FQI Insights
Reports and assessments on AI, quality management, inspection processes and technical problem solving.
68 reports found
Multiple filters are combined.
How quality teams distinguish measurements, simulation and estimated error components before AI proposes machine corrections.
A Ford case study exposes a gap in AI detection evidence. An FQI evaluation plan includes components that produce no alert.
AI evaluators can rank reports correctly yet accept the wrong ones. An FQI pilot separates ordering, decision thresholds and aggregate quality indicators.
Synthetic training examples must not compromise an independent model test. An FQI plan checks data lineage, separation and realistic defect frequencies.
A successful tool call can still lose crucial information. A bounded test process makes coverage and limitations visible.
Movable and flexible measurement systems need a verified geometry state. A bounded acceptance process links pose, correction and outcome.
Protect current outputs, detect persistent quality loss and assess suspended AI agents separately before readmission.
Traceability and decision correctness need separate evidence. A preprint exposes the risks of indiscriminate rule-based overrides.
An FQI verification process binds AI-generated metrics to their source, context and calculation before they enter quality reports.
A signed agent credential applies to a tested configuration, task, scoring rule and validity period—not to every use of the agent.
Pneumatic tubing can attenuate, amplify and delay pressure signals. The released evaluation method must therefore match the required result.
An initial defect suspicion steers camera angle and zoom for a targeted second image. Quality depends on controlled remeasurement before judgement.
The number of generated AI answers does not fully describe the process. Batching, call structure and measurement boundary belong in the released recipe.
Large AI datasets need more than a scale record: structure, provenance, factual correctness and demonstrated effect are separate release gates.
Technical pre-action checks can expose silent AI-agent execution errors. A preprint illustrates their value, false alarms and limits.
Formal verification can confirm rules without confirming the real safety objective. Why the specification itself needs independent qualification.
Why an unremarkable group comparison cannot replace release evidence: verify the metric, sensitivity, equivalence margin and excluded cases first.
GPT-6 Astra is real and sets new performance marks. Why 99.9 percent on ARC-AGI-3 still does not prove general intelligence.
A new registry shows how documented agent incidents can become a controlled input to test planning and CAPA—without treating incident shares as failure rates.
An industrial field study shows how a canonical measurement record, persistent transfer and ERP reconciliation can reduce manual breaks in the inspection chain.
A counter-test removes uncredited factors and checks whether an AI decision actually follows its explanation.
OpenAgentFlow routes heterogeneous AI agents through one shared, auditable checkpoint before each consequential action.
Invalidation contracts reveal what stored agent hints depend on and when a change makes them unsafe to reuse.
Quantisation can change language-model behaviour, so the actual deployed configuration must be part of release evidence.
RiskBlend prioritises test cases by behavioural changes between model versions, making limited expert review more effective.
SafeStep shows how delayed sensor data can degrade a live decision even when transmission remains technically correct.
A study of more than 33,000 support cases shows how structured operating data can complement model selection, onboarding and monitoring.
A new method adds conditional confidence intervals to updated failure probabilities and exposes uncertainty sources that remain open.
A new study compresses AI-agent execution traces into small state models. FQI derives an auditable approach to runtime monitoring, stop rules and change control.
3D scans quickly produce dense geometry data. They become defensible only through intended use, traceability, validation and controlled approval.
A recent study separates retention from retrieval in agent memory. FQI turns that distinction into an auditable control for complete decision chains.
A study combines anomaly detection and consistency screening for battery cells. FQI identifies the evidence still needed before release.
A recent study connects defect localisation, structured process knowledge and human review. For FQI, the decisive feature is the auditable evidence chain.
Rockwell's Singapore plant combines more than 50 digital and AI-enabled solutions with measurable improvements in quality, productivity and workforce qualification.
An ARC-AGI-3 experiment shows how context handling and compaction change an AI agent's performance. The quality of the complete system is what matters.
The Digital Omnibus postpones key high-risk deadlines. Quality management can use the window to build a robust AI inventory, risk controls and release process.
