Every defect frame logged at the edge builds an auditable quality and waste ledger for Digital Product Passports

Domestic and regional buyers can use a quality certificate long before an EU passport applies. The useful object is not a decorative badge. It is a signed, batch-specific summary linking the buyer's acceptance standard to inspection coverage, observed defects, dispositions and unresolved exceptions. The certificate should identify the roll or production batch, garment style, inspection stations, method versions and the person or function authorising release. It should also state whether inspection covered every item or a defined sample.

The certificate's strength comes from traceable exceptions. A PASS summary should remain connected to the events that support it, including holds, repairs and later releases. Deleting failed intermediate states would make the final result look cleaner while destroying the evidence needed to investigate a dispute. A corrected record should reference the earlier event and explain the correction. This is more useful than calling the database immutable: authoritatively preserving corrections is the control, not preventing any new information from being added.

The textile tracking prototype cited here combines a permissioned blockchain with product tracking and reports an end-to-end organic-cotton garment experiment.[S28] That demonstrates one way to share signed records among parties. It does not establish that blockchain is necessary for a single factory certificate. A smaller plant can begin with conventional relational records, cryptographic hashes of exported evidence and controlled signatures. The choice should follow the number of independent parties, their trust boundaries and their ability to operate the system.

Human verification remains part of the certificate method. The metal-textile inspection workflow routes doubtful cases to an inspector rather than asking the model to decide every case.[S34] The certificate should therefore distinguish automatic screening, inspector-confirmed disposition and later audit. If a buyer requires item-level assurance, unresolved critical exceptions block release. If the certificate covers a batch, it names the sampling rule and the lots excluded. The document must not imply that high model confidence equals product conformity.

A concise certificate can carry five evidence blocks: identity, acceptance standard, inspection coverage, exception closure and authorisation. Supporting thumbnails and raw events remain available under agreed access rather than crowding the buyer-facing page. Where a dimensional check, shade assessment or laboratory test falls outside the vision station's capability, the certificate marks it as separately evidenced or not performed. Empty fields are preferable to inferred compliance.

This approach has commercial value without inventing a market result. It gives a procurement team a reproducible way to compare deliveries and investigate failures. It also lets the factory answer a complaint by retrieving the exact production and inspection context. The certificate does not guarantee that no defect escaped, and it should say so. Its promise is narrower and stronger: the stated checks were performed on the identified scope, the exceptions were dispositioned, and the evidence can be inspected.

A traceable prediction is not merely an image plus a class. It also needs the model weights, threshold, preprocessing configuration, camera setup and illumination reference that produced it. Without those fields, the factory cannot reproduce why two stations treated similar fabric differently or why the same station changed behaviour after maintenance. Model governance is therefore part of product evidence, not a separate technical luxury.

Drift can begin outside the neural network. LED output can change, lint can obscure the viewport, roller vibration can alter focus, and a new dye lot can shift the background. A daily reference-image check should compare intensity and sharpness against an approved envelope before automatic PASS begins. A failed check should route work to human inspection and record an optics exception. Cleaning or replacing a light then requires a repeat reference capture, not an informal assumption that the station has recovered.

Model outputs need a separate audit. A weekly sample should include automatically cleared work as well as flagged frames, because reviewing only anomalies cannot estimate defect escapes. The edge-cascade study reports that 85 percent of forwarded frames were false alarms in its recall-prioritised benchmark and examines tighter calibration as a way to reduce inference work.[S07] That result is specific to its dataset. It nevertheless shows why the factory should track false alarms, escapes and review burden together rather than optimising one detection metric.

Exhibit 62 proposes a deliberately small operating rhythm. The sample sizes and intervals are illustrative starting points to be approved against production risk and volume.

The weekly sample in the table is not a statistical guarantee. Fifty illustrative frames may be enough to trigger an investigation but insufficient to estimate a low escape rate with useful precision. The quality lead should increase scrutiny after a product change, a critical complaint or a model update, and document the sampling rationale. Stratification should retain fabric family, colour, station, shift and defect severity so that an aggregate review does not hide a weak subgroup.

Candidate models should run beside the incumbent before release. The comparison uses held-out rolls or production periods and checks class-level recall, false alarms, review queue and processing time. Nearby frames from one physical defect must not leak between training and test partitions. The active-learning workflow can supply labels, but uncertain cases need consensus review rather than forced acceptance. Human exception review is valuable precisely because it preserves a route for disagreement with the model.[S34]

Release creates a new evidence version; it does not rewrite the old one. Events produced under the incumbent retain that version forever. Events produced under the candidate carry the new identifier only after approval. If performance deteriorates, rollback restores the complete approved package, including threshold and preprocessing. A later auditor can then understand which decisions were made under which configuration and why the factory changed course.

