حافظ على فائدة الذكاء الاصطناعي مع تغيّر الواقع

Breck et al.; IEEE Big Data · 2017

Peer-reviewed research · Abstract

A 2017 production-readiness rubric with 28 tests and monitoring needs; not a validated universal score or an AI certification.

Sambasivan et al.; CHI · 2021

Peer-reviewed qualitative research · Abstract

Qualitative interviews with 53 practitioners in high-stakes AI settings; not a representative estimate of failure prevalence or causal effect size.

Koh et al.; ICML · 2021-07-16, arXiv v3

Peer-reviewed benchmark research · Abstract

A historical benchmark of 10 datasets with naturally occurring distribution shifts; not a prediction of current enterprise or generative-AI performance.

Yao et al.; NeurIPS Datasets and Benchmarks · 2023-01-16, arXiv v2

Peer-reviewed benchmark research · Full text: Section 5.2, Tables 1 and 20; Appendix A.4.1 and Table 6

Historical temporal-shift benchmark. Exhibit uses one HuffPost ERM Eval-Fix comparison, averaged over three random seeds; dispersion across seeds is not a confidence interval for all future deployment conditions. No inference to current LLM accuracy.

Es et al. · 2025-04-28, arXiv v2

Research paper · Abstract

Automated, reference-free RAG evaluation methods; scores need local calibration and do not replace verification of domain-specific truth or authority.

Liu et al.; TACL · 2023-11-20, arXiv v3

Peer-reviewed research · Abstract

Experiments on specified historical models and tasks; position effects do not establish that every current long-context model behaves identically.

Google Cloud · Living guidance, accessed 2026-10-10

Vendor technical guidance · MLOps levels; automated validation and monitoring

Vendor architecture guidance describing automation levels and pipelines; does not prove that maximum automation is optimal for every organization.

Google SRE · 2016; accessed 2026-10-10

First-party engineering practice · Managing risk; error budgets

Site-reliability engineering guidance about reliability objectives and error budgets; applying it to AI quality is an analogy and must not make prohibited actions budgetable.

Yang et al.; NeurIPS 2024 Datasets and Benchmarks · 1 November 2024, arXiv v2

Peer-reviewed benchmark research · Section 3.1 Table 3; Appendix A.1.2 Table 7

Constructed benchmark distribution is not enterprise prevalence; historical model performance does not predict current deployment.