A use case earns admission only when a decision and baseline exist

Model admission begins with the physical or commercial decision that the factory intends to improve, never with an available dataset or a vendor demonstration.[S43] In the Cross-Industry Standard Process for Data Mining (CRISP-DM) lifecycle applied to manufacturing, business understanding is the opening phase and deployment is the closing phase that makes analytical work operationally real.[S43] ISO/IEC 5338 similarly frames artificial-intelligence lifecycle processes around defined system purpose, stakeholder requirements and operating context before technical build begins.[S13] Reversing that order creates an orphan model with no governed place in daily or weekly factory management.[S18][S43]

Every candidate proposal must state five operational facts before any training extract is approved: the specific decision to be taken, the named role accountable for taking it, the physical or commercial action that follows, the decision horizon, and the escalation route when the signal is contested.[S02][S43] Consider a model that predicts a sewing-line bottleneck four hours ahead. If the line supervisor has neither the authority to rebalance operators across styles nor a buffer of pre-cut bundles for an alternative operation, the score cannot alter production.[S43] Under GFT's admission rule, a candidate whose sole output is an interesting risk score, heat map or anomaly index with no pre-agreed operational response fails the first gate immediately.[S43][S44]

Defining the action and actor first fixes the technical constraints that the data spine must satisfy.[S08][S43] A cloth-inspection alert on a rolling frame requires a decision within seconds so the operator can mark the fault, log the defect code and cut or downgrade the roll before dyeing (wp03).[R29][S08] By contrast, a yarn-purchasing or style-allocation forecast operates on a weekly or monthly planning cycle aligned with sales orders and ginning deliveries (wp04).[S43] The first decision requires low-latency edge inference that survives a wide-area network outage; the second tolerates batch execution but demands reconciled order, inventory and lead-time history across enterprise and operations boundaries.[S08][S22] Naming the decision owner and horizon at admission prevents building streaming pipelines for a monthly planning meeting, or overnight batch reports for a line fault already sewn into three hundred garments by morning, an illustrative timing mismatch.[S02][S43]

No model can prove its worth unless the factory first records how the decision performs without it.[S36][S0902] Practical machine-learning engineering guidance states the rule directly: before formalising what a machine-learning system will do, track as much as possible in the current system.[S0902] World Bank evidence on micro, small and medium enterprise digitalisation reinforces the same requirement, showing that embedding monitoring and evaluation throughout design and implementation is essential for tracking impact and learning what works.[S36] Collecting defect rates, downtime minutes or forecast errors only from the day a model launches blends measurement visibility, supervisory attention and the model itself.[S36][S38]

To make claimed improvement falsifiable, the domain owner and data steward must freeze the governing key performance indicator (KPI) definition in the KPI dictionary before candidate development begins.[S02][S38] Following ISO 22400-2, that frozen record specifies the indicator's exact formula, input elements, unit, time behaviour, grain and exclusions.[S38] For a fabric-inspection candidate, the baseline states whether defect catch accuracy is measured per hundred linear metres or per roll, whether minor slubs within buyer tolerances are excluded, and how manual re-inspection confirms missed faults.[R29][S38] For a demand-forecasting candidate, the baseline locks the forecast horizon, the style-colour grain, and whether cancelled orders or stockouts are adjusted when computing actual demand.[S38][S0901] Changing the KPI formula mid-pilot invalidates the comparison and forces the candidate back to baseline collection.[S36][S38]

A factory baseline is a time-stamped operating record that captures the conditions under which the pre-model process ran.[S08][S38] In a Cameroonian cotton-to-garment operation, fibre characteristics shift across ginning lots, ambient humidity changes between seasons, grid power fluctuates, and operator proficiency varies across shifts.[R03][S38] Evaluating a candidate during a four-week run of plain-weave sheeting against a baseline window dominated by dyed twills confuses product mix with model skill, an illustrative confounding case.[S36][S38] The admission dossier must specify the baseline window, the stratification variables (shift, line, fabric construction, cotton lot and order class) from the operational vocabulary (wp14 Chapter 03), and the counterfactual check separating model contribution from external shifts.[S08][S36][S38]

A recurring mistake in industrial analytics is comparing a complex model against nothing, or against a weak straw man.[S43][S0901] Complexity is an operating liability that must earn its keep.[S16][S18] Every admitted candidate must be tested against two mandatory baselines during evaluation: current operating practice in the frozen baseline window, and the simplest useful rule or statistical comparator built on the same governed dataset.[S43][S0901][S0902]

