Small factory teams need reversible application deployment

Research on textile small and medium-sized enterprises recommends limited pilots to validate selected technologies before wider implementation. Its implementation model provides a useful starting point, but does not establish a universal factory schedule, staffing requirement or financial return. GFT's proposed extension is to give each pilot a decision owner, a comparison method and an explicit withdrawal condition before connecting it to operational work. The installation itself is an intermediate event; acceptance concerns the decision that the application helps a supervisor make.[S30]

A useful hypothesis describes a process that can be observed. For example, an illustrative inspection pilot might ask whether presenting suspected defects with fabric position helps an inspector locate and classify them without increasing unresolved holds. The baseline records the existing inspection method, product family, operating conditions and treatment of borderline cases. The candidate records the same boundary. A faster screen response is relevant only if the inspector can use it and the physical disposition of the fabric remains traceable. This is a proposed test design, not a reported inspection result.

GFT recommends agreeing success and withdrawal criteria together. Acceptance should require adequate evidence coverage, correct event identity, usable instructions and a confirmed fallback, as well as a beneficial operating measure. A withdrawal criterion might be unexplained discrepancies between inspected material and recorded decisions. The criterion should name who can stop the pilot and where unresolved material will go. If the team discovers that the baseline itself is unreliable, the decision is to repair measurement and repeat the comparison, rather than treat missing records as evidence of improvement.

The pilot brief should also state which work remains outside the test. GFT recommends naming excluded product families, unsupported machine states and the decisions that still use the accepted method. A quiet shift should not silently extend that scope. At handover, the outgoing supervisor should identify any material whose classification remains provisional and the record used to resolve it. A candidate that works only when its developer is present has established a narrower capability than an application operated independently by the shift team.

The comparison should retain product mix, shift, machine state and intervention notes. Under this proposed method, a change in fabric family during the candidate period is a reason to analyse the periods separately. A maintenance intervention is recorded alongside the software release rather than credited to it. Where comparable operating conditions cannot be obtained, the pilot can still establish interface compatibility and recovery behaviour. It should then report those results narrowly and leave the productivity hypothesis unresolved. Technical learning remains useful without a fabricated causal benefit.

For GFT's Cameroon textile programme, the proposed application strategy is to buy or license a proven capability before commissioning bespoke development, while retaining control over capture definitions and access to records. This is a design principle, not an assertion about purchases or supplier performance. The programme should configure the production context that differentiates the factory, then request custom code only where an observed requirement remains unmet. A product demonstration does not prove that local interfaces, intermittent connectivity or the factory's exception handling will work.

The illustrative gates in Exhibit 48 combine the pilot sequence with the return path. Each forward step requires retained evidence. A failed gate returns the application to the agreed operating baseline, not automatically to another development attempt. The diagram is a proposed procedure informed by the pilot literature and high-level risk guidance; neither source prescribes this exact workflow.[S30][S18]

A gate should leave behind an artefact that another shift can inspect: the baseline definition, interface test record, discrepancy log, operator acknowledgement or recovery record. The release owner collects these in a compact release pack rather than creating a parallel reporting system. Shadow operation means the candidate observes and records without controlling production decisions. Moving into an operator pilot changes that authority, so the approval must identify the affected decisions and the means of removing the candidate from them.

The information-exchange part of the manufacturing digital-twin framework identifies technical requirements between reference-architecture entities. That scope supports explicit interface contracts, but does not prove that a particular factory has compatible implementations. GFT proposes recording the accepted contract beside each release so that a deployment cannot silently change the interpretation of an event. A message that still parses may nevertheless have acquired a different unit, classification rule or timestamp meaning.[S06]

Under the proposed method, the release manifest identifies the application artefact, configuration, input schema, output schema, model where applicable and relevant calibration reference. It also identifies the approved predecessor and the compatibility conditions for returning to it. An artefact digest helps distinguish exact packages; it does not establish that the package was authorised or that its origin is trustworthy. GFT recommends separately recording the approved source of the artefact, verification result and release approver, with credentials held outside the manifest.

The provenance standard for characteristic master data has a narrower scope than a complete application history. Its published scope excludes configuration management and change control. GFT therefore distinguishes provenance of a material or asset characteristic from analytical lineage and from release authorisation. Connecting these records is useful, but citing the master-data standard cannot substitute for a change process or establish that model outputs have been reproduced.[S41]

GFT's proposed analytical record links a result to its source-event identity, source quality, application identity and the configuration that affected interpretation. Where an inference uses an image, the record should identify the retained image or governed reference permitted by the retention policy. Where an aggregate uses a rule, the result should identify the rule and aggregation boundary. A corrected result should refer to the superseded result and reason for correction. It should not overwrite the history that explains an earlier operational decision.

The illustrative layers in Exhibit 50 separate what was observed, how it was interpreted and who allowed the interpretation into use. No layer is sufficient alone. An approver can authorise an unsuitable package, and a correct calculation can be displayed outside its accepted operating scope. GFT recommends testing the links by selecting a retained decision and reconstructing it through the release pack, rather than accepting a manifest merely because every field is populated.

Data compatibility determines whether the return path is credible. GFT recommends introducing a new field without immediately removing the accepted field, then checking that both the candidate and predecessor can operate on the retained event set. A destructive migration should be postponed until its separate restoration and reconciliation procedure has passed. Returning an executable to its previous version cannot reverse an incompatible record transformation or recover observations that were never stored.

GFT recommends giving compatibility tests an explicit consumer boundary. A candidate dashboard may understand a changed classification while the accepted reporting service does not. Both must be exercised on the same retained events, including the transition period, before the team calls the release reversible. Compare the interpreted fields and decisions, not just record counts. Equal counts can conceal changed meaning. Where a compatibility adapter is introduced, give that adapter its own accepted artefact and return procedure rather than treating it as invisible migration work.

Configuration should pass the same practical discipline. Under the proposed method, a threshold edit, changed asset mapping or revised unit conversion is a release whenever it changes a decision. Such edits need an identifiable difference, an operator-visible consequence and a tested return path. Model retraining remains a separate lifecycle topic; this chapter requires only that the application can identify the approved model and withdraw its operational use without losing the underlying production evidence.