PCEI-30-01 deep review — September 30, 2026

The guide now maps 33 numbered objectives, grouped 5/5/6/7/5/5, separately from 36 exam items, and expands 20 answers. Three exact original files execute 21 behavioral test methods and 20 observation checks per normal/optimized mode. A fourth optional pandas/Matplotlib file passes syntax and selected lint only. The result remains reviewed with blockers.

Objective and credential review

The complete indexed canonical English syllabus, detailed bullets and minimum-qualified-candidate profile were read after direct retrieval and the monitor failed with Win10060. Item allocations differ from numbered objectives because 2.4,4.6 and5.2 have two items each. The existing six-paragraph summary snapshot remains historical evidence; manual confirmation is not a successful fresh hash observation or full PDF review. Objective2.2's title mislabels its ML-workflow content.

The complete credential page confirms the active30-01 outline,36questions/60minutesplusNDA/75%/English/TestNow,seven-year validity and no formal prerequisite, with PCEP+PCED-equivalent preparation recommended. The public exam/retake listings are USD69/86. Practice tables name Q3/Q4 2026 while older FAQ/PCAI text names Q1 2026; PAI101 is in development. Those conflicting statements do not prove actual product availability.

The complete testing policy, revisedNovember18,2025, distinguishes global browser proctoring from local-partner procedures and preliminary scores from verified results. The specific seven-day rule follows failure; generic passed-retake wording remains unresolved. No account, booking, purchase, diagnostic, device permission or runtime setting was changed.

Executed original evidence

The exact guide fences for ai_workbook.py, test_ai_workbook.py, observations.py and optional_visuals.py were matched to private source hashes and compiled. The first three execute in four recorded runs with warnings treated as errors, normal and optimized modes. Both complete observation JSON reports match. Public checks remain active under optimization. Selected RuffE4/E7/E9/F/W/E501 passes at79columns for all four files; this is not an all-rule or security audit.

The CSV fixture has four accepted and four rejected records. It preserves the original string and rejected parsed cells/record numbers/ending physical lines. Exact header/shape/category/ID/numeric rules distinguish row rejects from whole-input size/syntax failures. IDs are reserved only by accepted records. The four accepted hour values produce mean4.5 and median4.0. This is in-memory synthetic data, not a disk loader or privacy-ready reject log.

Original distance cases switch neighbors between Euclidean and Manhattan geometry and between raw and training-range-scaled features. Exact ties use lexical ID order as a teaching convention. Training0/10 scales10to1; incorrectly including held-out100 scales that same training10to0.1. Correct transformation leaves held-out100at10 without clipping/refitting. Constant-feature drift explicitly raises an error. This custom policy is not a complete reimplementation of scikit-learn, which was unavailable. See the pitfalls and MinMaxScaler contracts.

With two positives among ten cases, the all-negative baseline has0.8accuracy,zero recall and undefined precision. Two fixed-score cutoffs preserve0.8accuracy but yield recall0.5versus1 and hypothetical 5*FN+FP costs6versus2. Fictional groupB has no positive cases, so its recall is undefined despite perfect accuracy. There is no calibrated model, threshold-selection study, population fairness conclusion or real business-cost evidence.

Executed NumPy2.5.2 float64 operations distinguish rowL1norms[3,4]frommatrixL1norm5, rowL2norms[3,sqrt8], columnmeans[2.5,1] andflattenedmean1.75. The norm and mean pages were read completely; their examples were not copied or claimed as executed.

The original sigmoid neuron yields weightedinput1.1 and output0.7502601055951177. All three analytic derivatives match central finite differences within1e-8 in the tests. One learning-rate0.5step lowers this example's half-squared loss from0.031185007428680883to0.0251708674548701. This is not deep-network training, calibration, convergence or generalization evidence. ExistingCPython3.13.14 andRuff0.16.4 were used without installation.

Learning and responsible-use boundaries

Complete public Google MLCC, Microsoft ML-for-Beginners and Elements of AI landings were read, plus Google prerequisites and classification/neural/LLM introductory pages. Linked lessons, notebooks, quizzes and model trials were not completed. All guide hour ranges are author planning budgets. Microsoft’s12week/26lesson/52quiz headline is distinguished from the visible numbered sequence. Elements separates Introduction to AI from Building AI, with different programming expectations.

Selected O'Reilly public metadata and opening contents confirm Hands-On Machine Learning3rd,AurelienGeron,October2022,864pages. Remaining contents and paid interior were not reviewed; no entitlement or purchase was exercised. Publisher reading estimates are not learner completion claims.

The complete NIST AI RMF landing says1.0is under revision; theApril2026critical-infrastructure concept is not a final replacement. The complete Playbook landing offers selectable voluntary practices, not a universal checklist, and expects an update after the revision. Linked detailed actions/PDF remain unread. The fictional project card assigns owners, baseline, costs, stop rules and fallback without claiming compliance or a completed business case.

Selected pandas, Matplotlib and OWASP contracts support proposed exercises and their limits. The five-case prompt matrix lists expected behavior only, without fabricated model outputs. No delimiter, keyword filter or second model is represented as proof against prompt injection.

Remaining work and validation

There are28direct source receipts:20successful,7errors and1blocked. Indexed fallback reading is separate from direct reachability. Three freshly retrieved source bodies match PCEA byte hashes, allowing reuse of specifically recorded CSV/JSON/Ruff reading ranges. Other selected/complete/unread boundaries are enumerated individually in evidence.

pandas,Matplotlib,scikit-learn,torch andTensorFlow are absent. optional_visuals.py was never imported or run; no table/plot was rendered, exported or visually approved. The twelve larger labs, fullPDF, live prompt evaluation, framework exercises and independent human content/accessibility review remain pending. No dataset-file write, network/modelAPI call, service, package installation or external message was performed by these examples. The five GitHub certifications remain reserved for the joint workflow test.

Repository evidence ADLC_Docs/operations/2026-09-30-pcei-30-01-deep-review.json records exact sources, hashes, four receipts, computed observations, dependency inventory and source-reading limits. Repository tests, strict site generation and repository/site/catalog/whitespace checks are attached only after their successful completion.