Systems and records
Food quality hold release dataset: logging holds, tests and releases
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Store each hold, sample, result, decision and release as a dated event tied to a lot code, with the hold basis, specification version, evidence links and approver, and never overwrite a corrected record. A food quality hold release dataset built that way can be replayed through traceability links, shows premature releases as deviations, and teaches the reasoning behind quality decisions rather than final statuses.
Checklist
Food quality hold release
Key takeaways
- Record each hold, sample, result, decision and release as a separate dated event tied to a lot code.
- Name the hold basis, because internal, agency and customer holds have different release conditions.
- A closed set of states with legal transitions turns odd histories into findings instead of noise.
- Keep corrected records alongside the originals so the history can be replayed as it was known.
- Label a release logged before results as a deviation even when the result was later acceptable.
Why does a released lot need its whole hold history, not just its final status?
A food operations dataset should store every hold, sample, result, decision and release as its own dated event tied to a lot, with the evidence and the person behind each step. A final status of released tells you the lot left; it does not tell you why it was held, what was tested, whether results were back before release, or who approved it.
That missing middle is what an AI system needs to learn from. A model asked to recommend whether a lot can move, or to draft the disposition note for a quality manager, has to reason about the sequence: the trigger, the containment, the evidence gathered, the rule that applied and the approval. If the dataset only keeps the end state, the model learns that lots usually get released, which is the wrong lesson.
The record described here is a state transition log with links. Each transition names the lot, the from and to state, the reason, the evidence references, the specification version in force, and the approver. Replaying the log from the first event should reproduce the lot's state at any moment, including the moments when it should not have shipped.
Which rules create the hold, and what do they require you to write down?
Holds come from two places: a facility's own food safety plan when something goes wrong, and agency programs that make product wait for test results. Both require records, and both distinguish the immediate fix from the fuller decision about affected food.
Under FDA's preventive controls rule, 21 CFR 117.150 says that when a preventive control is not properly implemented or is found ineffective, the facility must take corrective action that identifies and corrects the problem, reduces the likelihood of recurrence, evaluates all affected food for safety, and, as necessary, prevents affected food from entering commerce. Written procedures must keep affected food out of commerce if the facility cannot ensure it is not adulterated, and corrective actions must be documented in records. Section 117.3 defines a correction as the narrower act of identifying and correcting a production problem without those further steps, so a dataset should never collapse the two into one label.
For meat, poultry and egg products under USDA inspection, FSIS guidance on controlling product pending FSIS test results lets establishments package and label sampled product but not ship it into commerce until acceptable results are available, and tested product moved off-site under the establishment's ownership cannot complete pre-shipment review or change hands while results are pending. FSIS Directive 9900.6, which covers laboratory sampling of imported meat, poultry and egg products, lists where test and hold instructions apply, including non-intact raw beef tested for STEC and ready-to-eat products tested for Listeria monocytogenes or Salmonella.
For the dataset, the practical consequence is simple: the hold reason must name its basis. A hold under the facility's own plan, a hold pending an agency sample and a hold for a customer specification have different release conditions and different evidence.
What should the domain reviewed quality state transition record include?
It should hold one row per transition, with enough linked evidence that a qualified reviewer could reach the same decision without asking anyone. The schema below is written so a food or agriculture quality team can adopt it as is; rename fields to match its ERP, LIMS and quality management system.
| Field | Contents | Typical source | Reviewer check |
|---|---|---|---|
| lot_id and lot_type | Internal lot or traceability lot code, product or ingredient, quantity and unit | ERP or warehouse system | Same code used in every linked record |
| transition | From state, to state, event time with time zone, system that recorded it | Quality or warehouse system audit log | Event time is when it happened, not when it was keyed |
| hold_basis | Own food safety plan, agency sample, customer specification, supplier issue | Hold ticket and plan reference | Basis matches the release condition used later |
| spec_version | Product specification and sampling plan version in force at the event | Specification and document control | Version was effective on the event date |
| samples | Sample IDs, sample time, sampled quantity, laboratory and method reference | LIMS or laboratory chain of custody | Sample links to the held lot, not a neighbor |
| results | Result per analyte or check, result time, reported against which limit | LIMS or certificate of analysis | Result time precedes any release |
| substitutions | Substitute ingredient or supplier lot, who approved it, and the approval date | Change control or deviation record | Approval exists before the substitute is used |
| decision | Release, rework, downgrade, divert, destroy, or keep on hold, with reason text | Disposition record | Decision fits the hold basis and results |
| approver | Role and pseudonymous person key, date of approval | Quality sign-off | Approver had authority for this decision type |
| outcome | Shipment, destruction confirmation, rework lot, complaint or recall link | Shipping, destruction and complaint records | Outcome matches the decision, with evidence |
Which states and transitions should the record allow?
