Short answer
A nonconformance report (NCR) records what was found wrong, where, on which lot, how it was judged, and what was done about it. As text with attached images, NCR archives can support defect reasoning, root-cause language and disposition models. They contain no images unless attached, no robot actions, and are often heavy with customer and supplier names that need careful handling.
What an NCR archive actually contains
Manufacturers record nonconformances because their quality system requires it. ISO 9001:2026, the sixth edition published on 16 September 2026, sets generic requirements for quality management systems, including documented information, performance evaluation and improvement (ISO, 2026). Organizations run those requirements through their own forms and software, so no two NCR archives share a schema.
A typical record narrates a deviation: part number, lot, quantity affected, who found it, where in the process, a description, a disposition (use as is, rework, scrap, return to supplier), containment actions and a corrective action reference. That mix of structured fields and free text is the asset. The images, measurement logs and drawings that explain the narrative usually sit in other systems and have to be linked deliberately.
This page covers sourcing and assessment. For the text itself see inspection report text, and for the follow-up records see corrective action records. Wider context is in the inspection cluster and the humanoid data section.
What NCRs can and cannot support
NCRs can support models that read defect descriptions and propose probable causes, models that classify severity or disposition from text and measurements, retrieval systems over failure history, and evaluation sets for defect reasoning. When images are attached, they add a text-to-image grounding signal.
They cannot teach perception on their own, because most NCRs carry no image or a single phone photo. They contain no manipulation data, no robot actions and no force or motion information, so they say nothing about how a humanoid should handle, rework or re-present a part. They describe what people concluded, not necessarily what happened: root causes in NCRs are often provisional, and the record may never be updated when a better cause is found.
Treat them as a reasoning and language layer that complements defect image datasets, and as a source of failure vocabulary for failure reasoning data.
Illustrative package: what a useful NCR extract looks like
The table is Illustrative; values would be synthetic placeholders.
| Field group | Example fields | Why it matters |
|---|---|---|
| Identity | ncr_id, opened_date, closed_date, site_id | Time splits and de-duplication |
| Part context | part_family, process_step, supplier_class, lot_id | Domain shift and leakage control |
| Finding | defect_code, free_text_description, detection_point | Label and language |
| Measurement | characteristic, nominal, tolerance, measured_value, unit | Grounding the narrative |
| Judgment | severity, disposition, approver_role | Decision supervision |
| Cause and action | cause_category, root_cause_text, corrective_action_id | Reasoning supervision |
| Attachments | image_ids, drawing_ref, gauge_report_ref | Multimodal linkage |
Ask for a field dictionary alongside the extract. Without one, a column named status may mean workflow state, disposition or inspection result.
Quality criteria to test before licensing
- Completeness of disposition. What share of records have a final disposition rather than an open status? Open records bias a model toward early-stage guesses.
- Cause quality. Sample 30 records and read the root-cause field. Check whether it states a cause or repeats the symptom ("part was scratched; cause: scratch").
- Code consistency. Defect codes change over time and between sites. Ask for the code tables and effective dates; see defect taxonomy and label schema.
- Language mix. Abbreviations, shop jargon and multiple languages inside one field are common and affect text models.
- Template bleed. Copy-pasted boilerplate inflates apparent volume and encourages models to memorize phrases.
- Image linkage rate. Count records with at least one attached image, and check that the image shows the failing characteristic, not just the whole part.
- Time and site coverage. One plant over one year is a narrow slice. State what is covered so evaluation claims stay honest.
Preparation the data holder usually needs to do
Parse. NCRs often live as PDFs, scanned forms or ERP exports. Parsing tables and handwriting introduces errors, so request the structured source where it exists and a sample of parsed output checked against originals.
Redact. Customer names, supplier names, part numbers, drawing numbers, employee names and sometimes pricing appear throughout the narrative. Redaction should be consistent, so the same supplier maps to the same token, which keeps the analytic signal while removing the identifier. Redaction reduces exposure, but it does not by itself settle whether the holder has the right to license the records, particularly where customer contracts restrict disclosure. See confidentiality in defect data. This is not legal advice.
Link. Join NCR IDs to image archives, gauge data and corrective-action records, and keep the join keys. A pile of narratives without links is a weak asset. Where images exist, the question becomes whether each one maps to the recorded outcome; see inspection images linked to outcomes.
Normalize judgment fields. Dispositions such as use-as-is, rework and scrap mean different things between plants, and severity scales vary. Map them to one scheme and keep the original value; the options are discussed in defect severity and disposition.
Split. Prepare time-based and site-based evaluation splits before delivery. NCRs about the same lot or the same supplier problem are near-duplicates; random splitting leaks them across train and test. See leakage in defect datasets.
Who holds this data
The quality department of a manufacturer is the primary holder. Contract manufacturers and tier-two suppliers also hold NCRs about the parts they ship, and customers hold supplier corrective action requests that mirror them. Third-party inspection firms hold reports on behalf of clients, which brings an extra rights layer; see third-party inspection firms as suppliers.
Suppliers often prefer to license a curated, redacted extract rather than the system of record. That is workable if the extract keeps field semantics and linkage intact.
Questions to ask the data holder
- Which system holds the NCRs, and what is the export format and field dictionary?
- How many records have a final disposition and a documented root cause?
- What share have attached images, and where are the originals stored?
- How have defect and cause codes changed, and over what dates?
- What redaction has been applied to customer, supplier and personnel names, and how is it verified?
- Do any customer agreements restrict disclosure of defect information about their parts?
- Can records about the same lot or incident be grouped for split design?
- Who can authorize licensing for model training, and does that cover derived evaluation sets?
Limitations to state in your requirement
Say plainly that the use is text and tabular reasoning, with images only where attached. If you need pixel-accurate defect imagery, specify it separately through pixel-level defect segmentation and linked outcome data. If you want a robot to act on findings, NCRs can supply the decision logic, but the motion has to come from other sources.
Discuss your requirements
If you are looking for quality records from US manufacturers, describe the part types, record fields and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. A request does not guarantee a match.
Discuss your humanoid training-data requirements with SourceX
Sources
- ISO 9001:2026 Quality management systems - Requirements — ISO 9001 sets quality-management requirements including documented information, performance evaluation and improvement, which is where nonconformance records usually originate.