1 to 2 missed defects per year is the verified outcome at a leading display panel manufacturer running HyperQ AI Vision — a number that exists only because every frame, every classification result, and every operator override was logged to the quality record. Remove the record and that detection rate becomes an assertion. No ISO auditor accepts assertions. No customer portal accepts assertions. The inspection may have happened. Without the record, there is no evidence that it did.
This post covers the data pipeline from camera frame to quality record: what metadata the pipeline must carry at each stage, how batch and lot linkage is established, what hash-chaining means for audit evidence integrity, and where the pipeline connects to your existing MES and eQMS systems. Quality engineers evaluating HyperQ AI Vision for audit-ready traceability will find the integration architecture discussion directly applicable.
The image is not the product — the record is
This distinction is not obvious when you are evaluating a vision system by its detection rate. The demo shows a model catching surface scratches that a human inspector missed. The natural follow-on question is whether the detection rate holds in production. That is the right question. The less obvious follow-on question is what happens to the evidence that a pass or fail was correctly issued.
A vision system that catches defects at 99% rate but produces no structured quality record has the same audit value as a manual inspection with no paperwork. The defect was caught. You cannot prove the defect was caught, when it was caught, which unit it was caught on, or what the disposition was. You cannot reconstruct the inspection history of a specific lot. You cannot demonstrate to a customer's audit team that inspection occurred at the required frequency and scope. You cannot produce the non-conformance record that closes the CAPA loop.
An AI vision deployment is a quality system, not a camera system. The camera is a sensor. The model is an inference engine. The quality record is the deliverable. Building the pipeline to produce that record correctly from the first deployment is considerably cheaper than retrofitting it after a customer audit identifies the gap.
What the pipeline carries at each stage
A complete vision data pipeline moves from raw image through several transformation stages, each of which adds structured data that the quality record needs.
Frame capture The pipeline starts at the camera. Each captured frame needs a timestamp (to millisecond precision), a camera identifier, a station identifier, and a production order or job number linking the frame to the batch being inspected. If the station is inline on a conveyor, the frame also needs a part sequence number or unit identifier — a serial number scanned at entry, a position counter on the conveyor, or a vision-read mark on the part.
Without the unit identifier at capture, the quality record can state that 11,520 units passed through station 3 on a given day. It cannot state the inspection result for unit 7,342. Lot-level traceability covers the first case; unit-level traceability requires the identifier at frame capture.
Inference and classification The AI model produces a classification result (pass, fail, review), a confidence score, a defect class (if fail), and a bounding box or region annotation localising the defect in the image. All four elements belong in the record, not just the pass/fail verdict.
The confidence score is particularly important for quality records. A classification issued at 0.97 confidence and a classification issued at 0.61 confidence are both "pass" results. They represent different risk levels. Defect-class analysis, model drift detection, and retraining prioritisation all benefit from the confidence distribution across classifications. Collapsing confidence to a binary verdict at the pipeline stage discards information you will want later.
Operator interaction Any operator override — a flagged unit cleared by a quality technician, a borderline classification reviewed and accepted — must be logged as a discrete record event with the operator's identifier, the timestamp of the override, and the original classification result. The override record is not a correction to the classification record. It is an additional record event that the quality system can query independently.
ISO 9001 auditors ask specifically about override rates. A vision deployment where operators override more than 3–5% of classifications is a system that needs retraining, not a system performing correctly. If the override events are not logged, the override rate is unknown and the performance claim is unverifiable.
Disposition and hold When a unit is classified as fail or hold, the pipeline needs to trigger a downstream disposition event in the MES: a stop signal to the line, a rejection bin assignment, a non-conformance record opened in the eQMS. This is where the vision pipeline integrates with existing quality infrastructure rather than standing alone.
The disposition record closes the loop on the classification. A classification record without a corresponding disposition record creates an open question: was this unit quarantined, reworked, or inadvertently shipped? Audit teams ask this question. So do customer complaints teams when a defect reaches the field.