Allow a small, closed set of states and say which transitions are legal; anything else in the history is a finding, not noise. The GS1 Core Business Vocabulary is a useful anchor because it already names many of these conditions. Its disposition list includes in_progress, conformant, non_conformant, non_sellable_other, available, recalled and destroyed, and its business steps include holding, inspecting and sampling. EPCIS 2.0 also allows a persistent disposition that an event can set and a later event can unset, which fits a hold that must stay attached to a lot until a release explicitly removes it.
Mapping your states to that vocabulary is optional, but it makes records from different plants or partners comparable, and it forces a decision about what each local status really means.
| State | Meaning in the record | Closest CBV term | Legal next states |
|---|---|---|---|
| Received or produced | Lot exists and is in the facility's control | in_progress | On hold, Available |
| On hold | Lot may not ship; hold basis recorded | non_sellable_other (persistent) | Sampled, Released, Rework, Destroyed |
| Sampled, results pending | Samples taken and linked; no result yet | non_sellable_other with sampling step | Conformant, Non-conformant |
| Conformant | Results meet the specification version in force | conformant | Released |
| Non-conformant | Results or checks fail the specification | non_conformant | Rework, Diverted, Destroyed, On hold |
| Released | Hold removed by an authorized approver | available | Shipped, Recalled |
| Destroyed or diverted | Lot left food use with confirmation | destroyed | None |
Can you replay a lot's quality decisions through its traceability links?
Yes, if every transition carries the lot identifier the traceability records use and every transformation links its input lots to its output lots. Replay means starting at a finished lot, walking back through transformations to the ingredient lots, and seeing each quality decision on the way, in time order.
FDA's Food Traceability Rule gives the backbone for foods on its Food Traceability List. Covered persons keep records of key data elements tied to critical tracking events such as shipping, receiving and transformation, the traceability lot code must appear in the records at each event once assigned, and records must be provided to FDA within 24 hours of a request or a reasonable time FDA agrees to. A hold and release history that uses the same lot code can be joined to those events instead of living in a separate spreadsheet.
Corrections need care. When a quality record turns out to be wrong, do not overwrite it. EPCIS models this with an error declaration on the original event, giving a reason such as incorrect_data or did_not_occur and pointing to the corrective event by its ID. A replayable dataset keeps both, so a model can learn that the first record existed and was later corrected, and a grader can tell which version an agent should have relied on at a given time.
Worked example: a ready-to-eat lot held for a pathogen test, with a substitute ingredient
This example is illustrative: the lot, times and counts are invented to show the record in use, and it is not guidance on how to run a test and hold program. A plant produces a ready-to-eat lot of 1,200 cases. The lot is sampled for Listeria monocytogenes under an agency program, so its hold basis is agency sample, test and hold. One seasoning was replaced earlier that day with an approved substitute from a second supplier because the usual lot was short.
Starting state: lot produced at 06:10, placed on hold at 06:15 with the agency sample basis, sample ID linked at 06:40, specification version 7 in force. Linked evidence: the sample chain of custody, the change control record approving the substitute seasoning at 05:30 by the quality manager, the substitute's supplier lot and certificate of analysis, and the production record showing which mixer batches used it.
Decision path: the result arrives two days later as negative, which the laboratory reports against the program's criterion. The lot moves from sampled to conformant, and the quality manager releases it the same afternoon with a note citing the result, the substitute approval and the spec version. Verified outcome: the shipping record shows 1,200 cases shipped after the release time, and no complaint or recall link appears in the following 90 days of records.
Exception: an otherwise identical lot from the next day shows released in the ERP at 14:00 while the laboratory result time is 16:30. The correct label for that case is a deviation, released before results, even though the result was later negative. A dataset that labels it a normal release teaches the opposite of the rule, and a model trained on it would treat a premature release as acceptable because the outcome happened to be fine.
Second exception: had the substitute seasoning been used at 05:00, before its 05:30 approval, the release decision would rest on an unapproved input. The case then belongs with the quality and food safety reviewers, not in a set of clean releases.
What separates a defensible release from a lucky one when grading?
