AI data market
Manufacturing quality data market: 31 open defect sets, zero root causes
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
The manufacturing quality data market is not short of defect images: a public list of 31 industrial defect datasets (October 2026) adds up to about one million images, frames and scans, and none labels a cause or a corrective action. The scarce, licensable layer is the chain from nonconformance to root cause to verified fix. Investors should score that chain, its rights screens and its preparation cost, not image counts.
Checklist
Manufacturing quality data
Key takeaways
- Public defect data is about location, not cause: 23 of 31 listed datasets label pixel masks, and none records an investigation or a fix.
- The investable layer is the linked chain of NCR, root cause, corrective action and dated effectiveness check, which exists only inside quality systems.
- FDA's QMSR, in force since February 2, 2026, puts device makers' quality systems under federal rules, but it does not make their records complete or licensable.
- Customer-owned designs, ITAR technical data and supplier confidentiality can remove the deepest archives; ask for each screen's drop rate.
- Decision context is a testable thesis until someone publishes a utility pilot of linked investigations against image-only data.
What is the manufacturing quality data market really selling?
It sells two different things that are easy to confuse: pictures of defects, and the record of what a plant decided about each defect and whether that decision worked. The first layer is abundant and largely public. The second sits inside quality management systems, and it is the only layer where a supplier of manufacturing quality data can claim something a buyer cannot download.
The decision layer has familiar names on the plant floor. A nonconformance report (NCR) logs the failed part and its disposition: use as is, rework, return or scrap. An 8D report or a corrective and preventive action (CAPA) file records the investigation, the cause the team settled on, the action it took and the check that showed whether the problem came back. A supplier corrective action request (SCAR) carries the same chain across a company boundary.
AI developers are already building agents that reason over this kind of operational record in a neighboring field. IBM's AssetOpsBench, an open benchmark for industrial asset operations, describes 141+ scenarios across nine asset classes that pair sensor data with Industry 4.0 records such as FMEA analyses, work orders and alerts, run in simulated industrial environments. Quality has no public equivalent built from real investigations, and that absence is the opening an investor should test rather than assume.
A census of 31 public defect datasets
Public defect data is plentiful and almost entirely about where a defect is, not why it happened. To check, we counted the dataset table in a large, actively maintained public index of the field, the awesome-industrial-anomaly-detection list kept by M-3LAB on GitHub, at its October 22, 2026 commit.
The table lists 31 datasets. Their annotation levels are segmentation masks (23), bounding boxes (4), point-level masks (2) and image-level labels (2). By source, 22 are described as real, 8 as synthetic and 1 as both. The 30 datasets that state a size add up to 1,004,946 images, video frames or point clouds as each one counts them, with a median of about 3,300; one video dataset, IPAD, accounts for 597,979 of that total.
Across the whole README, which runs to 1,478 lines and 1,103 paper and dataset entries, no line mentions a root cause, a corrective action, CAPA, 8D or a nonconformance. Individual datasets, and the benchmarks built from them, tell the same story at the file level. Amazon's VisA dataset holds 10,821 images, 9,621 normal and 1,200 anomalous, across 12 objects, and each image carries an image-level label and a pixel mask. MMAD, an ICLR 2025 benchmark that tests multimodal models as industrial inspectors, built 39,672 multiple-choice questions from 8,366 images drawn from 4 public datasets.
The census, by annotation level, with what each label teaches a model and what it leaves out:
| Label type in the public table | Datasets | What a model learns from it | What it cannot learn |
|---|---|---|---|
| Segmentation mask | 23 | The exact pixels of a scratch, dent or contamination | Why the defect occurred or what removed it |
| Bounding box | 4 | Roughly where a defect sits in the frame | Which process step or supplier produced it |
| Point-level mask (3D) | 2 | Where a shape deviates on a scanned part | Whether the deviation mattered to the customer |
| Image-level label | 2 | Whether a frame shows any anomaly | The disposition decision and its cost |
| Synthetic source (any label) | 8 fully, 1 partly | Defect appearance that can be generated on demand | Any real plant's decision history |
Why the cause-and-fix chain is the scarce layer
The chain from defect to verified fix is scarce because it only exists where a quality system forced people to write down a judgment and then test it. An image shows the symptom. The investigation shows which of several plausible causes survived the evidence, the action shows what the team was willing to change, and the effectiveness check shows whether reality agreed.