Batch and lot linkage
Lot-level traceability requires that every inspection record is linked to the lot number of the material being inspected. This sounds straightforward and is frequently implemented poorly.
The common failure mode: the vision system writes inspection records to its own database keyed by production order number, while the eQMS maintains non-conformance and lot records keyed by lot number. Production order to lot number mapping exists in the MES but is not automatically propagated to the vision system or the eQMS. When an auditor asks for all inspection records associated with lot 2024-0813-A, the quality team manually cross-references three systems.
The correct architecture writes the lot number into the vision record at frame capture. The lot number is available in the MES at job setup — it is the same lot number the MES assigned when the production order was opened. Propagating it to the vision station at the start of each job requires an API call or a manual scan at station setup. Either mechanism is straightforward; the decision to implement it correctly at project start avoids the manual cross-reference problem permanently.
A quality systems lead on a precision components line described the retrospective fix: "Three months after go-live the audit prep team spent two weeks manually cross-referencing the vision system database and the MES lot records to produce the lot history one customer had requested. The inspection data was all there. The linkage simply had not been built. We implemented the job-setup API call after that, but it was a two-week avoidable exercise." The cross-reference problem is not unique to that deployment. It appears consistently in vision integrations where the scoping conversation focused on detection rate and skipped the record structure.
For medical device manufacturing, where 21 CFR Part 820 and ISO 13485 require device history records (DHR), lot-level linkage is not optional. The DHR must include inspection results tied to each device or lot. A vision system that stores inspection records without lot linkage cannot contribute to a DHR without manual data entry — which recreates exactly the manual process the system was deployed to replace.
Hash-chaining evidence for audit integrity
An inspection record can be forged. A timestamp can be changed. A classification result can be edited. A quality system that relies on a flat database of inspection records with no integrity protection is vulnerable to exactly the kind of audit manipulation that regulatory bodies design their verification processes to detect.
Hash-chaining is the mechanism that makes a quality record tamper-evident. When each record in the pipeline is generated, it includes a cryptographic hash of the previous record in the chain. If any historical record is modified — its timestamp, its classification result, its operator identifier — the hash of the modified record no longer matches the hash the next record in the chain expected. The break in the chain is detectable by re-running the verification function. An auditor can verify the entire inspection record history of a lot by running the hash verification and confirming the chain is unbroken.
This is not theoretical. It is the same mechanism that makes blockchain ledgers tamper-evident, applied to a quality record database. The implementation does not require a blockchain — a standard SQL database with a trigger that computes and stores the chained hash at record creation time is sufficient.
For ISO 13485 medical device manufacturers, IEC 62304-regulated software manufacturers, and any facility subject to 21 CFR Part 11 electronic records requirements, hash-chained inspection records provide objective evidence of record integrity that an auditor can verify computationally rather than trusting the quality system administrator's assertion. HyperQ AI Vision includes hash-chained record output as standard. The chain can be exported in standard formats for external verification.
MES and eQMS integration points
The vision data pipeline integrates with existing quality infrastructure at three points.
Job setup (MES → vision system): At the start of each production run, the MES pushes the production order, lot number, part number, revision level, and any inspection programme parameters to the vision station. This step seeds the vision record with the identifiers needed for lot-level linkage and ensures the correct inspection model is loaded for the part being produced.
Real-time disposition (vision system → MES): When the vision system classifies a unit as fail or hold, it sends a disposition event to the MES. The MES processes the event according to its own configured workflow — stop the line, divert to rework, open an NCR. The vision system does not need to know the MES workflow; it needs to publish the disposition event in the format the MES expects. Standard integration formats are REST API, OPC-UA (for machine-level events), and database triggers where the vision system and MES share an infrastructure.
Quality record transfer (vision system → eQMS): At the close of each production batch, the vision system exports a structured inspection record — frame count, pass/fail distribution, defect class breakdown, confidence distribution, override count, hash-chain verification status — to the eQMS. The eQMS attaches this record to the lot's quality history, makes it available for NCR investigation, and includes it in the DHR where required.
