Research libraryThe Industrial Data Spine
Chapter 13 of 14 · Part IV · Draft web edition
The spine earns further investment when people use it, decisions improve and net operational benefit survives a comparison with the baseline.
The cloth is woven from this chapter's 4 exhibits: each weft is one exhibit's own data, normalised; each warp group is a section. The probe reads the exhibit's own value where it crosses. Procedural visualisation, not a measurement.
Read-out
Illustrative planning assumptions, not observed durations or benchmarks. Low is initial routine or shadow use, mid is stable task competence with a usable comparison, and high is a net-value review with adequate evidence. S40's two-to-three-year statement concerns TVET strategy and does not set these ranges.
Unit: illustrative months from approval. Status: illustrative. Sources: S35, S36, S40, S43. Illustrative figures are teaching examples, not measured factory results.
A data spine creates no operating value merely by going live. Value begins when a supervisor uses a trusted record to change a decision, when the resulting improvement survives comparison with the pre-intervention baseline, and when that benefit exceeds the recurring work required to keep the record or model useful. The launch is therefore the start of measurement, not its conclusion. GFT's recommendation is a recurring value routine with four linked tests: people perform the governed work, operational indicators improve under like-for-like conditions, net benefit remains after lifecycle effort, and trust or safety does not deteriorate.[S11] [S18] [S36] [S38]
Sustained use cannot be inferred from initial enthusiasm. In a World Bank study of an enterprise platform offered to SMEs in Viet Nam, close to 80 percent of supported firms used it after initial training, about 40 percent used it thereafter, and 35 percent were still using it about 18 months later. Firms that sustained adoption commonly assigned an internal point person to champion it.[S37] The study concerned business software in a different country and context. It is not a Cameroon factory forecast. It does show why activation and training completion are weak evidence of durable adoption.
This chapter turns that lesson into factory practice. It defines adoption as observable behaviour, measures outcomes through existing manufacturing indicators, separates cash effects from capacity and risk effects, and gives every deployment a fix, scale, maintain, postpone or stop decision. The method is an engineering governance framework. It is not investment, legal or tax advice. Site values, accounting treatment and worker-data obligations require the responsible factory functions and, where necessary, qualified local advisers.
13.1
Adoption needs an operating owner because friction crosses functional boundaries. A capture form may belong technically to IT, its defect codes to quality, its completion time to production and its use in a shift meeting to operations. Without one person accountable for follow-through, each function can defend its own part while the complete routine fails. The World Bank's SME finding supports a named internal champion, but it does not justify dependence on a heroic individual. The champion needs management backing, a deputy, allocated time and a route into existing production decisions.[S37]
GFT would assign the champion three duties. First, make the governed workflow usable during ordinary shifts, including handover, interruption and exception handling. Second, coordinate role-based practice and confirm that users can perform their tasks without assistance. Third, bring behavioural evidence and unresolved friction to the factory's existing management cadence. Domain stewards remain close to the work. They diagnose capture defects, help operators correct records and explain why a definition or interface must change. Technical custodians keep the service available, but they do not own adoption.
A private spreadsheet or paper tally after launch is evidence before it is a compliance issue. It may preserve a field missing from the governed record, provide a faster view at the line, allow work during disconnection or protect a local calculation that management still uses. Removing the sheet by instruction can hide the record without removing the need. The steward should ask which decision the workaround supports, what the official workflow cannot represent and which change would make the governed route more useful.
The review should also distinguish necessary local work from harmful duplication. A temporary offline contingency may be legitimate if it has an owner, a synchronisation rule and a reconciliation check. An unowned monthly extract that changes the KPI denominator is different. The former is a controlled operating mode. The latter fractures the baseline and prevents comparison. The champion records the cause, corrective action and verification result rather than reporting only the number of workarounds found.
13.3
Training should reproduce the decisions a role will face, not tour software menus. OECD case evidence from a German bakery SME says that converting traditional processes into digital ones requires sizeable effort, does not directly pay off and involves a longer learning process.[S35] ILO and UNESCO guidance separately favours tangible pilots with practical use and built-in evidence about what works. Its two-to-three-year horizon concerns digital transformation of technical and vocational skills systems, not a universal factory payback period.[S40] Together, the sources support practice around real work and patience with learning, not a generic promise of rapid return.
