EX267 deep review — September 28, 2026

Same-context AI review; independent human review pending. Study guide.

Scope and unresolved source differences

Read all 43 current main-page tasks, in twelve groups with 3/4/2/2/6/4/3/4/3/5/3/4 tasks. The exam and AI267 course retain OpenShift AI 3.3 and OpenShift 4.20. Objective and lifecycle monitor hashes are unchanged; accepted snapshots were not replaced.

The EX267V33K section in the public version PDF has 41 matching tasks after normalization. It omits the main page's Identify and allocate resources group: selectors/tolerations and placement of workbenches/model servers on specific nodes. All 43 canonical main-page tasks remain mapped. Same-day directly fetched PDF bytes were reused from EX280; EX267 sections were separately extracted, read and compared.

The exam also names Standard and Advanced serving modes. The 3.3 deployment guide uses RawDeployment and an Advanced settings wizard. Those labels do not establish a mapping to old Serverless/Knative or ModelMesh recipes. Retain the objective and verify the assigned environment. Both source discrepancies remain blocked findings with an October 5 library follow-up, not a vendor transition date or a request to change the exam scope.

Implementation and support repairs

Reviewed targeted implementation sections in twelve 3.3 books, rather than treating the documentation landing-page index as implementation evidence. Added Hardware Profiles versus legacy accelerator/size selection, connection protocol annotations, registry v1beta1, image publication versus workbench creation, placement diagnostics, modelcar layout/permissions and deployment strategy capacity/downtime.

The release notes cover 3.3.6 and label LLM Compressor integration Developer Preview. This remains an exam topic, but its presence in objectives is not a production-support guarantee. The registry guide limits its URI-registered deployment workflow to public OCI repositories; this is not generalized to all direct-serving storage mechanisms.

The pipeline guide supports Python 3.11+/KFP 2.14.3+, two compiler output formats, explicit cache controls and image-specific Elyra. Added hidden mutable-input cautions and the documented UBI 9/RHEL 9 requirement for custom pipeline containers on FIPS clusters. The LM-Eval guide distinguishes platform permission from job requests for internet/code access. The guardrails guide supports separate detector, Orchestrator and gateway checks.

Original example and actual execution

Added an original two-component KFP graph with typed Dataset, Metrics and Artifact ports. Preparation validates a synthetic binary-prediction dataset and emits canonical bytes plus their digest. Evaluation verifies that digest, scores overall and cohort accuracy, checks the declared population/minimum sizes, and emits evidence that explicitly still requires separate approval. No deployment step exists.

39 local checks passed using Python 3.13.14 and KFP 2.14.3. They execute both original component Python bodies with real KFP artifact/metric objects. Cases cover a passing population, 90% overall accuracy hiding a 0% cohort, missing/extra/undersized cohorts, duplicate IDs, field types, malformed thresholds, changed artifact bytes and exact boundaries. The compiler produced both plain PipelineSpec IR and Pipeline/PipelineVersion resources with the same embedded specification, correct dependencies/resources/cache settings, and rejected an incompatible Dataset-to-Model edge. The initial negative-test harness caught the wrong exception class; it was corrected to expect the SDK's InconsistentTypeException, and the complete 39-check suite passed.

These are local function/compiler results. No container image was pulled, no pipeline backend or cluster ran, and no model inference, training, LM-Eval, GPU, S3, database, registry, serving endpoint, TrustyAI service, guardrail deployment or durable restart was executed. The image string is a compilation exercise; deployment requires an approved pinned environment. All eight complete platform labs remain proposed. Synthetic quality thresholds are illustrative and do not establish production fairness or representativeness.

Blog and learning catalog

Ana Biazetti, Nelesh Singla and Matt Prahl's June 3, 2026 modular-components article informed early validation, narrow interfaces and small-data checks. Original code was independently authored. Its stability labels and agent workflow are not adopted as production guarantees or session instructions. Heather Sherwood's July 8 training announcement is lifecycle context, not current pricing or entitlement evidence.

Retained and strengthened the guide's existing 40 answers. OReilly and Udemy pages blocked rechecking; inherited runtimes are now explicitly unverified. Places to learn is the final section. Repository validation, 176 unit tests, strict site build, generated-site validation, learning catalog consistency and diff checks are recorded in the operational receipt only after execution.