Research libraryThe Industrial Data Spine
Chapter 11 of 14 · Part III · Draft web edition
Release authority follows demonstrated local fitness, safe fallback and accountable acceptance, not a strong aggregate validation score.
The cloth is woven from this chapter's 5 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 shadow-validation comparison across five operational indicators. Values are constructed to demonstrate paired pre-release evaluation under S38 KPI definitions, S43 lifecycle evaluation and S1102 baseline sanity checks, not measured results from a GFT facility.
Unit: illustrative index (0 to 100, lower is better). Status: illustrative. Sources: S38, S43, S1102. Illustrative figures are teaching examples, not measured factory results.
A machine-learning candidate that scores well on an offline test partition has not yet earned the right to alter factory work. Release is an operational change decision, not the automatic final step of a training script.[S11][S15] Promoting a model into production changes how inspectors examine cloth, how planners size buffers, or how supervisors allocate maintenance across shifts.[S43] Governing that change by a single aggregate accuracy figure exposes daily execution to unmeasured subgroup failures, false-alarm fatigue, silent pipeline defects and workflow dependency.[S19][S1101] Green Future Tech Limited (GFT), Hong Kong SAR, designed its three-part Cameroon textile programme (prepared 31 January 2024, with source-refresh notes retained in 2026) around the sector baseline in which about 95 percent of Cameroonian cotton fibre is exported raw and about 5 percent is processed locally, against a 50 percent local-processing target by 2030 (these inherited baseline entries were not refreshed in 2026 and their original publication dates are unrecorded).[R06][R07] Moving from raw-fibre export to competitive local spinning, weaving, dyeing and garment assembly requires decision tools that line managers trust under shop-floor constraints.[R06][S04] Validation and staged release keep every bought or locally trained model subordinate to operations through four controls: human-set thresholds tied to floor consequences, disaggregated validation across local slices, staged authority expansion from shadow mode to assisted operation, and rehearsed rollback to a known baseline.[S11][S16][S22]
Earlier chapters in this part established that a candidate enters development only after an operational decision, baseline and comparator are frozen, and that training becomes auditable only when code, dataset snapshots, feature definitions and execution lineage are recorded in a versioned registry.[S09][S15][S38] Those controls prove how a candidate was built, not that it is safe or useful on the floor.[S13][S1102] This chapter specifies the validation and release discipline that separates technical calculation from operational acceptance, prices the cost of wrong interventions, sequences deployment through conditional stages, and treats rollback and manual fallback as tested release artefacts.[S14][S15][S43]
11.1
An automated pipeline can compute precision, recall, mean absolute error or calibration curves, but it cannot decide what an error costs on a weaving shed floor or in a cutting room.[S11][S38] The National Institute of Standards and Technology (NIST) Artificial Intelligence Risk Management Framework (AI RMF 1.0) states the rule directly: human judgement must be employed when choosing the specific metrics related to AI trustworthiness characteristics and the precise threshold values for those metrics.[S11] Delegating threshold selection to software defaults or to the developer who trained the model confuses mathematical convenience with operational risk appetite.[S11][S14] In a factory spine, the domain operating owner and designated risk owner set acceptance metrics and cut-off values jointly before validation begins, locking those thresholds into the evaluation protocol so they cannot be adjusted retroactively.[S02][S11][S15]
Every industrial prediction error carries an asymmetric physical or commercial cost.[S11][S43] In fabric-defect detection, a false negative lets a stained or broken-pick roll pass into dyeing or cutting, multiplying wasted labour, dye liquor and cloth.[R29][S43] A false positive stops the frame or flags a clean roll for manual re-inspection, consuming operator minutes without destroying material.[R29][S38] In short-horizon demand forecasting, over-prediction ties up working capital in slow-moving greige cloth, whereas under-prediction triggers short shipments or overtime changeovers.[S38][S43] Because these consequences hit production, quality and commercial budgets, the accountable line manager defines the trade-off between missed events and false alarms.[S02][S11]
Setting human thresholds translates statistical scores into unit quantities managed under ISO 22400-2 indicator definitions.[S11][S38] An Area Under the Receiver Operating Characteristic curve (AUC) of 0.94 is not an operational threshold because a supervisor cannot staff around an abstract area score.[S38][S1102] An actionable threshold states required recall at a maximum false-alarm rate per thousand linear metres of cloth, or maximum mean absolute percentage error alongside a cap on single-order under-predictions during peak shipping weeks, recorded in the model registry with its operating assumptions.[R29][S11][S15][S38]