A line steward should practise timely capture, an unknown defect code, a missing lot reference and correction after quarantine. A quality supervisor should practise reviewing a model flag, recording a reasoned override and checking the later physical outcome. A planner should practise comparing a forecast with the current statistical rule and documenting why a schedule changed. A custodian should practise restoring the records and instructions needed to reproduce a reported figure. Each exercise uses realistic records with personal data removed or protected according to purpose.
Attendance does not demonstrate readiness. The practical test is whether the user can complete the governed task, recognise an exception and select the correct escalation route without prompting. A suitable assessment records the task, conditions, errors, assistance required and reassessment result. Failure leads first to inspection of the work instruction and interface. A confusing taxonomy or slow screen is a design defect, not a training deficiency to be repeated unchanged.
The champion should test competence again after a material process, schema or model change. This prevents an initial training certificate from becoming permanent evidence of capability. It also connects training to release control: a role that cannot execute fallback safely keeps the deployment in its current limited mode. NIST's risk framework places human judgement around metrics and thresholds, so users must also understand what a score means, when it is uncertain and when they are expected to challenge it.[S11]
Time is required for habits to stabilise and for a fair comparison to accumulate. The ranges in Exhibit 56 are illustrative planning assumptions for three different evidence problems. They are not benchmarks or commitments. A manual spine needs stable capture and repeated management use. A bought vision system needs local validation and override evidence even if no local training is planned. A history-dependent forecaster needs sufficient relevant order history and a comparator before its value can be judged.[S35] [S36] [S40] [S43]
The ranges prevent two opposite mistakes. The first is declaring value before users and baselines have stabilised. The second is continuing a weak deployment indefinitely because learning is still supposedly under way. Each review should state which evidence is missing, when it should become available and what decision follows if it does not. Calendar time never substitutes for an evidence gate.[S36] [S44]
13.6
Read-out
Derived provenance graph linking raw visibility event, versioned quality gate, curated record, KPI formula execution, shift review, frozen dataset snapshot, pipeline run and model decision, based on W3C PROV-O (S09), Google Cloud MLOps lineage requirements (S15) and W3C PROV Primer (S0403).
Unit: . Status: derived. Sources: S09, S15, S0403.
Read this exhibit in its chapter ↗13.6 · Introduction
Licences issued, dashboard views and stored rows describe system activity.
Exhibit 17, chapter 04: Raw event log
13.7 · Adoption metrics need denominators that resist gaming
Every behavioural measure can be distorted if its denominator is convenient.
Exhibit 17, chapter 04: Raw event log
13.8 · Override frequency needs a quality check
A low override rate can mean that a model is useful, but it can also mean that users accept its output without examining it.
Exhibit 17, chapter 04: Quality gate v1.2
13.6 · the bodyThe film above follows these sub-steps.
Licences issued, dashboard views and stored rows describe system activity. They do not establish that a record arrived in time for a decision, passed its quality checks or changed an operating action. The World Bank recommends embedding monitoring and evaluation throughout digitalisation programme design and implementation, which requires measures attached to the intervention's mechanism rather than a launch inventory.[S36] For the industrial data spine, GFT's proposed adoption set therefore uses five behavioural measures: timely capture, validated completeness, correction recurrence, active routine integration and documented override reasons.
Timely capture compares event time with the contract deadline for the decision. A shift total entered before handover may be useful even if a machine event arrived later through a verified offline queue. The measure must preserve that distinction rather than treating every replay as late work. Validated completeness counts required records that pass the applicable rules at first submission, with missing and estimated values disclosed rather than silently converted to zero. Data quality remains purpose-specific and managed across the lifecycle, consistent with ISO/IEC 5259-3.[S04]
Correction recurrence tests whether the same capture cause returns after correction. A falling number of corrected rows can signal improvement, but it can also result from a disabled rule or unreported workaround. Active routine integration checks whether shift, production and quality reviews cite governed records and close actions against them. Documented override reasons show whether a person recorded why a recommendation was rejected and whether the later outcome was reviewed. These are proposed factory measures, not universal thresholds. Each site sets its target after establishing a baseline.[S36]
Every behavioural measure can be distorted if its denominator is convenient. Timely capture improves artificially when expected records are removed from the schedule. Completeness improves when difficult fields become optional. Routine integration improves when actions are closed without verification. The data owner should therefore approve the denominator and its exceptions before the result is used in a review. Changes remain visible through the contract version and effective date.