A defensible release has its evidence in place, in the right order, approved by someone with authority; a lucky one merely ended well. Grade the reasoning and the sequence, because outcome-only grading rewards the premature release in the example above.
| Graded element | Pass condition | Evidence | Typical false pass |
|---|---|---|---|
| Hold basis identified | Basis matches the trigger and the release condition | Hold ticket, plan or program reference | Treating an agency hold like an internal one |
| Sequence | Every result time precedes the release time | LIMS timestamps, disposition record | Release logged before results, outcome later fine |
| Specification version | Decision uses the version in force on the event date | Document control history | Applying the current spec to an old lot |
| Inputs approved | Substitutes and supplier changes approved before use | Change control, production record | Approval dated after first use |
| Correction versus corrective action | Labels follow the 117.3 distinction | Deviation and corrective action records | One label for a quick fix and a full evaluation |
| Outcome linked | Shipment, destruction or rework confirmed by records | Shipping, destruction and rework records | Status of released used as the outcome |
Which hold cases belong with a food safety specialist, not in a general dataset?
Keep confirmed pathogen positives, recalls, regulatory correspondence and anything involving illness complaints out of a general training set unless a food safety specialist and counsel have reviewed the case and agreed on how it may be used. These cases are valuable but carry legal, reputational and supplier confidentiality weight that ordinary holds do not.
Narrow the scope rather than redact heavily. A set limited to one product family, one plant and one hold basis is easier to review and easier for a buyer to evaluate than a mixed export with gaps. Supplier names, specification limits and recipes may be confidential under supply agreements, so replace them with stable keys and keep the limit as pass or fail when the exact value is not needed for the task.
The dataset is for teaching and testing models to reason about records. It should never become the system that decides whether real product ships; that decision stays with the facility's qualified people under its food safety plan and any agency program. Rights to license the records are a separate question from whether they can be shared at all, and de-identification does not create that right.
- Refer to a food safety specialist: positive pathogen results, allergen cross-contact events, and any lot linked to a recall.
- Refer to counsel: records tied to agency correspondence, complaints or supplier disputes.
- Narrow the set: mixed product families, plants with different specification systems, or periods with incomplete laboratory links.
- Exclude: records that cannot be tied to a lot code or that lack an approver.
Licensing quality decision records through SourceX
Years of hold, test and release records with their evidence attached are hard to find outside food plants, which is why they interest teams building quality and operations models. A food or agriculture business can license selected quality records through SourceX, which covers sourcing, rights review, contracting, preparation, delivery and payment.
A food or agriculture business can begin with a short company assessment without uploading data. Nothing is delivered until the company has approved the scope and terms and signed the agreement.
Frequently asked questions
What is a food quality hold release dataset?
It is a log of lots moving through hold, sampling, results, decisions and release or disposal, with each step linked to its evidence and approver. It lets a model learn and be tested on the reasoning behind quality decisions rather than on final statuses.
What is the difference between a correction and a corrective action?
In FDA's preventive controls rule, a correction identifies and fixes a production problem. A corrective action goes further: it also reduces the chance of recurrence, evaluates affected food for safety and keeps affected food out of commerce as necessary.
Can test and hold product be labeled before results come back?
FSIS guidance allows establishments to package and label sampled product before results return, but not to ship it into commerce until acceptable results are available. The dataset should record the label time and the ship time separately.
Should corrected quality records be overwritten?
No. Keep the original record and the correction, linked to each other. EPCIS does this with an error declaration on the original event that points to the corrective event.
Do I need GS1 standards to build this record?
No. The CBV vocabulary is a convenient reference for naming states such as conformant and non_conformant, but any closed set of states works if the legal transitions are written down and applied consistently.
Sources
- 21 CFR 117.150 requires corrective actions that evaluate affected food, keep it out of commerce when safety cannot be ensured, and are documented in records. Source
- 21 CFR 117.3 defines a correction as fixing a production problem without the other steps of a corrective action procedure. Source
- FSIS guidance lets establishments package and label sampled product but not ship it until acceptable test results are available. Source
- FSIS Directive 9900.6 lists products to which test and hold instructions apply, including STEC and Listeria or Salmonella testing. Source
- FDA's Food Traceability Rule requires key data elements for critical tracking events, the traceability lot code at each event, and records within 24 hours of a request. Source
- The GS1 CBV disposition vocabulary includes in_progress, conformant, non_conformant, non_sellable_other, available, recalled and destroyed. Source
- The GS1 CBV business step vocabulary includes holding, inspecting and sampling. Source
- EPCIS 2.0 events can set and later unset a persistent disposition on the same objects. Source
- EPCIS error declarations mark an event as incorrect and point to the corrective event by ID. Source
See if your company qualifies
A short company assessment. No data uploads are needed.