AI uses for records
Food production rework dataset: the decisions a clean log misses
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Rework records reveal the exception decisions a clean production log hides: whether held product was eligible for reuse, which batch could take it and how much, whether allergens matched, who approved it and whether the finished lot was released. A food production rework dataset links hold, cause, disposition, receiving batch, traceability chain and release outcome for each case.
Record timeline
Food production rework
Key takeaways
- Clean production logs show the happy path; rework records show how a plant handles exceptions.
- Rework has a narrow regulatory meaning, so eligibility is the first field to annotate.
- Allergen matching and specification limits decide most dispositions, not yield.
- A case is usable for grading only when it replays from source lot to released finished lot.
- Recall, pathogen and inspection-linked cases stay out; unmatched allergens go to a food safety reviewer.
What decisions do rework records reveal that a clean production log misses?
Rework records show the judgment calls a production log hides: whether off-spec product was safe to reuse at all, which later batch could take it, how much of it, under what allergen and specification limits, who approved that, and whether the finished lot was released. A clean log records only that a batch ran and met target; the rework trail records why some of it nearly did not.
That difference is what makes a food production rework dataset useful for AI work. A model that only sees finished batches learns the happy path. A model that sees the hold, the disposition, the substitute use and the release decision learns how a plant handles exceptions, which is where planning, quality and traceability agents struggle.
This article treats rework as administrative and operational records. It does not give food safety advice, and every disposition rule mentioned here belongs to the plant's own food safety plan and its qualified staff.
What counts as rework, and what does not
In U.S. federal rules for human food, rework has a narrow meaning. 21 CFR 117.3 defines it as clean, unadulterated food that has been removed from processing for reasons other than insanitary conditions, or that has been successfully reconditioned by reprocessing, and that is suitable for use as food. Product that does not meet that definition is not rework, whatever the plant calls it on the floor.
The same part sets handling expectations. Section 117.80 says material scheduled for rework must be identified as such, that rework must be held in bulk or in containers that protect against allergen cross-contact and contamination, and that rework containing food allergens must be identified and held in a manner that prevents allergen cross-contact.
Rework also differs from a correction and a corrective action. Section 117.150 describes corrective action procedures that identify and correct a problem with a preventive control, reduce the likelihood of recurrence where needed, evaluate all affected food for safety and prevent affected food from entering commerce if its safety cannot be assured. A rework decision can sit inside that process, but the dataset should label which one it is.
For annotation, this gives three questions to answer before anything else: was the material eligible to be rework, was it identified and held as rework, and was its use tied to a corrective action? A record that cannot answer the first question belongs in a different category.
Where rework decisions are recorded in a plant's systems
Rework decisions are scattered across production, quality and inventory records, and only linking them shows the decision. Each record below holds one piece; none holds the whole case.
Odoo's public documentation illustrates the pattern in one system. Its quality alert form includes a Root Cause field and separate Corrective Actions and Preventive Actions tabs; scrapped items move to a virtual scrap location so losses are logged without affecting real stock levels; and unbuild orders return components to inventory while keeping component and finished counts accurate. Other ERP and MES products use different names for the same ideas.
| Record | What it shows | Typical system | Linking key |
|---|---|---|---|
| Batch or production record | Planned and actual inputs, including any rework added | MES, ERP manufacturing order | Batch or lot number |
| Hold or nonconformance record | Why product was stopped and its status while held | Quality system, hold tag log | Hold ID and source lot |
| Quality alert or deviation | Root cause, corrective and preventive actions | Quality app or deviation log | Alert ID and lot |
| Disposition decision | Rework, release, downgrade, scrap or destroy, with approver | QA disposition form | Hold ID |
| Rework inventory movement | Where rework was stored and when it was consumed | WMS or ERP stock moves | Rework lot or container ID |
| Scrap or unbuild entry | Material written off or broken back into components | ERP scrap and unbuild orders | Source order and lot |
| Traceability record | Input lots and new lot codes created by transformation | Traceability or ERP lot genealogy | Traceability lot code |
| Release or test record | Whether the receiving batch was released, and on what evidence | Lab system, release log | Finished lot number |
The rework cause, action and outcome annotation guide
The guide is the asset. Each row is one field an annotator fills for every rework case, with the allowed values, the evidence required and the rule for deciding. A case is one batch of material set aside for rework and everything that happened to it.