The review should pair each adoption indicator with a counter-check. Timely capture is read beside late corrections and physical sampling. Validated completeness is read beside quarantine volume and disabled rules. Override documentation is read beside the quality of the eventual decision. A high score that fails its counter-check opens an investigation rather than earning praise. This keeps the measure attached to factory work and discourages teams from optimising the report instead of the routine.
Adoption should also be compared across roles and shifts. An aggregate result can conceal a team that carries the capture burden for everyone else or a night shift that receives no steward support. Where the sample is too small for a conclusion, the register should say so and collect more evidence. The purpose is not to rank workers. It is to find where the operating design needs correction.
A low override rate can mean that a model is useful, but it can also mean that users accept its output without examining it. A high rate can mean poor model fitness, a confusing interface or a legitimate response to conditions absent from training. NIST warns that people may assume AI systems work well in every setting and calls for measurement to reflect real operational context.[S11] The adoption review should therefore sample agreements as well as overrides and compare both with delayed ground truth where available.
Three outcomes matter: a justified override catches a model error, an unjustified override rejects a sound recommendation, and an unchallenged model error passes without human intervention. The third is especially important because a simple override rate hides it. Sampling should include different lines, product families, shifts and data-quality conditions. Results lead to interface correction, threshold review, further training or suspension, not an instruction to agree with the model more often.[S11]
Interlude · the cloth half-woven
Woven: sections 13.0 to 13.6Still to weave: 13.9 to 13.17
Read-out
Inherited GFT programme evidence base prepared 31 January 2024. R06 and R07 were not refreshed in 2026 and their original publication dates are unrecorded. Figures describe sector-level baseline and policy ambition, not a current GFT factory measurement or causal productivity claim.
Unit: percent of cotton fibre. Status: sourced. Sources: R06, R07.
Read this exhibit in its chapter ↗13.9
M1
M1: Timely capture 62, First-pass yield 84, Override quality 78 (illustrative percent on a shared zero to 100 scale) · illustrative
A validation score answers whether a candidate meets a technical test.
M2
M2: Timely capture 74, First-pass yield 85, Override quality 82 (illustrative percent on a shared zero to 100 scale) · illustrative
M3
M3: Timely capture 85, First-pass yield 88, Override quality 86 (illustrative percent on a shared zero to 100 scale) · illustrative
M4
M4: Timely capture 88, First-pass yield 89, Override quality 71 (illustrative percent on a shared zero to 100 scale) · illustrative
M5
M5: Timely capture 89, First-pass yield 87, Override quality 64 (illustrative percent on a shared zero to 100 scale) · illustrative
M6
M6: Timely capture 91, First-pass yield 90, Override quality 88 (illustrative percent on a shared zero to 100 scale) · illustrative
Illustrative six-month sequence. Values are invented to show diagnostic divergence. Timely capture and first-pass yield improve while audited override quality falls in M4 and M5, then recovers after intervention. These are not measured GFT or client outcomes.
Unit: illustrative percent on a shared zero to 100 scale. Status: illustrative. Sources: S11, S30, S36, S38. Illustrative figures are teaching examples, not measured factory results.
13.9 · the bodyExhibit 57 above; its figures are illustrative where marked.
A validation score answers whether a candidate meets a technical test. It does not answer whether the factory produces better work. Operational evaluation should reuse the KPI dictionary and its frozen formulas so that the baseline, intervention period and subsequent review retain the same meaning. ISO 22400-2 presents manufacturing KPIs through formulas, corresponding elements, time behaviour, units and other characteristics, making those definitions the appropriate reference point for comparison.[S38]
ISO 22400-2 presents manufacturing KPIs through formulas
S38
The selected indicators should follow the deployment's mechanism. Quality assistance may affect first-pass yield, rework and scrap. Shared stoppage evidence may affect availability and changeover duration. Maintenance advice may affect unplanned stops and repair time. Order and forecasting tools may affect schedule adherence and work in progress. Metering and allocation may affect energy, water or material intensity. A benefit claim should name the physical pathway from record to action to result; an improvement without that pathway is a hypothesis to investigate, not a validated effect.[S36] [S38]
The central question is what would have happened without the intervention under comparable conditions. Product mix, material grade, order urgency, equipment state and staffing can change at the same time as a release. A before-and-after average that ignores those changes can credit the tool for an easier mix or blame it for a harder one. The evaluation should hold the baseline formula constant and compare matched lines, shifts, styles or operating strata where the available evidence permits.[S36] [S43]
A simple rule helps: report the total observed change separately from the portion attributed to the intervention. The first is an operational observation. The second requires a comparison method and confidence statement. If a like-for-like comparison is impossible, the register should retain the result as unverified contribution rather than forcing a causal claim. This is particularly important when several changes share one period, such as a new maintenance routine and a model release.