In one sector, regulation makes sure some of these records exist. FDA's Quality Management System Regulation, in force for medical device manufacturers since February 2, 2026, incorporates ISO 13485:2016 by reference, so device manufacturers covered by 21 CFR Part 820 run a documented quality management system under federal rules. That creates a floor of record-keeping in one sector; it says nothing about whether the records are complete, linked or licensable.
The public record shows only the tip of the chain. FDA's medical device recall database records an FDA-determined cause for recalls, with categories such as Device Design, Process control and Software design, and lists the firm's action. FDA also notes that cause determinations can change until a recall is terminated. A category word on a recalled product is a useful signal that regulators care about causes, but it covers only defects that escaped into the market, and it carries none of the evidence a model would need to learn how the cause was found.
Sector opportunity scorecard: five quality record layers
The scorecard rates each record layer on how easily a buyer can substitute public data, how heavy the rights and preparation load is, and what evidence is still missing before anyone should call it an opportunity. Read it from the top down: value rises as public substitutes disappear, and so does the burden of proof.
The ratings are editorial judgments built from the census and the sources in this article. No row is backed by measured buyer demand or utility, because none is public; the evidence-gap column names the measurement that would move each rating up or down. A company claiming a manufacturing quality data asset should be able to show where its records sit on these rows:
| Record layer | Public substitute today | Rights and preparation load | Opportunity read and evidence gap |
|---|---|---|---|
| Defect images and masks | Plentiful: 31 listed datasets, about 1.0 million items, 8 fully synthetic | Low preparation; customer part geometry may be visible | Weak alone. Gap: any proof private images beat public ones for a named task |
| NCR with disposition | None found in the census | Part numbers, customer and supplier names to remove | Moderate. Gap: share of NCRs that link to an investigation |
| Supplier corrective action (SCAR) | None found | Two companies' confidential information in one file | Promising, rights-heavy. Gap: whether supply agreements allow outside use |
| Root-cause investigation (8D, CAPA) | Recall-level cause category only, for recalled devices | Free text naming people and suppliers; controlled technical detail in defense work | Strong if evidence-backed. Gap: share of causes supported by attached evidence |
| Corrective action with effectiveness check | None found | Dates and recurrence windows must survive de-identification | Strongest and scarcest. Gap: share with a dated check and a recurrence window |
Rights screens that remove records before any buyer sees them
Rights screening removes manufacturing records in three predictable places, and an investor should ask for each removal rate separately. The first is customer-owned design information: drawings, specifications and part geometry that a contract manufacturer holds but does not own. The second is export control. The third is the other company named in a supplier dispute.
Export control deserves its own line in diligence. Under the U.S. International Traffic in Arms Regulations, technical data includes information required for the design, development, production, manufacture, assembly, operation, repair, testing, maintenance or modification of defense articles, including blueprints, drawings, photographs, plans, instructions and documentation (22 CFR 120.33). A defect photograph or an inspection record from a defense program can fall inside that definition, so a supplier with defense work may hold quality records it cannot license at all, however complete they are.
Supplier corrective actions raise a quieter problem. A SCAR records one company's finding about another company's part, often under a supply agreement with confidentiality terms, so the supplier's consent or a contractual right may be needed before the file leaves the building. De-identification helps with names, but it does not create a right to license information that belongs to someone else. These are general observations, and counsel reviews any specific case.