At a glance
The rework cause, action and outcome annotation guide
Keep the specification version with the case. Recipes, rework limits and allergen matrices change, and a decision that was right under last year's specification can look wrong under this year's.
| Field | Allowed values | Evidence required | Annotation rule |
|---|---|---|---|
| Case identity | Rework case ID, source lot, product family, plant line pseudonym | Hold record and batch record | One case per set-aside quantity, never per shift |
| Eligibility | Eligible rework, not eligible, unknown | Hold reason and QA note against the plant's rework policy | Unknown stays out of grading sets |
| Cause | Packaging fault, weight or fill, appearance, process deviation, labeling error, changeover surplus, other | Quality alert root cause, operator note | One primary cause; secondary causes listed separately |
| Allergen profile | Allergens present in the rework, matched or not to receiving product | Formula and allergen matrix version | Unmatched allergen use is always flagged for a food safety reviewer |
| Identification and holding | Labeled and held as rework, labeled but held improperly, not labeled | Hold tag, location record | Record what the evidence shows, not what policy requires |
| Action | Reworked into product, reconditioned, downgraded, scrapped, unbuilt, destroyed | Disposition form, stock moves | Use the action that was executed, not the one proposed |
| Receiving batch and quantity | Receiving lot, rework quantity as a share of batch, specification limit | Batch record and specification version | Flag any share above the specification limit |
| Approval | Role of approver, date, specification version cited | Disposition sign-off | Roles only; remove names |
| Traceability link | Input lots, rework lot, new lot code if one was assigned | Lot genealogy or traceability records | The chain must replay from source lot to finished lot |
| Verified outcome | Released, held again, rejected, scrapped later, complaint linked, unknown | Release log, later hold, complaint record | Outcome comes from a record dated after the action |
Can a rework case be replayed from source lot to released finished lot?
A rework case is only usable when a reviewer can replay it from source lot to finished lot and see every quality decision on the way. If the chain breaks, the case can still teach cause classification, but not disposition decisions.
Traceability rules make this concrete for some foods. Under 21 CFR 1.1320, a firm must assign a traceability lot code when it transforms a food on FDA's Food Traceability List, and must not set a new code for other activities such as shipping. Rework that is blended into a new batch of a listed food is a transformation event, so the replay should find the input lot codes and the new code in the plant's traceability records. For foods not on the list, the plant's own lot genealogy plays the same role. FDA has said it will not enforce the traceability rule before July 20, 2028, following a congressional directive, so a plant's records may not follow it yet.
Records matter on the quality side too. Section 117.190 lists records documenting corrective actions among the implementation records a facility must keep, and section 117.150 says corrective actions, and corrections where appropriate, must be documented. Where rework came out of a corrective action, that documentation is part of the chain.
- Start from the hold: source lot, quantity, cause and time stopped
- Find the disposition: action, approver role, specification version
- Follow the stock: rework location, consumption date, receiving batch
- Check the code: input lots and new lot code on the finished product
- Confirm the release: test or release record dated after blending
- Close the loop: any later hold, rejection or complaint on the finished lot
Worked example: granola bars with failed seals (illustrative)
This example is illustrative. The plant, product and counts are invented to show how the guide is filled, and the numbers are not recommendations.
Starting state: a packaging line producing an oat and almond granola bar has a seal fault for 40 minutes. The quality technician places the affected output on hold. The bars themselves are intact and were never exposed to an insanitary condition, so the QA lead classifies the unwrapped bars as eligible rework under the plant's rework policy and records seal failure as the root cause in a quality alert.
Action: the bars are unwrapped, ground and labeled as rework containing tree nuts. Two days later, 3 percent of a new batch of the same oat and almond formula is made up of rework, inside the plant's illustrative 5 percent specification limit. The batch record shows the rework lot consumed, and the traceability record lists the source lot and rework lot as inputs to the new lot code.
Verified outcome: the finished lot is released after its routine checks, and the release record is dated after blending. The case is graded as eligible, correctly identified, allergen matched, within limit, traceable and released.
Exception: the same rework is proposed for a plain oat bar run on the next day to use it up quickly. The plain bar does not contain tree nuts, so the allergen profile is unmatched. The correct label is not eligible for that receiving product, and the right action is to wait for a matching run or scrap the material. A model that recommends the faster blend because the specification share is within limit has missed the allergen boundary, which is the decision the case exists to test.
- Cause: packaging fault (seal failure), recorded in a quality alert
- Eligibility: eligible; identification: labeled as rework containing tree nuts
- Action: reworked into same formula at an illustrative 3 percent share
- Outcome: released, with release record dated after blending
- Exception: unmatched allergen in receiving product, so wait or scrap
Grading a model's rework disposition
Grade the disposition against the rules and evidence, then check the outcome separately. A disposition that broke an allergen or eligibility rule is wrong even if the finished lot happened to be released without issue.
A practical rubric has five checks: eligibility judged from the hold reason, allergen profile matched or the case routed to a reviewer, quantity within the specification limit, traceability chain intact, and action matching the approved disposition. The outcome check then asks whether the finished lot was released and stayed released.
