solution approaches. The point that survives contact with a food plant is that when setup times depend on the preceding job, optimising the order is a different problem from optimising batch sizes, and an average setup time answers neither.
Which allergens count is a regulatory question with a different answer in each market. The EU lists fourteen in Annex II of Regulation 1169/2011, including celery, mustard, lupin and sulphites, which do not appear on the United States list. The FASTER Act of 2021 added sesame as the ninth major allergen in the United States from January 2023. A plant exporting to both writes its matrix against the union of the lists, and a plant that inherited its matrix before 2023 should check whether sesame is still being treated as an ingredient rather than a transition.
Ascending allergen load, and where the cycle breaks
The standard answer is to run the cycle in ascending order of allergen load. Free-from products first, then single allergen products, then the multi-allergen ones, with the single validated clean at the end of the cycle before the range starts again. Every transition inside the cycle adds allergens, so every transition inside the cycle is a flush.
That works, and it has two consequences that get discovered later.
The first is that cycle length is now set by the clean rather than by demand. One validated clean per cycle at five and a half hours is a manageable weekly cost. Two cycles a week doubles it. The scheduler who wants to be responsive to a customer is asking for a second clean, and the honest way to have that conversation is with the hours on the table rather than as an argument about flexibility.
The second is that the free-from range gets made once per cycle, in a batch sized to cover the whole cycle. If free-from is eight per cent of volume on a weekly cycle, the batch is small and awkward and the line runs it inefficiently. If the cycle stretches to a fortnight to save the clean, the free-from stock built on day one has to last twenty eight days, and the shelf life consequences of that decision are covered in AQ2. Batch sizing against the new setup cost belongs to EE3, and what the recovered hours are worth in money belongs to EE1.
The cycle also breaks in a way that is specific to allergens rather than to setups generally. A production wheel assumes the range is stable. Introduce one new SKU with an allergen profile that sits between two existing groups and the ordering is no longer total, so the ascending rule has no unique answer and someone has to decide which clean to pay. That decision should be made once and written into the matrix rather than re-argued every week by whoever is on shift.
Carry-over travels through rework and through the room
Sequencing controls the line. It does not control two other paths that carry allergens into product, and both of them are invisible in the schedule.
Rework is the first. A batch of milk-containing sauce that fails on viscosity gets held and blended back into a later batch at five per cent. If the rework record in the ERP carries a generic material code, the allergen profile of that rework is lost, and the later batch inherits an allergen its specification does not declare. The fix is unglamorous and effective: rework materials carry their own allergen profile, and the planning system treats a rework material as ineligible for any product whose profile does not already contain it. Most systems can hold that as a simple restriction once someone decides to maintain it.
Shared upstream equipment is the second. Mixers, sieves, hoppers, conveying air and dust in a dry blending area move allergens between products that never share a filler. A schedule built around the filling line can be entirely correct and still be wrong, because the constraint lives two stages upstream where nobody models capacity.
There is a third path worth naming even though sequencing cannot help with it. A large share of allergen incidents originate in packaging rather than in cross-contact, where the wrong film or the wrong label runs on a correct product. FDA's Reportable Food Registry summaries have consistently placed undeclared allergens at the top of the reported hazard categories, and label control at changeover is a repeated root cause. Packaging material planning is VV4's subject and traceability after the event is QQ4's, but the scheduling implication is direct: a changeover is the moment when both the product and the packaging change, and compressing changeover time by removing verification steps increases the probability of the more common failure.
What precautionary labelling buys and what it costs
Precautionary allergen labelling, the may contain statement, gets decided by the quality and commercial functions, and it arrives in the plan as a constraint that has been removed.
If a product carries a may contain declaration for milk, the transition from a milk product into it stops requiring a validated clean and becomes a rinse. On a line with a mixed range, applying precautionary labelling across the board can remove nearly all the expensive cells from the matrix. That is why it is tempting, and why it spreads.
The cost is commercial and it is growing. Retailer own-label specifications increasingly forbid precautionary labelling on products positioned as free-from, so applying it removes those listings. The FAO and WHO expert consultation on risk assessment of food allergens published reference doses in 2022, and the Allergen Bureau's VITAL 3.0, released in 2019, turns a measured cross-contact level into a labelling decision against those doses. Both push the same way: a precautionary statement should follow from a quantitative assessment rather than from convenience, and blanket application is becoming harder to defend.
For a planner the practical effect is that precautionary labelling policy is an input to the schedule and belongs in the matrix as a flag per SKU. Products where the customer forbids it must be scheduled in a verified clean window. Products where it is permitted can sit anywhere in the cycle. Where the customer set differs by retailer, F8 covers how those specification differences propagate.
Where this stops
The clean time in the matrix is a quality output and a scheduler cannot optimise it downwards. If the validated procedure says five hours and twenty minutes, five hours and twenty minutes is what the plan uses until the procedure is revalidated. Everything here arranges work around a fixed cost rather than reducing it, which is a different exercise with a different owner.
Matrices also go stale. A recipe change, a new piece of equipment, a change in the cleaning chemical or a new allergen on the regulated list all invalidate cells. A matrix that has not been reviewed since it was built is a liability, because it is trusted and wrong. Annual review against the current specification list is the minimum, and any recipe change should trigger a check of the affected rows.
The approach also earns less on simple lines. A plant with three allergen profiles and two changeovers a week has perhaps forty minutes of sequencing benefit available, and the effort belongs elsewhere. The gains concentrate above four profiles and five changeovers a week.
Exact sequencing gets hard at scale. Ordering fourteen jobs with sequence-dependent setups is small enough to solve exactly. Ordering two hundred SKU and line combinations across a multi-line site is not, and the practical route there is a heuristic, which J21 covers.
The last limit is the one worth being blunt about. Better sequencing reduces the number of times you pay for a clean, and it does nothing about whether any given clean worked, so the residual food safety risk stays with verification quality rather than moving to the plan.
Take one line for an afternoon, group its SKUs into allergen profiles, fill in the ordered transition times with the quality manager signing each cell, and then count how many hours of last month's changeover total sat in the three most expensive cells.
More on this in industry on the DemandLabs platform.
