NCA-GENL deep review — September 29, 2026

The NCA-GENL guide now maps 33 detailed public topic bullets, adds 40 answered original prompts, three reasoning scenarios and eight proposed infrastructure activities. An actual NumPy CPU workbook passed 23 checks. Independent human review and GPU/LLM execution remain pending.

Scope and monitor repair

The current canonical page replaced its former training/weight table with detailed preparation sections. Current page order is Core Machine Learning and AI Knowledge 30% (10 bullets), Data Analysis 14% (5), Experimentation 22% (7), Software Development 24% (7) and Trustworthy AI 10% (4). Weights remain the same; the data-analysis title, presentation, detail and recommended materials changed. The former introductory 50-question statement is absent; the detail panel still lists 50–60. The unchanged lifecycle snapshot records one hour, English, USD 125, basic knowledge and two-year validity. No new effective date was invented for the page redesign.

The previous monitor failed because it required the former “Exam Blueprint” heading. It now accepts the exact exam-code preparation heading while retaining the legacy branch. New regression cases cover detailed extraction/footer exclusion, another exam's heading and malformed or empty weighted sections. Forty focused monitor tests passed. Prior objective/status bytes and the historical review/audit reference were preserved before explicit baseline acceptance; the post-acceptance monitor reported unchanged.

Teaching and current documentation

The guide separates fitting, inference, prompted examples, learned adaptation and retrieval. Current NIM documentation introduces 2.x version assumptions and distinct offering/support decisions. Selected release notes motivate request-contract regression checks and separate container from host-driver remediation. Triton repository/readiness/inference concepts remain distinct. No serving software or models were launched.

The local example executes least-squares fitting, held-out prediction and losses, then fixed-vocabulary cosine retrieval. Training fits exactly, but held-out MAE is 6 and MSE 72 with one error of 12. Eligibility filtering excludes forbidden, other-tenant and outdated documents before ranking. An unfiltered top hit demonstrates disclosure risk; filtering only that hit misses an eligible result. Precision 0.5 and recall 1.0 describe retrieval under the original fixture labels, not generated-answer quality.

All data and decisions are synthetic. NumPy 2.5.2 was already installed; no package installation was needed. The code performs no file writes, network calls, account actions or model downloads. It is not an LLM, learned embedding, complete RAG system, authenticated permission service or production benchmark. Scikit-learn references support selected evaluation concepts, but that library was not installed or executed.

Resource access and remaining limits

Twenty-two distinct sources were freshly fetched: 21 returned HTTP-success evidence and one commercial catalog was blocked. HTTP success did not establish readable course metadata. Four NVIDIA course endpoints returned a shared blank template with placeholder workshop entries; another returned a maintenance/moved notice. Their earlier runtimes and prices remain unverified. The current preparation page instead recommends public Fundamentals of Deep Learning materials and Rapid Application Development with LLMs. The public repository README lists six modules; its notebooks and compute setup were not executed.

The old NIM introduction URL returned an HTML meta-refresh with no substantive text. Its receipt was preserved, the actual target followed and the registered source updated to the overview. The 2023 prompting article's early-access service and timing examples were treated as historical. The 2025 RAG article supplies conceptual context without adopting performance or privacy claims as guarantees. NIST's AI RMF portal distinguishes voluntary guidance, the published generative-AI profile and ongoing revision/concept work.

Read boundaries and exact local execution receipts are stored in ADLC_Docs/operations/2026-09-29-nca-genl-deep-review.json. Repository tests, repository/catalog validation, strict site build, generated-site validation and diff checks are attached when completed. The five reserved GitHub guides and disabled notification pilot remain untouched; no new accounts, permissions, installations, services or external messages were used.