A usable baseline contains more than a historical average. It records the KPI version, source records, quality state, observation window, operating conditions and current comparator. It also records known weaknesses, such as incomplete stoppage capture or an inconsistent product boundary. If those weaknesses would change the value decision, they become corrective work before a strong benefit claim is allowed.[S04] [S36] [S38]
The baseline owner should be independent enough from the deployment team to challenge a favourable reconstruction. Where new information improves an old record, the team should preserve both the original reported baseline and the corrected version, with lineage explaining the difference. This allows reviewers to distinguish a genuine operating change from an accounting repair. It also prevents a weak starting point from being selectively adjusted until the intervention appears valuable.
Uncertainty belongs in the outcome as well as the baseline. A narrow difference between two highly variable periods may not support a scale decision even when its arithmetic is correct. The register should state the sample support, important uncontrolled conditions and the additional observation needed. This is a decision-quality requirement, not a demand for statistical complexity that a small team cannot maintain.
The illustrative panels in Exhibit 57 show why these measures should not be read alone. Capture and yield can improve while override quality deteriorates. A programme that reports only the first two panels would miss the emerging oversight problem. The values are invented to demonstrate the review method and are not GFT results or factory benchmarks.[S11] [S30] [S36] [S38]
In this illustrative sequence, timely capture rises from 62 to 91 percent and first-pass yield ends at 90 percent, while override quality falls from 86 to 64 percent before recovering. The correct response is not to celebrate the average direction. It is to inspect the model, alert burden and user decisions during the weaker months, then verify whether the corrective action restored informed review. The exhibit cannot establish why the measures moved; it tells the team where to look.
timely capture rises from 62 to 91 percent and first-pass yield ends at 90 percent
illustrative
13.13
Exhibit 587 paragraphs step it
Read-out
Illustrative reconciliation: 100 - 28 - 12 - 14 - 11 - 9 = 26. Cost categories reflect lifecycle maintenance in S18 and the application-specific processing warning in S39. Proportions are invented for method demonstration, not benchmarks or GFT results.
Unit: illustrative annual value index, gross benefit equals 100. Status: illustrative. Sources: S18, S36, S39. Illustrative figures are teaching examples, not measured factory results.
A real operational improvement can still produce a weak investment once the work around the model is counted. Research on machine-learning systems identifies ongoing maintenance costs arising from dependencies, configuration, hidden feedback and changes in the external world.[S18] A 2025 NIST semiconductor workshop report adds a sharply bounded warning: manufacturers reported that data processing for a single application can cost as much as 70 percent of total application cost. That is a reported upper bound in a particular manufacturing context, not an average for SMEs or a prediction for a textile factory.[S39]
The value review should therefore begin with validated gross operating benefit and deduct the recurring effort needed to sustain it. Preparation includes entry, cleaning, labelling and reconciliation. Integration includes schema, master-data and interface maintenance. Human review includes alert handling, false interventions and override documentation. Model operation includes monitoring, delayed-label review and revalidation. Continuity and support include recovery tests, access review, incremental infrastructure and licences. Initial implementation cost should also remain visible, but it must not be used to hide the recurring run cost.[S18] [S36] [S39]
Exhibit 58 expresses the reconciliation as an illustrative index. It is deliberately not a return ratio. The purpose is to make every deduction visible before the remaining value is discussed.[S18] [S36] [S39]
In the illustration, recurring lifecycle work consumes 74 of the 100 gross-benefit units and leaves 26. The appropriate management question is not whether the gross case sounded attractive. It is which deductions can be reduced through reuse and better capture without weakening control, and whether the remaining net value justifies the deployment's risk and scope. A negative result should trigger redesign, postponement or retirement rather than a search for additional unverified benefits.[S44]
Direct cash effects include measured reductions in material waste, external rework, energy purchases or paid overtime, supported by quantities and prices for the stated period. Freed capacity includes machine or labour hours released by the change. It becomes additional revenue only when saleable demand and the remaining production constraints allow those hours to be used. Until then, it should be reported in hours or physical units, not converted into speculative cash.