A model that clears a single accuracy hurdle can still degrade operations if its input feed is incomplete, its latency stalls a work cell, or its gains on high-volume styles are paid for by severe errors on specialty lots.[S04][S19][S1101] GFT structures pre-release acceptance around four mandatory threshold families: technical discrimination and calibration, operational net benefit, serving data fitness, and bounded subgroup behaviour.[S04][S11][S1101] Each candidate must pass all four families simultaneously.[S11][S14]
Exhibit 46 sets out the derived structure of these four threshold families, mapping each family to its primary metric class, required disaggregation dimensions, mandatory failure action and accountable owner.[S04][S11][S19][S1101]
Read-out
| Technical performance | Recall, precision, calibration error | Defect class, fabric weight, weave structure | Block export; refine features or threshold | Model builder and model owner |
|---|---|---|---|---|
| Operational benefit | Net review minutes, first-pass yield, scrap delta | Line, shift, operator experience tier | Hold in shadow or reject candidate | Domain operating owner |
| Serving data fitness | Completeness, timeliness, valid range share | Capture device, lighting state, offline buffer | Quarantine feed; repair capture before re-test | Data steward and data custodian |
| Subgroup floor | Worst-slice error cap, minimum sample support | Low-volume style, night shift, degraded mode | Restrict operational scope or stop release | Risk owner and domain owner |
Derived threshold framework combining human metric and threshold authority from S11, ML data-quality management from S04, upstream cascade awareness from S19, and disaggregated condition reporting from S1101.
Unit: governance specification. Status: derived. Sources: S04, S11, S19, S1101.
As Exhibit 46 illustrates, each threshold family triggers a specific remedy when a candidate falls short.[S04][S11] A failure on serving data fitness sends the team back to capture hardware, lighting calibration or schema enforcement rather than hyperparameter tuning, preventing upstream data cascades.[S04][S19] However, Exhibit 46 is a governance contract rather than a universal numeric table: it cannot establish the exact numerical cut-offs for a specific loom or inspection table, which must come from local baseline records.[S11][S38]
11.4
Aggregate evaluation scores weight every test record equally, so they are dominated by the factory's most frequent styles, best-lit day shifts and cleanest sensor feeds.[S04][S1101] Mitchell and colleagues show through their model-card framework that benchmarked evaluation must be disaggregated across conditions, groups and intersectional combinations relevant to the application domain so unsuitable contexts are exposed before release.[S1101] On a production line, disaggregation prevents releasing a model that performs well on standard plain-weave greige cotton but fails on dyed twill, heavy canvas, humid afternoon shifts or frames running after a yarn-lot changeover.[S04][S19][S1101]
Under ISO/IEC 5259-3, managing data quality for analytics and machine learning requires verifying that validation datasets represent actual operating conditions.[S04] When evaluating a candidate, a factory team partitions held-out validation and shadow datasets along six operational axes wherever applicable:[S04][S38][S1101]
Method
Disaggregated slice validation and minimum support check
To evaluate local operating slices without mistaking sampling noise for proven fitness, the validation protocol executes four sequential checks on the frozen held-out snapshot and shadow log:[S04][S11][S15][S1101]
style_group x line_id and defect_severity x shift_code), linked to governed identifiers in the KPI dictionary.[S02][S38][S1101]Intersectional slices are equally important.[S1101] A vision candidate may pass recall thresholds across all shifts combined and across dark-dyed fabrics combined, yet fail when a dark-dyed style runs on Line B during the night shift under artificial lighting.[R29][S1101] Pre-release slice tables that cross-tabulate style family by line and shift catch that failure before a customer rejects a lot.[S19][S1101]
Disaggregating a validation set reduces the sample size in each cell.[S04][S1101] If a specialty style appears only eight times in a snapshot, a single misclassified roll shifts the apparent error rate by 12.5 percentage points (an illustrative ratio of 1 divided by 8), risking either false confidence or premature rejection.[S04][S11]
11.7
Read-out
Derived stage-gate path from operational decision and frozen baseline through data and rights readiness, simple comparator test and risk-adjusted lifecycle value check to four explicit exit outcomes, based on the lifecycle phases in S43 and the SME postponement and stop criteria in S44.