Avoid grading on scrap avoided or yield recovered. Those numbers reward reuse, and the cases that matter most are the ones where scrapping was the right decision. A balanced set includes rework, downgrade and scrap outcomes in proportions that reflect the plant's real history, not only the successful blends.
Rework cases that need a food safety reviewer or should stay out
Some rework cases should never leave the plant, and others need a qualified food safety reviewer before they are used for anything. Decide per case, using the eligibility and allergen fields.
Keep out cases linked to recalls, pathogen findings, regulatory inspections or ongoing investigations. Route any case with an unmatched allergen, a process deviation affecting a preventive control or an unknown eligibility status to the plant's preventive controls qualified individual or equivalent reviewer before it is labeled.
Check ownership as well as privacy. A co-packer may hold batch records for formulas that belong to the brand owner, and supply agreements may restrict how those records are used. Removing operator names addresses one privacy issue, but it does not give the plant the right to license the records; access to the system, privacy permission and the contractual right to license are separate questions for counsel to review.
- Recall, pathogen or inspection link: exclude
- Unmatched allergen or preventive control deviation: food safety reviewer first
- Customer-owned formula or specification: confirm rights under the supply agreement
- Operator and approver names: replace with roles
- Specification version unknown: keep for cause analysis only
Licensing rework and quality decision histories
Linked rework histories are uncommon because the pieces sit in different systems. A food manufacturer that can assemble them in the shape above holds decision records that AI developers building planning, quality and traceability tools find hard to get elsewhere.
A food manufacturer that wants to license its rework and release history can do so through SourceX, which runs the sourcing, rights review, contracting, preparation, delivery and payment. On volume, SourceX's general guide is that buyers want years of consistent plant records, often two or more, numbering in the tens of thousands within a single system; a rarer process can qualify with less, and each plant's case is weighed separately.
Batch and quality records leave the plant only after the company approves scope and terms and signs, and the opening step is a brief assessment that needs no batch data.
Frequently asked questions
What is a food production rework dataset?
It is a set of rework cases from a food plant, each linking the held material, its cause, the disposition, the receiving batch, the traceability chain and the release outcome. Its value comes from the decisions it records, not from the production volumes.
How is rework defined in U.S. food regulations?
21 CFR 117.3 defines rework as clean, unadulterated food removed from processing for reasons other than insanitary conditions, or successfully reconditioned by reprocessing, that is suitable for use as food. Material that does not meet that definition should not be labeled rework in a dataset.
Why do allergens matter so much in rework records?
Rework carries the allergens of the product it came from. Federal rules require rework containing food allergens to be identified and held in a way that prevents allergen cross-contact, so a dataset should record the allergen profile and whether it matched the receiving product.
Does blending rework into a new batch require a new traceability lot code?
For foods on FDA's Food Traceability List, the traceability rule requires a traceability lot code when a food is transformed, though FDA has said it will not enforce the rule before July 20, 2028, and blending rework into a new batch is generally a transformation. The plant's food safety and regulatory staff should confirm how the rule applies to its products.
Can rework records be shared without exposing customer formulas?
Sometimes. Formulas, specifications and batch records may belong to a brand owner under a co-packing or supply agreement, so the plant should confirm its rights first. Product families and specification versions can often be generalized, but counsel should review the agreements before any sharing.
Sources
- 21 CFR 117.3 defines rework as clean, unadulterated food removed from processing for reasons other than insanitary conditions, or successfully reconditioned by reprocessing, that is suitable for use as food. Source
- 21 CFR 117.80 requires material scheduled for rework to be identified as such, rework to be held so as to protect against allergen cross-contact and contamination, and allergen-containing rework to be identified and held to prevent cross-contact. Source
- 21 CFR 117.150 describes corrective action procedures that identify and correct a problem, reduce recurrence, evaluate affected food for safety and keep it from commerce if safety cannot be assured, and requires corrective actions to be documented. Source
- 21 CFR 117.190 lists records documenting corrective actions among the records a facility must keep to document implementation of its food safety plan. Source
- 21 CFR 1.1320 requires a traceability lot code when a firm transforms a food on the Food Traceability List and bars new codes for other activities such as shipping. Source
- Odoo's quality alert form includes a Root Cause field and Corrective Actions and Preventive Actions tabs. Source
- Odoo moves scrapped items to a virtual scrap location so losses are logged without affecting real stock levels. Source
- Odoo unbuild orders dismantle manufactured products and return components to inventory while keeping counts accurate. Source
- FDA says the Food Traceability Rule's compliance date was proposed to move to July 20, 2028, and that it will follow a congressional directive not to enforce the rule before that date. Source
See if your company qualifies
A short company assessment. No data uploads are needed.