Avoided risk is different again. A lower probability of a customer claim, safety event or equipment failure may matter greatly, but its monetary estimate depends on likelihood, consequence and assumptions. Keep that estimate separate from ledger savings and disclose its basis. Do not add cash, unused capacity and avoided risk into one unqualified total. The register should let management see each category and decide which is relevant to the resource decision.[S11] [S36]
13.15
A deployment can improve an operating indicator and still be unacceptable. False alerts can create unnecessary stops or inspections. A capture routine can distract an operator or encourage an unsafe shortcut. Worker-level records can become a disciplinary instrument even when their original purpose was process diagnosis. NIST treats positive and negative impacts as part of AI risk management, so harm and trust evidence belongs in the same review as benefit rather than in a separate annual exercise.[S11]
The trust review should cover false-intervention burden, safety near misses, unresolved override friction, worker concerns and contestation outcomes. Each measure needs a defined event and investigation route. A concern count alone is ambiguous: fewer complaints can mean fewer problems or less confidence that reporting will help. The more useful evidence is whether concerns were understood, whether the response changed the system and whether the same cause returned.[S11] [S30]
The ILO code of practice states that decisions concerning a worker should not be based solely on automated processing of that worker's personal data. It also says that personal data collected by electronic monitoring should not be the only factors in evaluating performance. The code is non-binding guidance, not a statement of Cameroonian law.[S30] GFT's operating recommendation is to keep worker-affecting decisions under informed human review, provide an accessible challenge route and correct disputed records without retaliation.
Meaningful review requires time, authority and context. If a station record shows low output, the reviewer should be able to inspect material, machine, style and measurement conditions before attributing fault. If a worker challenges the record, the steward should preserve the original evidence and append the resolution. A system that cannot explain or correct its worker-affecting use has not passed the value gate, regardless of its apparent financial return.[S11] [S30]
13.16
Value evidence becomes useful only when it changes a resource decision. GFT's proposed benefit register gives every active deployment one row and every review one explicit outcome. The adoption champion maintains it with the operating owner, using the factory's existing management forum rather than creating a parallel programme office. Each row links back to the admission case, the frozen baseline and the released artefact so the team evaluates the same mechanism and scope that was approved.[S36] [S43]
The register records the accountable owner, baseline definition and period, expected mechanism, behavioural adoption, observed operating effect, comparison method, gross benefit, recurring lifecycle cost, evidence confidence, adverse effects and next decision. It also records missing evidence. This prevents an uncertain estimate from being silently promoted to a measured saving between reviews. Where the KPI definition or production boundary changes, the old and new results remain distinguishable.[S36] [S38] [S41]
Four actions are available. Scale expands scope only after adoption, net value, recovery and trust checks pass. Maintain keeps a beneficial deployment within its current scope where expansion adds no proven value. Fix contains the deployment and assigns dated corrective work when friction, cost or override quality threatens the case. Postpone or stop returns to the verified manual or simpler comparator when the necessary resources, skills, costs or expected benefits do not support continuation. Research with SMEs explicitly recognises postponement as a valid investment decision under those conditions.[S44]
Read-out
| Line log: stops 14.2% of shift | 92% timely; no open concern | Stops down 2.4 pts; high | Scale | Reconcile on next line |
|---|---|---|---|---|
| Vision: downgrade 1.8% of metres | 88% override quality | Net XAF 6.0M/yr; medium | Maintain | Repeat matched-style audit |
| Energy log: steam 9.4 kg/kg | 71% complete; manual meters | Break-even; low | Fix | Test meter and recost |
| Bearing alert: 6 stops/quarter | 48% override quality | Negative; medium | Stop | Compare fixed maintenance |
| Forecast: MAPE 28% at 8 weeks | Shadow use; 7 months history | Unproven; low | Postpone | Review after more history |
Derived register structure from embedded evaluation, governed KPI definitions, lifecycle cost and stop decisions. All deployment values are illustrative, not observed GFT results. Net value means validated operating benefit less recurring lifecycle effort. Confidence reflects comparison and input quality, not model accuracy alone.
Unit: derived register with illustrative row entries. Status: derived. Sources: S18, S30, S36, S38, S44.