Unit: . Status: derived. Sources: S43, S44.
Read this exhibit in its chapter ↗11.7 · Introduction
A candidate model is never evaluated against an implicit alternative of zero capability.
Exhibit 40, chapter 09: 1. Decision and frozen baseline
11.8 · The candidate must beat both current practice and a transparent comparator
Every validation report places three options side by side on the same historical and shadow windows:
Exhibit 40, chapter 09: 1. Decision and frozen baseline
11.9 · Review burden, false interventions and latency enter the net acceptance balance
A statistical gain that slows the line or exhausts operators is an operational regression.
Exhibit 40, chapter 09: 2. Data, rights and capacity check
11.7 · the bodyThe film above follows these sub-steps.
A candidate model is never evaluated against an implicit alternative of zero capability.[S43][S1102] Even a manual factory has a working decision process with an observable yield, speed and error cost.[S36][S38] Google's machine-learning engineering rules advise keeping the first model simple to establish baseline metrics and behaviour, and running sanity checks on held-out data right before exporting any model to serving.[S1102] In the CRISP-DM manufacturing lifecycle, the evaluation phase verifies that the model meets business objectives before deployment.[S43]
Industry literature in GFT's register notes that manual fabric defect inspection is 60 to 70 percent accurate and misses 20 to 30 percent of defects, whereas computer-vision benchmarks exceed 90 percent accuracy across more than 40 defect types at speeds up to 60 metres a minute (inherited Grade-C entry, not refreshed in 2026 and original dates unrecorded).[R29] Those figures justify a pilot, but local lint, cotton-seed trash, vibration and weave variations can push a vendor model's precision below a trained two-person inspection perch.[R29][S04][S19] Only local comparison settles the question.[S43][S1102]
A statistical gain that slows the line or exhausts operators is an operational regression.[S22][S38][S43] If a vision system catches four additional true defects per shift but triggers forty false stops requiring three minutes each, it consumes 120 minutes of inspection capacity per shift (an illustrative product of 40 stops and 3 minutes) and drives operators to ignore alerts.[S18][S19][S38]
Exhibit 47 illustrates how a candidate model is compared against current baseline practice across five normalised operational burden and error indicators during shadow validation.[S38][S43][S1102]
In Exhibit 47, lower index values represent less wasted time, fewer false stops or lower error over the same shift population.[S38][S43] If a candidate cut missed defect share from 26 to 11 on the illustrative index while raising false intervention rate from 42 to 75, the chart would expose an unacceptable trade-off hidden by a composite F1-score.[S11][S43] Because Exhibit 47 is an illustrative demonstration, it cannot establish actual factory gains; live decisions require measured units from a frozen shadow run.[S38][S43]
…cut missed defect share from 26 to 11 on the illustrative index while raising…
S11, S43 · illustrative
Interlude · the cloth half-woven
Woven: sections 11.0 to 11.7Still to weave: 11.10 to 11.22
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 ↗11.10
Switching a live process directly from manual execution to closed-loop model control introduces unnecessary operational and safety risk.[S16][S22] Offline snapshots cannot replicate live camera fouling, network micro-outages, clock drift or end-of-month dispatch pressure.[S19][S22][S42] Microsoft's MLOps maturity model advises building capability gradually rather than attempting full automation upfront, while NIST SP 800-82 Revision 3 prioritises OT performance, reliability and safety.[S16][S22] Combining these rules creates a staged release ladder.[S16][S22]
