GH-300 and AI-103 deep review — September 27, 2026¶
This review follows the AB-100 work and updates the complete GH-300 guide and AI-103 guide. It combines objective mapping, current primary-source research, useful blog selection, corrections, and original learning exercises. It is a same-context AI review; independent and human review remain pending.
Scope and evidence¶
| Check | GH-300 | AI-103 |
|---|---|---|
| Official objective baseline | August 7, 2026; unchanged | April 16, 2026; unchanged |
| Detailed objectives mapped | 41 in 14 groups | 64 in 14 groups |
| Weighted domains | 6 | 5 |
| Current guide citations | 69 registered: 67 reachable, 2 access-blocked | 52 registered: 50 reachable, 2 access-blocked |
| Current learning exercises | 7 labs and 13 answered scenarios | 8 labs and 12 answered knowledge checks |
| New blog readings | 2 | 2 |
Both the objective and delivery-status comparisons were unchanged. The accepted snapshots were preserved. The GH-300 blueprint's duplicated high-level features label is not an additional weighted domain. AI-103's historical Content Understanding terms remain explicitly mapped to retired and current product generations.
The public-access run covers 131 unique registered sources: 115 already associated with these guides plus 16 selected references. Fourteen references are newly registered; two existing Foundry references are additionally associated with AI-103. There were 127 successful HTTP responses and four blocked training requests. A second content fetch and visible-text capture distinguishes HTTP reachability from article readability. Course, channel, video, partner, and assessment shells do not establish access to their contents. Some Microsoft pages include sign-in notices above a publicly readable article body.
The blocked requests were the GH-300 O'Reilly video and Udemy hands-on course, and AI-103 Udemy courses by Alan Rodrigues and Luke Ginn. These results concern automated access, not proof that the pages are unavailable to learners. Paid lesson interiors and sign-in-only assessment questions were not inspected. Existing commercial durations and update dates are labeled historical in both guides.
The machine-readable evidence in ADLC_Docs/operations/2026-09-27-gh-300-ai-103-deep-review.json preserves prior catalog/review records, before/after guide hashes, each objective's position and hash with its guide mapping, source fetch observations, source-intake decisions, applied findings, and validation results. Objective wording is not republished in that ledger.
GH-300 decisions and changes¶
The guide now distinguishes an enforced feature policy from an overridable review-effort default. GitHub's September 23 release supports that distinction. Requesting an automatic review, setting review effort, and satisfying a merge rule are separate actions.
The organization-instruction section adds the supported JetBrains plugin/session boundary from the September 22 release. The guide also explains the forthcoming October 22 default-availability transition while preserving explicit feature choices, delegated organization settings, and preview exclusions. An announcement is not evidence that a future rollout has completed.
The Terraform workflow example previously used path filtering even though learners were instructed to require its result. A documentation-only pull request could therefore remain pending. The example now runs for all pull requests, includes the merge-queue event, and explains project-directory and version assumptions, following required-check troubleshooting. The workflow was reviewed as documentation; it was not executed against a Terraform project.
Four worked examples connect acceptance criteria, context selection, customization mechanisms, and productivity measurement. Two new labs exercise independent boundary tests and policy/merge controls. Five additional scenarios have answer explanations. The productivity example explicitly avoids treating sums of stage medians as measured end-to-end medians, and the new review-stage metrics are scoped to their qualifying human-review population.
The September 24 training event is now labeled past. The guide points learners to current credential and recertification information without converting a credential extension into a blueprint change.
AI-103 decisions and changes¶
The catalog prerequisite now matches the official Python requirement. The guide adds an identity matrix and a migration failure drill: copying an agent definition does not transfer downstream role assignments. The Foundry migration guidance supports the newer identity/endpoint explanation while continuing to describe legacy Agent Applications as supported during transition. No retirement deadline is invented.
Six worked examples follow a fictional maintenance assistant through identity, retrieval, tools, vision, speech, and document extraction. The original local Python example computes explicitly defined set-based retrieval precision and recall. It distinguishes relevance from authorization, ranking, provenance linkage, and claim-level groundedness. Empty-label questions receive a separate abstention treatment.
The tool example shows why HTTP success can coexist with acting on the wrong asset. It maps failure questions to current agent evaluators, while keeping approval and authorization in application controls. The current evaluation workflow informs version pinning, trajectory inputs, bounded polling, and accounting for missing or errored cases.
The SDK section adds an implementation checklist for compatible package versions, endpoint contracts, token audiences, runtime identities, and renamed RBAC roles. A new integrated lab can be completed with synthetic offline fixtures before optional cloud work. All six previous knowledge checks now have answer checkpoints, and six additional answered checks cover the new examples.
The release sweep included the July/August Foundry roundup published September 9, current SDK/agent documentation, and Content Understanding release notes. It did not treat every announcement as another exam objective or add redundant guide links.
Blog intake¶
| Source | Attribution and date | Accepted use | Limits and corroboration |
|---|---|---|---|
| Context engineering | Christina Warren; January 12, updated January 14, 2026 | Conceptual context-packet and customization reading | Public body read. The article's plural prompt-file suffix is corrected explicitly; official customization documentation governs filenames and supported surfaces. |
| Context handling and routing | Joe Binder; June 17, updated June 18, 2026 | Explain cached context and deferred tool definitions | Public body read. VS Code implementation, internal measurements, and rollout plans do not imply universal client behavior or guaranteed gains. |
| Debugging RAG agents | Chang Liu; November 20, 2025; no separate update date observed | Separate retrieval evaluation from response evaluation | Public body read. Historical clients and APIs require current documentation; agentic retrieval must earn its cost on the learner's own evaluation set. |
| Content Understanding August updates | Peyton Fraser, Krishnakumar Muthukrishnan, Joe Filcik; August 12, 2026; no separate update date observed | Identify GA/preview experiments and extraction tradeoffs | Public body read. Current release/mode documentation governs implementation; internal benchmark numbers are not acceptance criteria. |
These articles are supplemental expert resources, below exam blueprints and implementation documentation in authority. Reading estimates in the guides include time for notes and are editorial estimates. No exam-integrity concerns were found in the reviewed articles. Search results advertising dumps or recalled items were excluded. No article code, illustrations, paid lessons, or real exam questions were reproduced; the examples, fixtures, labs, and answer explanations are original.
Recurring maintenance¶
The existing weekly maintenance workflow continues to compare blueprints, lifecycle evidence, registered source health, and dated events. This pass adds GH-300 follow-ups for September 28 policy/retention changes, October 2 model removals, and October 22 default feature availability in config/certification-maintenance.json. Each event requests an evidence review rather than assuming a scheduled change happened.
For later deep reviews, use the same sequence: objective comparison; targeted product/release research; bounded blog intake; useful guide changes; source registration and evidence; repository/site validation. Automation can identify drift and prepare a review queue. Human or agent review must still decide whether a changed page affects the explanation, exercise, or answer. See automation and source freshness.
Validation and limits¶
Validation results are recorded after the checks complete in the operations ledger. Local Python arithmetic can validate the retrieval exercise; it does not validate Azure integration. Neither guide's new IDE, enterprise-policy, Terraform, or cloud exercises were run against vendor services. The guides remain unofficial AI-assisted learning resources with human review pending.