Stopping does not erase the learning. The registry should retain why the deployment failed, which records remain reusable and which assumption would need to change before reconsideration. A stopped model may leave a valuable governed dataset and a clearer baseline. Continuing to pay for an unsupported model, by contrast, consumes the same small team's capacity for capture, maintenance and process improvement.[S18] [S44]
13.17
Value measurement should start before every input is perfect, but it should show where the evidence is weak. GHG Protocol guidance describes collecting data, assessing quality and improving quality as an iterative process, with lower-quality inputs replaced over time. Although that guidance addresses Scope 3 emissions, the iterative principle is useful for factory value records when its original context remains explicit.[S41]
Although that guidance addresses Scope 3 emissions
S41
GFT's register should distinguish four evidence states. Measured contributions come from direct observations under the relevant KPI definition and comparison. Partially measured or allocated contributions combine observations with documented allocation rules. Estimated or proxy contributions depend on defaults, samples or assumptions. Model-derived contributions carry their validation limits and uncertainty. These labels describe the evidence behind a value claim, not a general score for the deployment.[S04] [S36] [S41]
The improvement decision follows consequence. A low-confidence input that has little influence on a decision may be adequate for the next review. A proxy that determines whether a deployment scales deserves priority for a better meter, sample or reconciliation routine. This focuses spending on the evidence gap that can change a decision rather than on indiscriminate data collection. Each iteration should record what changed and whether earlier results need recalculation.[S36] [S41]
The same discipline can connect the paper's companion tools without pretending they share one denominator. The kpi-cascade and oee-line reviews can preserve operational definitions. The inspection-economics and cold-start-forecast reviews can expose preparation, review and comparator costs. The footprint review can carry measured, allocated and estimated states into resource intensity. The benefit register is the common decision layer, while each specialised method remains distinct.[S36] [S38] [S39] [S41]
The chapter's test is therefore cumulative. Launch proves availability. Adoption proves governed behaviour. Operational comparison tests the claimed mechanism. Lifecycle accounting tests net value. Trust review tests acceptability. The benefit register turns those findings into a decision, and quality disclosure makes the strength of that decision visible. Expansion is justified only when the complete chain holds.
Check the record
Every knot is a source this chapter cites; select one to read its record, or filter the register by grade. Publication dates and inherited-source limits stay explicit. Read the paper's full evidence summary.
13 sources · 91 citation marks on this pageMost cited: S36, 21 times
Grade AGrade BGrade CKnot size: citations on this page
All 13 records shown
c13-S04 · Grade A
ISO/IEC 5259-3:2024: Artificial intelligence, Data quality for analytics and machine learning, Part 3: Data quality management requirements and guidelines
ISO/IEC 5259-3 defines the requirements and guidance for managing the quality of data used in analytics and machine learning (ML).
2024. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S11 · Grade A
Artificial Intelligence Risk Management Framework, AI RMF 1.0, NIST AI 100-1
Human judgment should be employed when deciding on the specific metrics related to AI trustworthiness characteristics and the precise threshold values for those metrics.
2023-01-26. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S18 · Grade B
Hidden Technical Debt in Machine Learning Systems
Using the software engineering framework of technical debt, we find it is common to incur massive ongoing maintenance costs in real-world ML systems.
2015. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S30 · Grade A
Protection of workers' personal data: An ILO code of practice
Decisions concerning a worker should not be based solely on the automated processing of that worker's personal data.
1997. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S35 · Grade A
Employee empowerment through digital tools: The case of a German bakery SME
However, following managing director Holste the change of traditional processes into digital ones requires sizeable efforts that do not directly pay off but involve a longer learning process.
2022-11-21. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S36 · Grade A
The Missing Element of Firm Digitalization: Lessons for the EU Member States: Best practices for enhancing MSMEs Digitalization and Managerial Capacity Building
Embedding monitoring and evaluation throughout each step of the program design and implementation process is crucial for generating knowledge, tracking impact, and improving learning for the further improvement of interventions pursuing the digitalization of MSMEs.
2023. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S37 · Grade B
Challenges with digital technology adoption by SMEs: A case study from Viet Nam
Firms that successfully took up the platform assigned a point person who championed adoption within the firm.
2025-05-20. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S38 · Grade A
ISO 22400-2:2014: Automation systems and integration, Key performance indicators for manufacturing operations management, Part 2: Definitions and descriptions
The KPIs are presented by means of their formula and corresponding elements, their time behaviour, their unit/dimension and other characteristics.