while NIST SP 800-82 Revision 3 prioritises OT performance
S16, S22
Every new deployment or major version upgrade begins in silent shadow mode: the model ingests live events or frames, computes features and predictions, and logs timestamped outputs and latency, while its outputs remain hidden from operators and cannot trigger machine or workflow actions.[S15][S16][S22]
Once the shadow log passes all four threshold families over a predefined window (such as an illustrative window of two to four weekly cycles across all shifts and main style families), the operating owner may promote the candidate to operator-assist mode, where the model displays recommendations or bounding boxes while the operator retains authority to confirm, modify or reject each action.[S11][S16][S30]
Advancing from shadow to operator-assist mode is an explicit gate decision, and advancing further to bounded automation (acting automatically only within narrow pre-approved limits) is optional.[S16][S22][S42] Decisions involving complex rework trade-offs, customer concessions or safety-adjacent settings remain permanently in operator-assist mode.[S11][S22][S30]
Exhibit 48 shows an illustrative 16-week staged release sequence from candidate freeze through shadow validation, rollback rehearsal, operator-assist operation and conditional bounded automation.[S15][S16][S22]
Exhibit 48 shows an illustrative 16-week staged release sequence from candidate freeze through shadow validation
S15, S16, S22 · illustrative
Read-out
Illustrative 16-week staged release schedule derived from S15 pipeline lineage and rollback controls, S16 incremental MLOps maturity, and S22 OT reliability constraints. Durations are planning illustrations, not fixed commitments.
Unit: illustrative weeks from candidate freeze. Status: illustrative. Sources: S15, S16, S22. Illustrative figures are teaching examples, not measured factory results.
As Exhibit 48 shows, the rollback rehearsal occurs in Week 5 to Week 6 before Gate 1 approves operator-assist activation in Week 7, and Gate 2 in Week 11 to Week 13 audits human overrides and KPI movement before any expansion.[S11][S15][S27][S38] Because Exhibit 48 is an illustrative schedule, it cannot fix calendar durations across use cases; monthly planning models require longer shadow windows than high-speed inspection cameras.[S16][S38][S43]
11.13
1. Detect breach or fault
1. Detect breach or fault. Starting point. Leads to 2. Suspend active model output (Threshold or steward trigger). · derived
A model release is safe only if the factory can undo it faster than a bad prediction series can damage production.
2. Suspend active model output
2. Suspend active model output. From 1. Detect breach or fault (Threshold or steward trigger). Leads to 3. Revert to v(k-1) or manual baseline (Registry pointer or SOP switch). · derived
3. Revert to v(k-1) or manual baseline
3. Revert to v(k-1) or manual baseline. From 2. Suspend active model output (Registry pointer or SOP switch). Leads to 4. Notify floor and log shift boundary (Confirm stable line state). · derived
4. Notify floor and log shift boundary
4. Notify floor and log shift boundary. From 3. Revert to v(k-1) or manual baseline (Confirm stable line state). Leads to 5. Quarantine and reconcile affected lots (Tag interval in event log). · derived
5. Quarantine and reconcile affected lots
5. Quarantine and reconcile affected lots. From 4. Notify floor and log shift boundary (Tag interval in event log). Leads to 6. Diagnose cause and revalidate (Preserve failed snapshot). · derived
6. Diagnose cause and revalidate
6. Diagnose cause and revalidate. From 5. Quarantine and reconcile affected lots (Preserve failed snapshot). End point. · derived
The method behind it: Warning.