How to measure preparation burden on a CAPA sample
Preparation burden is measurable before any deal: draw a random sample of closed CAPA files and time how long it takes to rebuild each one into a linked, de-identified case. The number that matters is not records per plant but linked cases per hour of preparation, because a buyer pays for the chain, not the row.
A workable protocol uses 50 CAPA files drawn at random from the last 24 months, with the random seed written down so the draw can be repeated. For each file, record whether it links to an NCR, whether the stated cause cites evidence (a test, a measurement, a teardown), whether the action is described specifically enough to repeat, and whether a dated effectiveness check exists. Then record the minutes needed to pull the linked records out of the QMS, MES and ERP, and to remove part numbers, customer names and people.
Report the four linkage shares with their denominators, the median minutes per complete case and the share of files dropped by each rights screen. The linkage shares are the headline: a plant where most CAPAs close as operator error with no evidence and no check has records, but not decision context. No plant's results from this protocol are published here; the design is offered so investors can ask for it by name.
Do plant counts or image volume predict supply?
No. Plant counts and image volume measure how much manufacturing exists, not how many linked, rights-clear decision records can leave a plant. The census already shows the gap: about a million public defect items and not one public record of a cause or a fix.
The strongest alternative explanation for the thesis runs the other way. Perhaps buyers simply do not want decision records yet, and the market for manufacturing quality data will stay a perception market, served by public images, synthetic defects and each manufacturer's own camera data. The evidence is consistent with that reading too: MMAD tests models as quality inspectors using images, and the census found 8 datasets that are fully synthetic, a sign that defect appearance can be manufactured on demand.
The way to separate the two readings is a utility pilot, not a market sizing. Give a diagnosis or CAPA-drafting task to two systems, one tuned on linked investigations and one on images plus public text, and score both against held-out investigations whose verified causes are known. Until someone publishes that comparison, decision context is a testable thesis rather than demonstrated demand.
Five ways the decision-context thesis could fail
The thesis fails if any of five things proves true, and each can be checked:
- Demand stays perceptual. Public benchmarks such as MMAD still frame models as inspectors of images; if buyers never fund diagnosis or action-choice tasks, the chain has no paying use.
- Rights screens remove most of an archive. Customer-owned designs and ITAR technical data (22 CFR 120.33) can strip a large share of a contract or defense manufacturer's records, and no public data shows how large that share usually is.
- Records are thin. CAPAs closed without evidence or without a dated check carry little more than an NCR, and a supplier's archive can look large while holding few complete chains.
- Labels move. FDA notes that recall cause determinations can change until a recall is terminated; internal root causes get revised too, so a dataset needs the history of the label, not just its final value.
- Regulation is narrow. The QMSR floor applies to medical device manufacturers; other sectors keep quality records by customer requirement or habit, and depth varies widely.
Questions that expose a thin quality archive
A thin archive gives itself away when the supplier is asked to show the chain instead of the archive size. Counts of NCRs, plants or years say little until each is tied to an investigation, an action and a check.
SourceX is one company pursuing this thesis, and the same questions apply to it. It sources datasets on request from U.S. businesses rather than holding stock, and the supplying company approves the scope and terms and signs before it delivers any records. Its supply and deal figures are not public; company and commercial numbers go to investors directly. The current raise, run under Rule 506(c), is limited to accredited investors, and anyone weighing SourceX against these questions can ask for materials at sourcex.si/investors.
Six questions that work with any company claiming a manufacturing quality data asset:
- On a described random sample, what share of NCRs link to an investigation, a specific action and a dated effectiveness check?
- Which buyer task is the dataset built for: detection, diagnosis, action choice or supplier management, and what evidence shows a buyer wants it?
- How are customer-owned designs and export-controlled technical data found and removed, who signs off, and what share of files did each screen drop?
- What did preparation cost per complete linked case on a pilot, in hours rather than records?
- Has any buyer run a utility test of linked investigations against image-only data, and can the protocol be shared?
- How does the dataset record causes that were revised after closure?