A new FMEA-AHP method adds organisational impact to the classic criteria. The approach is transparent, but has not yet been independently validated.
The new joint SPC manual harmonises methods and terminology. Its connection to FMEA, the Control Plan and MSA/VDA 5 is what matters.
DataSpace shows substantial performance differences with the same model under different agent orchestration—a case for measurement system analysis and design of experiments.
How mobile inspection and a digital representation connect monitoring data—and which controls quality teams need before operational use.
Unilever reports measurable gains from digital twins, but for quality the decisive factor is a controlled loop of measurement, model, approval and response.
A peer-reviewed study shows how quality management, data science and AI agents can form one controlled improvement system.
A study shows how false information propagates through multi-agent systems. Quality management requires independent evidence and source provenance.
A vision-language model can be semantically strong yet spatially unreliable. Hybrid inspection architectures reduce risk but do not replace validation.
The sixth edition of ISO 9001 is under publication. Organisations can assess expected changes systematically without rebuilding their QMS prematurely.
A case study reports strong classification performance and time savings but incorrectly assigns problem description to D1 rather than D2. An FQI fact check.
DIN EN ISO/IEC 42001:2026-08 brings the European AI management-system standard into a German edition. FQI examines how it can connect to existing QMS structures.
Forty percent active AI use meets only 21% multi-tier traceability. In supplier quality, a reliable evidence chain must come before the risk model.
A controlled study of 48,000 scenarios reveals common AI reasoning failures, with direct consequences for evidence, challenge and approval in 8D D4.
Hybrid approaches combining SPC and machine learning make sense when stability, prediction, uncertainty and response plans remain clearly separated.
A peer-reviewed study translates unplanned product requirements into manufacturing operations. Controlled knowledge and hard release gates determine practical readiness.
AssertMate combines static analysis with multiple AI perspectives. Quality management still depends on test oracles, inspection criteria and evidence of effectiveness.
Article 50 makes one point clear: human review of AI-generated content needs criteria, competence, release authority and traceable evidence.
HopRefusalBench reveals weaknesses in recognising unsolvable tasks. Quality management must specify stopping, escalation and safe fallback processes.
TinyGLASS reports strong anomaly detection on an intelligent image sensor, but quality must still distinguish anomaly, defect and nonconformity.
A research model connects condition-based maintenance with product quality. The key is to assess failure risk and quality loss together.
How digital threads, data quality, and quality governance turn AI into a dependable loop—and why vendor metrics are not proof of impact.
A recent review shows how standards, interoperability and automation can move metrology into a continuous digital thread. FQI examines the quality-management implications.
A recent preprint shows how a few real defect images can support synthetic training data for inspection ramp-up—and what quality validation is still required.
A current industrial roadmap traces the maturity path from data and metrology through integration and system testing to controlled operation.
When AI creates work products, response metrics are insufficient. Process impact, defect escape, rework and controllability become decisive.
Real AI testing incidents show that test environments, orchestration, permissions, monitoring and external evaluation partners must be controlled.
A research approach shows how large language models can use historical events to create traceable hazard scenarios—and what FMEA can learn from it.
Learning agents modify prompts, tools and capabilities. Quality management must control changes, corrective actions and rollback.
A recent study shows how LLMs, ontologies and knowledge graphs can make existing FMEA knowledge reusable across manufacturing lines.
Fraunhofer projects connect vibration and AI with tool life and workpiece quality. The decisive point is a proven CTQ link, not just a condition forecast.
A US draft shows how AI systems can be tested for their intended use and linked to risk analysis, test planning, audit and release.
EMAIN reports new progress in zero-shot anomaly detection. For quality management, the decisive issue is how such inspection chains are validated, monitored and released.
No matching reports found. Remove individual filters or change your search.
FQI Insights collects expert reports and assessments of current AI developments in the context of quality management, FMEA, 8D, Six Sigma, CAPA and inspection processes.