Derived rollback and fallback sequence combining S15 artefact and pipeline lineage, S22 OT safety and reliability preservation, and S27 practised incident-response procedures.
Unit: operational control sequence. Status: derived. Sources: S15, S22, S27.
11.13 · the bodyExhibit 49 above; its figures are illustrative where marked.
A model release is safe only if the factory can undo it faster than a bad prediction series can damage production.[S15][S22] When deployment overwrites scripts in place and removes paper check-sheets, a bumped camera lens or corrupted feature leaves the floor unable to restore operations without hours of downtime.[S18][S19][S22] GFT treats technical rollback (reverting to prior approved version $v_{k-1}$) and operational manual fallback (reverting to the non-model baseline workflow) as mandatory release artefacts built and exercised before approval.[S15][S22][S27]
When an upstream sensor fails or a first-time model is suspended, the line must drop back cleanly to manual or deterministic baseline operation.[S22][S27] Following NIST SP 800-61 Revision 3 guidance on periodic incident exercises, a pre-release fallback drill verifies that workflows survive suspension without lost records.[S22][S27]
Exhibit 49 traces the derived six-stage control path from anomaly detection through suspension, reversion, floor communication, lot reconciliation and investigation.[S15][S22][S27]
In Exhibit 49, suspending the model (Steps 2 and 3) must never leave an OT controller or terminal in an undefined state, and Step 4 logs the exact timestamp and lot boundary so KPI calculations separate model-assisted intervals from fallback intervals.[S09][S22][S38] Step 5 reconciles lots processed just before suspension.[S04][S27] Because Exhibit 49 is a process specification, it cannot guarantee recovery times unless timed physically on the line before approval.[S22][S27]
11.16
When the analyst or vendor who trained a model also decides whether it is ready for production, optimism bias enters the gate.[S11][S14] Both ISO/IEC 23894 and the NIST Cybersecurity Framework 2.0 place risk acceptance with accountable operational leadership rather than technical builders alone.[S14][S23]
…and the NIST Cybersecurity Framework 2.0 place risk acceptance with accountable operational leadership rather…
S14, S23
A small Cameroonian spinning, weaving or garment plant needs no separate model-risk department to prevent self-approval: the analyst or vendor prepares the evidence package, while the quality or production manager (operating owner) and plant or maintenance manager (OT and risk approver) hold veto and sign-off rights.[S02][S11][S14][S16][S42]
Read-out
Derived consequence-to-authority matrix based on S11 human threshold authority, S14 integrated AI risk management, S22 OT safety constraints, S23 ERM governance, and S30 worker-data protection. Coordinates represent ordinal governance tiers rather than empirical probabilities.
Unit: derived governance score (0 to 10 scale). Status: derived. Sources: S11, S14, S22, S23, S30.
