Short answer
Defect type, severity and disposition answer three different questions: what is it, how bad is it against a standard, and what was done with the part. Datasets that fold them into one label cannot be reused when a standard, customer or product class changes. Specify each field separately, record the standard and product class that defined it, and keep the decision-maker's role.
Three labels that are often merged
A line records "solder bridge" in the defect field. Is that the type, the grade or the outcome? In many production records it is a bit of all three, and the field gets used differently by each shift.
Separating them lets you ask separate questions of the data. Type supports detection and classification. Severity supports grading and thresholding. Disposition supports learning decisions, and it also carries business context: a minor defect may be scrapped because the lot is late and rework is too slow, and a major one may be accepted by a customer waiver.
This page is the specification for the severity and disposition layers. The type layer is covered in defect taxonomy and label schema, and the dependency on written criteria in acceptance standards as label definitions.
Why the standard has to travel with the label
Severity is only meaningful relative to a rule. IPC-A-610, the visual acceptability standard for electronic assemblies, defines three product classes: Class 1 general, Class 2 dedicated service and Class 3 high performance, the strictest (IPC, 2024). The same condition on an assembly can pass in one class and fail in another. A dataset that records "reject" without the class cannot be reused for a buyer whose product sits in a different class.
Sampling context matters too. ISO 2859-1 sets out acceptance sampling by attributes indexed by acceptance quality limit (AQL) for lot-by-lot inspection, and applies to finished products, components, materials, records and procedures (ISO, 2026). A lot-level accept or reject under a sampling plan is a different event from a part-level disposition. Do not mix them in one column; see AQL sampling context.
What these labels can and cannot support
They can support severity grading models, threshold calibration, cost-sensitive decision learning, disposition prediction and evaluation of whether a model's decision matches a quality engineer's.
They do not support, on their own:
- Detection or localization, which need images and defect type or masks; see pixel-level defect segmentation.
- Cause analysis. A scrap decision does not explain why the defect occurred.
- Transfer to a different standard. Labels defined by one customer's specification may be wrong for another.
- Physical handling. Moving rejected parts to a bin is a manipulation task that needs robot data, not grade labels.
Illustrative specification
All values are examples the buyer replaces.
# Illustrative: severity and disposition fields, synthetic codes
defect_type: # what is it; owned by the taxonomy
code: "SOLDER_BRIDGE"
severity: # how bad, relative to the standard
scale: [cosmetic, minor, major, critical]
definition_source: "internal_spec_rev_C"
product_class: "class_2"
measured_value: {name: "bridge_length_mm", value: 0.4}
threshold_value: 0.0
disposition: # what was done
value: "rework" # accept | rework | scrap | use_as_is | hold | return_to_vendor
level: "part" # part | lot | sample
authority_role: "quality_engineer"
deviation_waiver: false
decided_at: "2025-02-11T09:14Z"
context:
line_id: "line_B"
lot_id: "lot_5521"
inspector_role: "operator"
The fields do the work. definition_source and product_class tie severity to a rule. level stops lot-level sampling results being read as part-level calls. deviation_waiver flags dispositions that departed from the standard. authority_role tells you whether a person with authority to overrule made the call.
Parameters the buyer decides
| Parameter | Options | Consequence |
|---|---|---|
| Severity scale | 3 to 5 ordered levels, or numeric | Fewer levels agree better; more levels capture nuance |
| Anchor definition | Written criteria with example images | Without anchors, labels drift between raters |
| Disposition set | Accept, rework, scrap, use-as-is, hold | Match the buyer's decision space |
| Level | Part, lot or sample | Keep separate columns |
| Measured values | Required, optional or none | Measurements make labels auditable |
| Multi-defect parts | Worst defect, all defects or per-defect | Decide before collection |
| Rater roles | Operator, inspector, engineer | Affects agreement and trust |
Acceptance tests
Thresholds are for the buyer to set.
- Field completeness. Share of records with type, severity and disposition all present, reported separately.
- Consistency with the standard. Re-grade a stratified sample against the written criteria; report disagreement by severity band. Methods are in inter-inspector agreement and label noise.
- Monotonicity checks. Larger measured size should not map to lower severity without a note.
- Disposition plausibility. Cross-tabulate severity against disposition; unexplained "critical accepted" or "cosmetic scrapped" rows need reasons.
- Standard versioning. Every record resolves to a standard revision and date.
- Borderline coverage. Include samples close to each threshold; see borderline samples.
- Hold-out. Reserve lots or lines for evaluation, never mixed into training.
Limitations and additional preparation
Existing archives rarely have all three fields cleanly. The usual route is to start from existing dispositions and have trained reviewers regrade severity against written criteria on images, which is annotation work performed on rights-cleared data. The annotators supply labels; they do not create rights in the underlying images. Legacy cleanup is addressed in relabeling legacy defect archives.
Dispositions reflect commercial decisions that no relabel can undo. A model trained on them learns the plant's behavior, which may be what you want or may not. Customer part images carry design and relationship information; see confidentiality in defect data.
Using the labels downstream
The three fields feed different models. A detector uses type and location. A grader uses severity, ideally with measured values as auxiliary targets. A decision model uses disposition, but should be trained knowing that disposition depends on factors outside the image, such as waivers, lot pressure and customer. Training a decision model on raw dispositions without those context fields teaches it the plant's habits.
When reporting results, report them by field. A model with strong type accuracy and weak severity accuracy is a different problem from one with the reverse, and a single "defect accuracy" number hides it. Where labels come from several customers with different standards, report by customer or product class and do not pool them without a mapping.
Also plan for revision. Standards change, so keep the label fields stable and store the definition separately. Then a revised standard means regrading with a new definition_source, not replacing the dataset. This keeps old and new gradings side by side for comparison.
Questions to ask suppliers
- Are defect type, severity and disposition stored in separate fields?
- Which standard, product class and revision defined severity, and does it vary by customer?
- Is disposition recorded at part, lot or sample level?
- Who decided: operator, inspector or engineer, and could the decision be overruled?
- Which dispositions were waivers or use-as-is decisions?
- Are measured values stored with the severity call?
- How many records have been regraded by a second person?
- Can you provide a sample extract with synthetic or redacted identifiers so I can test the schema against mine?
Discuss your requirements
If you need severity and disposition labels that survive a change of standard, tell SourceX what your evaluation set needs to cover, including the standard, product class and decision set. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. Annotation can be arranged case by case, and a request does not guarantee a match. See also the humanoid training data guide, the inspection, defect and failure data hub, and the page on inspection images linked to outcomes.
Tell SourceX what your evaluation set needs to cover
Sources
- IPC-A-610 Acceptability of Electronic Assemblies — three product classes with different acceptance strictness.
- ISO 2859-1 Sampling procedures for inspection by attributes - Part 1: Sampling schemes indexed by acceptance quality limit (AQL) for lot-by-lot inspection — AQL-indexed attribute sampling schemes.