Engineering guidance for machine-learning systems notes that basic heuristics should launch first when data are scarce; a simple heuristic captures a large fraction of the attainable benefit while establishing the measurement pipeline, whereas a premature complex model is difficult to maintain.[S0902] Forecasting theory enforces the same rule through four standard benchmarks: the historical mean, the naive method (setting future forecasts to the last observation), the seasonal naive method (setting each forecast to the last observation from the same season of the prior cycle), and the drift method.[S0901] As Hyndman and Athanasopoulos state, any new forecasting method must be compared against these simple benchmarks to ensure that it beats them; if not, the new method is not worth considering.[S0901]

In a factory context, simple comparators are transparent and low-cost: a seasonal naive or moving-average model with confirmed contract backlog for demand planning (wp04); fixed run-hour or metre-count preventive intervals from manuals and stoppage logs for maintenance (wp02 and wp15); and statistical process control (SPC) limits or four-point manual perch sampling for quality control (wp03 and wp06).[R29][S05][S38][S43][S0901]

Suppose a gradient-boosted tree improves mean absolute scaled error by 2 percent over a seasonal naive comparator in an offline holdout test, an illustrative result.[S0901] In a small factory, that 2 percent statistical gain may produce zero change in yarn purchase orders because supplier minimum order quantities and container increments are coarse.[S43] Meanwhile, the complex model introduces feature-pipeline fragility and maintenance costs that the seasonal naive formula avoids.[S18][S42] Under GFT's admission standard, the domain owner and model owner set a minimum operational superiority margin before training begins.[S11] If the candidate cannot beat the simple comparator by enough to change physical decisions and recover its extra lifecycle cost, the factory deploys the simple comparator and closes the model candidate.[S44][S0901]

Not every factory model carries the same consequence when wrong.[S11][S14] NIST's Artificial Intelligence Risk Management Framework (AI RMF 1.0) stresses that human judgement must be employed when choosing trustworthiness metrics and setting their precise threshold values in light of consequence.[S11] ISO/IEC 23894 similarly guides organisations to integrate AI risk management into ordinary operating and risk functions.[S14] At admission, GFT classifies every candidate into one of four consequence tiers (wp14 Chapters 02 and 11).[S11][S14]

Tier 1 (Advisory planning): Models generating batch demand projections or maintenance lists that a planner reviews during normal hours with full access to source records.[S11][S14] If a Tier 1 model stalls, the planner reverts to the spreadsheet or seasonal naive comparator with no threat to safety, product integrity or worker status.[S16][S0901] Admission requires domain-owner and steward sign-off, a frozen KPI baseline, and a documented manual fallback.[S02][S38]

Tier 2 (Quality and customer-affecting disposition): Models influencing roll grading, cutting, shade-lot grouping, or buyer carbon and compliance disclosures (wp03, wp05 and wp16).[S04][S14] A false negative allows defective cloth into dyeing or cutting, multiplying scrap cost; excessive false positives halt inspection frames and breed alert fatigue.[R29][S19] Admission requires independent challenge by the quality lead and model owner (wp14 Chapter 02), explicit error-cost weights, a stratified validation plan across styles and lighting, and a mandatory shadow-mode trial.[S04][S11][S16]

Tier 3 (Worker-affecting analytics): Any model touching identifiable workers, skill assessments, ergonomics, shift allocation, pay incentives or discipline.[S29][S30] Principle 5.5 of the International Labour Office (ILO) code of practice on workers' personal data states that decisions concerning a worker should not be based solely on automated processing of that worker's personal data.[S30] While the ILO code is non-binding guidance, GFT adopts Principle 5.5 as a binding internal rule alongside GDPR data minimisation and Cameroon's statutory prior-authorisation requirement in Law No. 2024/017.[S29][S30][S31] Ranking individual workers automatically for pay penalties or termination is prohibited at admission.[S30] Legitimate ergonomic or training analytics require role-aggregated inputs where feasible, documented legal review, and substantive human review where a supervisor has time, context and authority to reject any output.[S29][S30][S31]

Tier 4 (Control-connected OT automation): Any candidate proposed to write setpoints, adjust line speeds, or interface directly with programmable logic controllers (PLCs), drives, boilers or dyehouse valves.[S22] NIST Special Publication 800-82 Revision 3 underscores that operational technology (OT) systems have strict performance, reliability and safety requirements where latency or unintended state changes can damage equipment or injure staff.[S22] In a small factory building its initial data spine, GFT rejects closed-loop Tier 4 model actuation at admission and redesigns the candidate as an open-loop Tier 2 advisory display decoupled from the control network.[S16][S22] Any future control-connected candidate requires a formal process-hazard and OT safety review, independent hardwired interlocks, and joint sign-off by the factory director, OT custodian and safety officer.[S22][S23]