The vision model change control post covers the governance layer above these integration points — how model version changes are recorded in the quality system so that inspection records are always tagged with the model version that produced them.
Retention, storage, and access control
Inspection records accumulate quickly. A line running 11,520 units per day generates 11,520 classification records per day, plus associated frame thumbnails if those are stored for review. Over a year, that is roughly 4.2 million classification records for a single line.
Retention rules depend on the product and the regulatory regime. ISO 9001 typically requires quality records to be retained for the product lifetime or a defined minimum — often 7 years for automotive, 10 years for medical devices. A 4.2-million-record-per-year database retained for 7 years is approximately 30 million records before indexing. This is not a large database by modern standards, but the storage architecture — local vs cloud, backup frequency, access control by role — needs to be designed at deployment time.
Common decisions:
- Store classification records and metadata locally on the edge device for 30–90 days (sufficient for immediate NCR investigation), export to a central eQMS or quality data warehouse for long-term retention
- Store full-resolution inspection images only for classified-fail units and for a random sample of classified-pass units (typically 1–5%); this keeps storage volumes manageable while retaining the evidence needed for process audits
- Restrict image and record access by role: quality engineers need read access to records and images; operators need write access for override records; production managers need aggregate dashboard access but not per-unit image access
- Log all access to inspection records with the same hash-chain integrity as the records themselves — an auditor who can verify the record was not modified also needs to verify that access to the record was appropriately controlled
Frequently asked questions
Q: What is the minimum record a vision system must produce for ISO 9001 compliance? ISO 9001 clause 8.6 requires objective evidence of conformity with acceptance criteria before product is released. Minimum: a record of what was inspected, when, against what criteria, and what the result was. In practice this means a per-unit or per-lot inspection record with timestamp, acceptance criteria version, result (pass/fail), and the identity of who (or what system) issued the result. Model version should be included so the record can be traced back to the configuration that produced it.
Q: Can the vision system serve as the sole inspection record for a medical device DHR? The vision system can contribute to the DHR but typically does not replace all manual inspection records. Where the vision system covers 100% inspection of a defined defect class at required resolution and confidence, that coverage can substitute for manual inspection of that class — provided the system is validated under ISO 13485 clause 7.5.6 (validation of processes for production and service provision where the resulting output cannot be verified by subsequent monitoring). Full vision system validation for DHR purposes requires documented software validation, calibration records, and competency records for operators who perform overrides.
Q: How does the pipeline handle inspection of multiple part numbers on the same line? Model switching at the MES-to-vision job setup event handles this. When the production order changes (a different part number begins), the MES signals the vision station to load the appropriate model version. The record for the new batch is tagged with the new production order and part number from the start of that job. Mixed-lot risk (frames from part A classified with part B's model) is eliminated by tying the model switch to the job setup event rather than allowing operator-triggered model selection.
Q: Where do operator override records go in the eQMS? Override records should be treated as quality events — distinct from the original classification record, tagged with the operator identifier and timestamp, and linked to the original classification record by unit and batch identifiers. In eQMS systems that support NCR workflows, an override above a defined threshold (e.g., more than three overrides of fail classifications in a single shift) should trigger a review flag. The override record should not modify the original classification record; the two records coexist in the quality history.
Q: How long does integration with an existing MES typically take? Integration scope depends heavily on the MES vendor and its API support. MES platforms with documented REST API support integrate at the job setup and disposition event points within 1–2 engineering days once the API credentials are available. Legacy MES systems without API access require database-level integration or OPC-UA bridging, which adds project time. HyperQ's standard deployment includes MES integration scoping as part of the 4–8 week contract-to-live timeline.
If your eQMS or MES team is mapping inspection data requirements for an upcoming quality system review or audit preparation, send us your current record structure and the compliance standard you are working toward — ISO 9001, ISO 13485, or 21 CFR Part 11. We will return a pipeline integration map within two weeks showing where vision records attach to your existing quality schema, with no contract required until the integration meets your audit requirements.
Send your quality record schema and get a pipeline integration map