As Exhibit 50 shows, no deployment is permitted in the lower-right quadrant (Uncontrolled release risk).[S11][S14] Advisory planning (x = 2.8, y = 3.5) requires steward verification and operating-owner sign-off, whereas worker-affecting analytics (x = 7.8, y = 8.6) and control-connected OT (x = 9.2, y = 9.5) sit in the upper-right quadrant (Mandatory multi-role gate) with strict builder-approver separation and safety interlocks.[S11][S22][S30] These ordinal coordinates illustrate governance tiers and cannot replace a machine-specific hazard assessment.[S14][S22]
11.19
Factory data spines touch human work whenever logs describe identifiable operators.[S29][S30] Alongside GDPR Recital 39 data minimisation and the statutory prior-authorisation requirement in Section 19(1) of Cameroon's Law No. 2024/017 of 23 December 2024 (requiring local legal counsel for interpretation), Principle 5.5 of the International Labour Office (ILO) code of practice states that decisions concerning a worker should not be based solely on automated processing of that worker's personal data.[S29][S30][S31] GFT enforces this non-binding ILO principle as a mandatory release gate.[S11][S30]
Alongside GDPR Recital 39 data minimisation and the statutory prior-authorisation requirement in Section 19(1)…
S29, S30, S31
An "Approve / Override" button is not meaningful review if a supervisor must clear fifty scores in three minutes without seeing machine-fault context or faces pressure not to override.[S11][S19][S30] Validation of any worker-affecting candidate must verify four conditions before release:[S11][S29][S30]
The operator-assist interface logs each recommendation, reviewer action, reason code and review duration under least-privilege access controls.[S15][S25][S29] Gate 2 and monthly reviews audit overrides: near-100 percent overrides signal poor utility, while 0 percent overrides with two-second review times warn of mechanical rubber-stamping.[S11][S30][S36]
11.22
Related · chapter 076 paragraphs step it
Read-out
Illustrative permission mapping across seven factory roles on two governance dimensions: x measures operational and model write authority (0 = read/append local only, 10 = full pipeline or infrastructure write), y measures scope of sensitive and identifiable record visibility (0 = anonymised or single-asset only, 10 = plant-wide or personal-data review). Positions are illustrative design targets based on S25, S29 and S0701, not empirical measurements.
Unit: illustrative score (0 to 10). Status: illustrative. Sources: S25, S29, S0701. Illustrative figures are teaching examples, not measured factory results.
Read this exhibit in its chapter ↗A model release recorded only in a code repository leaves operators uninformed.[S12][S13] Translating ISO/IEC 42001:2023 management-system requirements and ISO/IEC 5338:2023 lifecycle processes into shop-floor practice requires publishing a concise release package alongside every promoted version.[S12][S13][S1101]
Combining ISO/IEC 42001, ISO/IEC 5338 and Mitchell and colleagues' model cards, the release package pairs a registry manifest with a one-page workstation sheet containing eight elements:[S11][S12][S13][S15][S1101]
In OECD case-study evidence on SME digitalisation, converting traditional processes into digital routines requires structured learning rather than yielding an instant payoff, and ILO-UNESCO guidance emphasises tangible pilots with built-in feedback and user preparation.[S35][S40] Model activation is aligned to a clean shift, batch or planning boundary and recorded as a provenance event, preceded by a workstation briefing from the steward and adoption champion covering shadow examples, override logging and the manual fallback sheet.[S09][S27][S35][S37][S38]
With the release package active, the candidate operates under controlled authority on the floor. Yet validation at release is only a snapshot: cotton fibre properties vary across harvests, dye lots shift, cameras accumulate lint, and ordering patterns change.[R01][S18][S21] Keeping the deployed model subordinate to operations over time requires continuous production monitoring, drift diagnosis and rehearsed incident response, examined next.[S17][S21][S27]
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.
37 sources · 415 citation marks on this pageMost cited: S11, 48 times
Grade AGrade BGrade CKnot size: citations on this page
All 37 records shown
R01 · Grade C · inherited record
Inherited evidence R01
Cameroon seed-cotton production 2022-2024: 330,000 t (2022), 314,455 t (2023), about 340,000 t (2024).
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R06 · Grade C · inherited record
Inherited evidence R06
About 95 percent of cotton fibre exported raw against 5 percent processed locally; industrialisation target of 50 percent local processing.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R07 · Grade C · inherited record
Inherited evidence R07
SND30 restates the 50 percent local-processing target by 2030.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R14 · Grade C · inherited record
Inherited evidence R14
Cameroon minimum wage (SMIG) XAF 41,875 a month in 2023, raised to XAF 43,969 a month from 23 February 2024, about US$76, a rise of 5 percent.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R29 · Grade C · inherited record
Inherited evidence R29
Manual defect inspection 60 to 70 percent accurate, missing 20 to 30 percent of defects; computer vision above 90 percent across more than 40 defect types; line-scan inspection up to 60 metres a minute.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
R32 · Grade C · inherited record
Inherited evidence R32
CFA franc to US dollar reference rates; franc pegged to the euro at 655.957. The GFT study's method note places the franc near 590 to 610 per US dollar across the 2023 study window and near 563 to 577 more recently, and treats currency conversions as estimates.