Methods note: census rules, versions and gaps
The census counted the dataset table in README.md of github.com/M-3LAB/awesome-industrial-anomaly-detection at commit d89ca8f (committed October 22, 2026), checked on October 27, 2026. A data row is a table line whose first cell starts with a link; annotation level is the sixth cell as written; a source cell counts as real or synthetic when it contains only that word and as both when it contains both; size is the Total cell when it is an integer, and one row (DTD-Synthetic) with no total is left out of the sum. Paper and dataset entries are README lines whose text begins with a plus sign and a space. The decision-context search, ignoring case, matched root cause or root-cause, corrective, CAPA and 8D as whole words, and nonconformance or non-conformance, over every line of the README. Anyone can repeat the count by cloning the repository at that commit and applying these rules.
The list is one curated index, so the census measures what one large public catalog contains, not every defect dataset in existence; sizes are as the list reports them and mix images, frames and scans. VisA, MMAD and AssetOpsBench figures come from their own repository READMEs at the commits cited in the sources. Nothing here measures private record quality, buyer demand or the utility of linked investigations, and no SourceX data was used; the scorecard ratings are editorial and should be replaced by measured values as they become available.
Frequently asked questions
How big is the manufacturing quality data market?
No public source sizes the market for licensed manufacturing quality records, and this article does not estimate one. What can be measured is the substitute: a large public index lists 31 defect datasets with about one million images, frames and scans, none with causes or corrective actions. Sector output and plant counts say little about how many linked, rights-clear records exist.
Why don't public defect image datasets satisfy AI buyers working on quality?
They answer where a defect is, which is the detection task. Diagnosis, disposition and action choice need the investigation and its outcome, and the census found no public dataset that labels either. A buyer building a quality engineering assistant needs those records from manufacturers.
What counts as decision context in a manufacturing quality record?
It is the linked chain around a defect: the NCR and its disposition, the investigation and the evidence behind the chosen cause, the corrective action, and a dated check showing whether the problem recurred. A record missing the cause evidence or the check carries much less of the reasoning a model is meant to learn.
Can aerospace or defense quality records be licensed for AI training?
Sometimes, but export control comes first. Under 22 CFR 120.33, technical data for defense articles includes photographs and documentation needed for production, testing and repair, so defect images and inspection files from those programs may be controlled. Counsel reviews each program before any record is offered.
What should an investor ask a quality-data supplier first?
Ask for linkage shares on a described random sample: how many NCRs connect to an investigation, a specific action and a dated effectiveness check. That one question separates an archive from a decision dataset faster than any count of records or plants.
Sources
- The M-3LAB awesome-industrial-anomaly-detection README lists 31 datasets with annotation level, source and size columns; per the census rules in the methods note, 23 use segmentation masks, 8 are synthetic, stated sizes add to 1,004,946 items, and no README line mentions root cause, corrective action, CAPA, 8D or nonconformance. Source
- The VisA dataset contains 10,821 images (9,621 normal, 1,200 anomalous) across 12 objects, with image-level labels and pixel-level masks. Source
- MMAD (ICLR 2025) generated 39,672 multiple-choice questions in 7 subtasks from 8,366 samples of 38 product classes drawn from 4 public datasets, to test multimodal models in industrial inspection. Source
- AssetOpsBench describes 141+ scenarios across 9 asset classes for industrial asset operations agents, using sensor data and Industry 4.0 records such as FMEA, work orders and alerts in simulated environments. Source
- FDA's Quality Management System Regulation took effect February 2, 2026 and incorporates ISO 13485:2016 by reference. Source
- FDA's medical device recall database records an FDA-determined cause for recalls, with categories such as Device Design, Process control and Software design, and notes that cause determinations can change until a recall is terminated. Source
- Under 22 CFR 120.33, ITAR technical data includes information required for the design, production, manufacture, repair, testing or maintenance of defense articles, including drawings, photographs and documentation. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.