The same discipline applies to exports. A monthly test should confirm that event identifiers, defect classes, timestamps and dispositions survive the journey from edge buffer to central store without duplication or silent type conversion. A quarterly recovery drill can restore a sampled batch and verify its hashes, references and images. These tests protect the quality ledger from a subtle failure: the model may still be accurate while the audit trail becomes incomplete.

The data spine should begin with a canonical event contract rather than a database decision. Each event needs a durable identifier, event type, product or lot reference, parent references, station, timestamp and time zone, evidence status, method version and disposition. Measured values carry units. Missing values carry a reason. The Excel Quality Log, edge buffer and central relational store then become different implementations of the same meanings. A factory can change technology without changing what a defect, repair, rejection or accepted unit means.

Schema control requires an explicit migration path. A new optional field can be added with a version note. A changed definition cannot quietly reuse an old field name. For example, if waste begins to include reusable remnants after previously excluding them, the report must show the definition break and restate the affected comparison. Controlled vocabularies for defect class, severity, disposition and waste destination prevent stations from inventing local labels that later cannot be reconciled.

Offline operation creates additional audit risks. An edge station may capture events while disconnected, then synchronise after another station has already written related records. Durable event identifiers and idempotent import rules prevent duplication. Original capture time must survive alongside receipt time. A late record should not overwrite a later human disposition without a visible conflict review. These controls let production continue without making the central ledger falsely appear complete at every moment.

Corrections should be append-only in effect. The original entry remains available, the correction references it, and the authoriser and reason are recorded. A role-based permission model restricts who can alter master identifiers, approve dispositions or export evidence. The textile tracking research addresses integrity across multiple supply-chain parties, but a small factory still needs these basic controls even without a distributed ledger.[S28] Technology cannot compensate for shared accounts, unlogged spreadsheet edits or uncontrolled label reissue.

Audit tests should be routine and short. Can a selected garment be traced to its bundle, lay and rolls? Do roll issue and disposition masses reconcile? Does every automatic PASS carry a valid model and station configuration? Are orphan identifiers, missing images and unresolved critical events reported? Can an exported batch be restored and read independently? Each failed test receives an owner and containment action. A passport export drawn from a ledger that fails these checks would only make weak evidence easier to distribute.

Retention should follow evidence purpose and applicable obligations rather than an arbitrary promise to keep everything forever. Raw frames consume storage and may contain sensitive production detail. Event records, dispositions and calibration results may need a different retention period from thumbnails. The factory should define legal, buyer and operational holds, protect backups and test deletion as well as retrieval. Differentiated access, contemplated by the ESPR framework, reinforces the case for separating internal evidence from external disclosure.[S1301]

The anti-greenwashing rule is simple: do not let a measurement say more than its method permits. A seam-puckering study shows that computer vision can make grading more consistent than a subjective visual method.[S21] That supports objective inspection evidence. It does not prove how the garment will perform after repeated washing, wearing or repair. Durability claims need an appropriate test protocol and sample, while life-cycle claims need their own boundaries and data.

The same caution applies to provenance and waste. A supplier lot reference is not by itself proof of farm origin. A lower waste balance is not evidence of lower total environmental impact if production, energy or product lifetime moved adversely. UNEP's roadmap presents circularity as a value-chain programme rather than a single factory metric.[S30] The factory should therefore publish each claim beside its denominator, period, evidence owner and unresolved exclusions. Vision becomes one credible strand in a wider evidence system, not a substitute for it.

A practical release review can ask five questions. Is the identified product the same one described by the claim? Was the method valid when the evidence was captured? Are exceptions and corrections visible? Can another reviewer reproduce the calculation? Does the wording distinguish observation, declaration, inference and estimate? If any answer is no, the claim should be narrowed or held. This discipline protects buyers from overstatement and protects the factory from building a polished passport on unstable evidence.

The resulting ledger is valuable even before external passport obligations settle. It connects defect prevention to physical yield, quality release to retrievable evidence and supplier dialogue to identified lots. It also shows where cameras stop and other methods must begin. That is the correct preparation for Digital Product Passports: not a promise of automatic compliance, but an inspectable record capable of supporting the fields that final product rules and buyers actually require.[S27][S1302]