Publication date unrecorded. Inherited source-text record. The original publication has not been independently reopened for this web edition.
c11-S02 · Grade A
ISO 8000-150:2022: Data quality, Part 150: Data quality management: Roles and responsibilities
This document specifies the key considerations for organizations that are establishing appropriate roles and responsibilities for data quality management.
2022. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-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 ↗c11-S09 · Grade A
PROV-O: The PROV Ontology
It provides a set of classes, properties, and restrictions that can be used to represent and interchange provenance information generated in different systems and under different contexts.
2013-04-30. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-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 ↗c11-S12 · Grade A
ISO/IEC 42001:2023: Information technology, Artificial intelligence, Management system
ISO/IEC 42001 is an international standard that specifies requirements for establishing, implementing, maintaining, and continually improving an Artificial Intelligence Management System (AIMS) within organizations.
2023. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S13 · Grade A
ISO/IEC 5338:2023: Information technology, Artificial intelligence, AI system life cycle processes
This document defines a set of processes and associated concepts for describing the life cycle of AI systems based on machine learning and heuristic systems.
2023. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S14 · Grade A
ISO/IEC 23894:2023: Information technology, Artificial intelligence, Guidance on risk management
The guidance also aims to assist organizations to integrate risk management into their AI-related activities and functions.
2023. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S15 · Grade B
MLOps: Continuous delivery and automation pipelines in machine learning
Information about each execution of the ML pipeline is recorded in order to help with data and artifacts lineage, reproducibility, and comparisons.
2024-08-28. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S16 · Grade B
MLOps Maturity Model
Use this model to develop your MLOps capability gradually instead of facing the full complexity of mature implementation upfront.
2026-06-03. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S17 · Grade B
Model monitoring in production
You can set thresholds for these metrics to trigger alerts about model or data anomalies via Azure Machine Learning or Azure Event Grid.
2026-01-27. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-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 ↗c11-S19 · Grade B
Everyone wants to do the model work, not the data work: Data Cascades in High-Stakes AI
Data cascades are pervasive (92% prevalence), invisible, delayed, but often avoidable.
2021. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S21 · Grade B
Learning under Concept Drift: A Review
Concept drift describes unforeseeable changes in the underlying distribution of streaming data over time.
2020-04-13. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S22 · Grade A
Guide to Operational Technology Security, NIST SP 800-82 Revision 3
This document provides guidance on how to secure operational technology (OT) while addressing their unique performance, reliability, and safety requirements.
2023. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S23 · Grade A
The NIST Cybersecurity Framework 2.0
Governance activities are critical for incorporating cybersecurity into an organization's broader enterprise risk management (ERM) strategy.
2024-02-26. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S25 · Grade A
Security and Privacy Controls for Information Systems and Organizations, NIST SP 800-53 Revision 5
Employ the principle of least privilege, allowing only authorized accesses for users (or processes acting on behalf of users) that are necessary to accomplish assigned organizational tasks.
2020-12-10. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S26 · Grade A
#StopRansomware Guide
Maintain offline, encrypted backups of critical data, and regularly test the availability and integrity of backups in a disaster recovery scenario [CPG 2.R].
2023. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S27 · Grade A
Incident Response Recommendations and Considerations for Cybersecurity Risk Management: A CSF 2.0 Community Profile, NIST SP 800-61 Revision 3
Procedures can be tested or exercised periodically to verify their accuracy and can be used to help train new personnel.
