New Generation Artificial Intelligence Major Project verification testing provides independent, standards-based evaluation of AI algorithms, platforms, and application systems developed under national-level research initiatives. Accredited third-party laboratories verify whether project deliverables satisfy the technical indicators defined in task statements, covering model performance, robustness, safety, and reliability. This article explains the testing scope, sample types, evaluation methods, applicable standards, and frequently asked questions.
Technical Overview
The New Generation Artificial Intelligence Major Project is a national-level science and technology initiative established to accelerate breakthroughs in fundamental AI theory, key common technologies, open-source platforms, and intelligent application systems. Research tasks under the initiative typically span multi-year development cycles with clearly defined milestone targets, quantitative technical indicators, and formal acceptance requirements.
Verification testing is the independent technical evaluation performed at key project milestones to determine whether the delivered algorithms, software, hardware, and integrated systems actually achieve the performance targets stated in the project task statement. The process generates objective, reproducible evidence that supports project acceptance review, risk management, and continuous improvement.
Verification activities are typically conducted by accredited third-party laboratories operating under ISO/IEC 17025 quality management systems, with reference to international and national standards for AI system quality, trustworthiness, and software testing, including ISO/IEC 25010, ISO/IEC 25059, ISO/IEC 22989, ISO/IEC 23894, ISO/IEC 24028, and GB/T 41867-2022. The evaluation combines functional verification, quantitative performance measurement, robustness and adversarial testing, and documentation review.
Testing Samples
AI Algorithms and Models
Deliverables in this category include deep learning models and algorithm packages for computer vision, natural language processing, speech recognition, multimodal perception, knowledge reasoning, and decision optimization. Samples are submitted together with trained model files, inference code, environment configurations, and declared performance baselines on specified benchmark datasets.
AI Software Platforms and Development Frameworks
Samples include machine learning platforms, distributed training frameworks, data annotation tools, model management and deployment pipelines, and open-source AI development environments. Verification covers platform functionality, interface correctness, multi-user concurrency, and compatibility with mainstream hardware and operating environments.
Intelligent Computing Hardware and Embedded Systems
This category covers AI accelerator chips, edge computing modules, intelligent sensors, robots, and unmanned equipment in which AI inference is embedded in dedicated hardware. Testing addresses computational throughput, energy efficiency, latency under real-time constraints, and hardware-software co-verification.
Domain-Specific AI Application Systems
Samples include integrated intelligent systems built for vertical industries, such as medical image diagnostic software, autonomous driving perception stacks, intelligent manufacturing quality inspection stations, and smart city management platforms. Verification is performed against both the project indicators and the functional requirements of the target deployment scenario.
Testing Items
Verification programs are tailored to the deliverable type and the task-statement indicators. Typical testing items include:
- Functional verification: confirmation that declared functions, interfaces, user roles, and operational workflows operate as specified
- Model performance evaluation: accuracy, precision, recall, F1-score, mAP, top-k recognition rate, word error rate, BLEU score, and task-specific metrics measured on standardized datasets
- Robustness and adversarial resilience: performance retention under data perturbation, noise injection, adversarial samples, and distribution shift
- Generalization capability: cross-dataset, cross-scenario, and cross-domain performance consistency
- Performance efficiency: inference latency, throughput, memory footprint, and resource utilization under representative workloads
- Reliability and stability: endurance testing, fault recovery behavior, and degradation analysis under sustained operation
- Safety and compliance: fail-safe behavior, output confidence handling, and conformance to personal information protection requirements such as GB/T 35273-2020
- Trustworthiness attributes: transparency of model documentation, bias assessment, and explainability support as guided by ISO/IEC 24028
Testing Methods
Evaluation is executed through a combination of laboratory testing and, where required, on-site verification in the actual or simulated deployment environment.
- Benchmark-based evaluation: model outputs are scored against curated ground-truth datasets using reproducible metric computation procedures, with fixed random seeds and version-controlled test data
- Black-box and white-box testing: interface-level functional testing is supplemented by code review and structural coverage analysis where source code access is granted
- Adversarial and stress testing: inputs are systematically perturbed using noise, occlusion, brightness variation, and adversarial attack algorithms to quantify robustness boundaries
- Load and endurance testing: concurrent user simulation and long-duration operation are used to measure throughput stability, resource leakage, and recovery behavior
- Standards conformance assessment: test design and reporting are aligned with ISO/IEC 25059 for AI system quality, GB/T 41772-2022 for face recognition system testing, ISO 21448 for safety of the intended functionality in autonomous systems, and related GB/T and ISO/IEC documents applicable to the sample type
- Scenario-based verification: for application systems, task-specific scenarios are reconstructed to confirm that end-to-end performance matches the acceptance indicators
Applications
Verification testing under this program supports national research initiatives and industrial deployment across key intelligent technology domains:
- Smart manufacturing: visual defect detection, predictive maintenance, and process optimization systems
- Intelligent and connected vehicles: perception, decision planning, and cockpit interaction subsystems
- Smart healthcare: medical imaging computer-aided diagnosis and clinical decision support software
- Smart cities: urban governance platforms, intelligent transportation, and public safety analytics
- Intelligent agriculture and environment monitoring: crop recognition, yield prediction, and remote sensing analysis
- Financial technology: risk control models, anti-fraud systems, and intelligent customer service engines
- Brain-inspired computing and new-generation robotics: neuromorphic hardware evaluation and embodied intelligent system verification
FAQ
What deliverables can be submitted for verification?
Acceptable samples include trained algorithm models, software platforms, AI chips and embedded modules, and complete application systems. Each submission should include version identifiers, environment requirements, user documentation, and the technical indicator table from the project task statement.
Which standards are applied during evaluation?
Testing references ISO/IEC 25010 and ISO/IEC 25059 for software and AI quality models, ISO/IEC 22989 for terminology alignment, ISO/IEC 23894 for risk considerations, ISO/IEC 24028 for trustworthiness, and applicable GB/T national standards such as GB/T 41867-2022 and GB/T 41772-2022, selected according to sample type.
What documentation is required before testing begins?
Typical submission materials include the task statement or contract extract with quantified indicators, functional specification, test-ready deployment package, self-declared baseline results, and data compliance statements for any personal information involved.
How long does a verification cycle take?
Duration depends on deliverable complexity and the number of indicators to be verified. A single model-level evaluation is generally completed within a few weeks, while integrated application systems involving on-site scenario reproduction may require a longer schedule agreed in the test plan.
What is issued upon completion?
An accredited third-party laboratory issues a formal verification test report documenting the test environment, methods, datasets, measured results, and pass/fail conclusions against each declared indicator. The report serves as objective technical evidence for project milestone review and final acceptance.