2014. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S39 · Grade A
Artificial Intelligence with Open and Scaled Data Sharing in Semiconductor Manufacturing: Workshop Report
Also, manufacturers have reported that the cost of data processing for a single application can be as high as 70% of the total cost.
2025-11-18. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S40 · Grade A
The Digitization of TVET and Skills Systems
We suggest that digital transformation strategies have a better chance of success if they are developed to support tangible pilots that can demonstrate the practical use of emerging technologies with a short timeframe (two to three years maximum) and inbuilt mechanisms to deploy technology that provides ongoing data on what works and what does not to the sponsoring stakeholders.
2020. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S41 · Grade B
Technical Guidance for Calculating Scope 3 Emissions, version 1.0
Collecting data, assessing data quality, and improving data quality is an iterative process.
2013. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S43 · Grade B
Ensuring the Robustness and Reliability of Data-Driven Knowledge Discovery Models in Production and Manufacturing
Briefly, CRISP-DM divides the data mining-related knowledge discovery process into six phases: 1) business understanding, 2) data understanding, 3) data preparation, 4) modeling, 5) evaluation and 6) deployment.
2021-06-14. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c13-S44 · Grade B
Supporting SME companies in mapping out AI potential: a Finnish AI development case
One must remember that in IT investments, the decision to postpone, e.g. an AI project, is also valid if resources, skills, costs, or expected benefits are not met.
2024-07-27. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗S11 National Institute of Standards and Technology. 2023-01-26.
S18 Advances in Neural Information Processing Systems. 2015.
S36 World Bank. 2023.
S38 International Organization for Standardization. 2014.
S37 World Bank. 2025-05-20.
S35 OECD. 2022-11-21.
S40 International Labour Organization and UNESCO. 2020.
S43 Frontiers in Artificial Intelligence. 2021-06-14.
S44 The Journal of Technology Transfer, Springer Nature. 2024-07-27.
S04 ISO/IEC. 2024.
S30 International Labour Office. 1997.
S39 National Institute of Standards and Technology. 2025-11-18.
S41 GHG Protocol, World Resources Institute and World Business Council for Sustainable Development. 2013.
Keep the thread · chapter 13 of 14
Move the shuttle along the rack to preview any chapter: its sections, exhibits and reading time. Draft web edition. This research is not investment, legal or tax advice; illustrative models are not measured factory results.
The paper's rack · 14 boltsBolt length: reading time
01Part I: Govern meaning before scale
The binding constraint on industrial AI is often the broken chain between work, record and decision, not the absence of an algorithm.
02Part I: Govern meaning before scale
A workable governance model fits into existing factory roles and makes every definition, exception and release somebody's explicit decision.
03Part I: Govern meaning before scale
Shared identifiers, event semantics and KPI formulas let unlike systems describe the same work without hiding their differences.
04Part I: Govern meaning before scale
A small set of enforced checks, visible exceptions and complete lineage is more useful than an unmeasured claim that the data are clean.
05Part II: Build the minimum durable spine
A durable architecture preserves source evidence, adds governed context once and lets each decision consume a controlled product.
06Part II: Build the minimum durable spine
A value becomes useful evidence only when the factory can establish what it describes, when it happened and which state or transformation it represents.
07Part II: Build the minimum durable spine
The data spine must protect production, people and legitimate access at the same time, with controls proportionate to a small team.
08Part II: Build the minimum durable spine
A factory proves resilience by operating through disconnection, restoring its records and leaving a vendor without losing its history.
09Part III: Control models as production change
The first model gate is operational and economic: nobody should train before the action, baseline, data rights and stop conditions are explicit.
10Part III: Control models as production change
A model becomes a release candidate only when another competent person can reconstruct what data, code, configuration and evaluation produced it.
11Part III: Control models as production change
Release authority follows demonstrated local fitness, safe fallback and accountable acceptance, not a strong aggregate validation score.
12Part III: Control models as production change
A useful monitor detects change, establishes whether it matters and routes a named person through a rehearsed response.
13Part IV: Prove value and scale deliberately
The spine earns further investment when people use it, decisions improve and net operational benefit survives a comparison with the baseline.
14Part IV: Prove value and scale deliberately
GFT should scale the spine only as ownership, meaning, recovery, reproducibility and net value are demonstrated in production.