2025. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S29 · Grade A
Regulation (EU) 2016/679, General Data Protection Regulation
The personal data should be adequate, relevant and limited to what is necessary for the purposes for which they are processed.
2016-04-27. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-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 ↗c11-S31 · Grade A
Law No. 2024/017 of 23 December 2024 relating to personal data protection in Cameroon
The processing of personal data shall be subject to prior authorization by the Personal Data Protection Authority.
2024-12-23. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S35 · Grade B
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 ↗c11-S36 · Grade B
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 ↗c11-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 ↗c11-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 ↗c11-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 ↗c11-S42 · Grade B
Is MLOps different in Industry 4.0? General and Specific Challenges
Also the starting-level of the technical infrastructure as well as the aimed at degree of automation influence very significantly the overall complexity of implementing MLOps in Industry 4.0.
2022. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-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 ↗c11-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 ↗c11-S1101 · Grade B
Model Cards for Model Reporting
Model cards are short documents accompanying trained machine learning models that provide benchmarked evaluation in a variety of conditions, such as across different cultural, demographic, or phenotypic groups (e.g., race, geographic location, sex, Fitzpatrick skin type) and intersectional groups (e.g., age and race, or sex and Fitzpatrick skin type) that are relevant to the intended application domains.
2019-01-14. Register access date: 2026-10-30. Source provenance is supplied by the paper evidence register.
Open the recorded source ↗c11-S1102 · Grade B
Rules of Machine Learning: Best Practices for ML Engineering
Do sanity checks right before you export the model. Specifically, make sure that the model's performance is reasonable on held out data. Or, if you have lingering concerns with the data, don't export a model.
2025-08-25. 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.
S15 Google Cloud. 2024-08-28.
S43 Frontiers in Artificial Intelligence. 2021-06-14.
S19 Association for Computing Machinery, CHI 2021. 2021.
S1101 arXiv, Conference on Fairness, Accountability, and Transparency. 2019-01-14.
R06 Cameroon Trade Hub. Publication date unrecorded; inherited source-text check.
R07 Ecofin Agency. Publication date unrecorded; inherited source-text check.
S04 ISO/IEC. 2024.
S16 Microsoft Learn, Azure Architecture Center. 2026-06-03.
S22 National Institute of Standards and Technology. 2023.
S09 World Wide Web Consortium. 2013-04-30.
S38 International Organization for Standardization. 2014.
S13 ISO/IEC. 2023.
S1102 Google Developers. 2025-08-25.
S14 ISO/IEC. 2023.
S02 International Organization for Standardization. 2022.
R29 MDPI; industry inspection literature. Publication date unrecorded; inherited source-text check.
S36 World Bank. 2023.
S18 Advances in Neural Information Processing Systems. 2015.
S42 SCITEPRESS, IN4PL 2022. 2022.
R14 WageIndicator. Publication date unrecorded; inherited source-text check.
R32 Wise; XE. Publication date unrecorded; inherited source-text check.
S30 International Labour Office. 1997.
S27 National Institute of Standards and Technology. 2025.
S26 CISA, MS-ISAC, NSA and FBI. 2023.
S35 OECD. 2022-11-21.
S23 National Institute of Standards and Technology. 2024-02-26.
S25 National Institute of Standards and Technology. 2020-12-10.
S29 European Parliament and Council of the European Union. 2016-04-27.
S31 Presidency of the Republic of Cameroon. 2024-12-23.
S12 ISO/IEC. 2023.
S17 Microsoft Learn, Azure Machine Learning. 2026-01-27.
S40 International Labour Organization and UNESCO. 2020.
S37 World Bank. 2025-05-20.
S44 The Journal of Technology Transfer, Springer Nature. 2024-07-27.
R01 Business in Cameroon. Publication date unrecorded; inherited source-text check.
S21 arXiv, author manuscript of an IEEE Transactions on Knowledge and Data Engineering article. 2020-04-13.
Keep the thread · chapter 11 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.