Buying or licensing a pre-trained vision system beats building from scratch on speed, capital risk and Cold-Start accuracy

Data ownership should be explicit enough to survive a change of supplier personnel. The factory retains all raw and annotated captures, roll and garment identifiers, operator verifications, defect coordinates, timestamps and derived production-quality records. The supplier retains its pre-existing software and baseline model. Any supplier use of factory data for training, product improvement or demonstration requires a separate written permission with a defined scope and expiry. Confidentiality language alone does not answer the ownership question.

The contract should also separate four artefacts that are often bundled ambiguously: baseline weights, factory-tuned weights, inference code and training code. A buyer may reasonably accept that training code remains proprietary while requiring delivery of the tuned weights and a local inference package. Where weights cannot be exported, the fallback should be a perpetual right to the last accepted version, an offline installer, complete configuration records and an escrow or continuity arrangement proportionate to the operational risk. The objective is continued inspection and migration, not unnecessary transfer of the supplier's entire research platform.

Local inference needs a precise contractual definition. Normal production must not require per-frame cloud calls, remote authorisation for every session or transmission of full production images. Remote support should be optional, logged and controlled by the factory. If connectivity is unavailable, the accepted version must continue to operate and store events for later synchronisation. Licence servers, update services and analytics dashboards should be listed as separate dependencies so that a claimed edge system cannot conceal a remote point of failure.

An open interface schedule should accompany the commercial schedule. It defines the quality-event fields, identifier formats, timestamp convention, defect dictionary, confidence representation, thumbnail reference, model version and operator disposition. It also defines export in a documented machine-readable form and the destination Performance-Management System tables. A bulk export offered only on request is weaker than a routine local interface available on every shift. The buyer should test a sample export before final acceptance.

Version control requires the same discipline as detection performance. The supplier identifies the accepted model, runtime and defect-mapping versions. Updates arrive with release notes and a rollback path. No update changes the operating threshold or defect definitions silently. The factory can test a candidate on a held-out sample and run it in shadow mode before promotion. The contract should also require notice when a component or model reaches end of support, with enough time to procure a replacement or migrate.

Service terms should describe spares, response routes, local technician training and the division of responsibility for optics, mechanics, computing and software. Acceptance payments follow measured gates, with a retention or warranty mechanism for faults that appear only during sustained production. Exit terms include a final export, removal of supplier access, delivery of configuration and weights where agreed, and a practical handover test. These clauses are procurement recommendations, not legal advice; local counsel should adapt ownership, liability, escrow and continuity provisions to the governing law and the actual transaction.

The licensed baseline is the beginning of local learning, not its end. Every verified defect and false alarm can improve the factory's understanding of its fabrics. Reverse-distillation research has evaluated unsupervised fabric anomaly detection across MVTec AD, AITEX, TILDA and a textile-manufacturing dataset, showing the relevance of methods designed for patterned textures and domain transfer.[S12] Network-pruning research further shows that trained fabric models can be optimised for embedded hardware, with dataset-specific rather than uniform accuracy outcomes.[S02] These findings support a later tuning stage, not an immediate from-scratch programme.

An illustrative planning window of six to twelve months can provide a review cadence, but the calendar must not be the promotion gate. The useful trigger is dataset coverage: enough independently verified examples across priority defect classes, fabric families, lots, shifts and operating conditions. Rare critical classes may remain unsuitable for supervised tuning even after a year. They can continue under one-class screening and human review while better-represented classes move to a local detector.

Data preparation determines whether tuning evidence is credible. Deduplicate near-identical frames, resolve operator disagreements and preserve the original disposition history. Split training, validation and final test sets by roll or production lot so that the same fabric run cannot appear on both sides of the evaluation. Keep a time-based holdout to test new production. Record optical and model settings with every sample. Without these controls, a larger dataset can merely produce a larger leakage problem.

Exhibit 45 sets out the proposed transition. The candidate model does not replace the accepted baseline until it proves itself on frozen data and in shadow operation.

The comparison uses the same class-level recall, false-alarm and speed measures as procurement. It also checks calibration by fabric family and verifies that a gain on common defects does not conceal worse performance on a critical rare class. Shadow mode records what the candidate would have done while the accepted system continues to control disposition. Promotion requires a named operational owner, a signed model package, updated documentation and a tested rollback. A failed candidate returns to analysis without interrupting production.

This stage changes the make-against-buy question. The factory now possesses something it lacked at launch: a verified local corpus and operating measures. It may ask the incumbent to tune the model, commission another supplier to run a challenger, or fine-tune an exportable model internally. The choice should be made on evidence and rights rather than habit. In all three cases, the camera and data spine can remain stable if the original interfaces were designed correctly.

Buy-before-build therefore does not freeze the factory into permanent dependence. It purchases a measurable starting capability, creates time to learn and reserves the option to build where local evidence justifies it. The discipline is to buy the reusable inference engine, own the measurement system and make every later model earn its place on the line.