AB-100, GH-300 and AI-103 refresh — October 28, 2026¶
This is the next user-priority group after the first detailed Microsoft batch. The user requested these three guides and staging, committing and pushing after each validated grouping. GitHub remains a separate certification vendor. Three read-only researchers inspected complete guides, current exam evidence, selected product references and specific learning resources; the coordinator applied the changes. The operations ledger contains exact guide/objective bindings, all 179 accepted objective mappings, source observations, findings and limits. AB-100's 74 future objectives are mapped separately.
| Guide | Accepted teaching baseline | Objectives | Source gate |
|---|---|---|---|
| AB-100 | Retained July 22, 2026 | 74 | Blocked: prerequisite/source conflicts and future outline boundary |
| GH-300 | August 7, 2026 | 41 | Blocked: current CLI exclusion documentation conflicts |
| AI-103 | April 16, 2026 | 64 | Sources/objectives checked; human and cloud execution pending |
The accepted objective/status snapshots remain unchanged. A monitor result does not complete a guide review, and these research receipts do not establish human review, live product execution or complete independent semantic audits.
AB-100¶
The live study guide now displays November 16 future scope. It has the same 74 detailed tasks with three small branding/capitalization edits, but materially different plan/design/deploy weights: 35–40 / 30–35 / 20–25 rather than the retained July 25–30 / 25–30 / 40–45. Both study routes are explicit. The future text is saved under a separate future filename; it does not replace the accepted baseline. Localized appointment scope must still be checked. The exam/credential prerequisite disagreement involving MB-280 and PL-200 remains unresolved.
New teaching covers effective classic/advanced connector-policy coverage, unsupported connector classes, propagation and rollback evidence. A platform-elimination example compares actual knowledge/authentication/channel requirements with the harness matrix; ALM support alone cannot satisfy missing capabilities. Shared Toolbox defaults are part of the release dependency, even if agent packages and endpoints stay unchanged. The new fifteenth exercise requires an effective-control matrix and evaluation-identity coverage; two more answers bring the total to 26. The newer policy article is accepted for selected reading, while a bootcamp workshop with setup dependencies and conflicting generalizations is not endorsed as a generally runnable lab.
GH-300¶
The current six-domain, 41-objective outline is unchanged. The current GitHub-hosted outline API corroborates the Microsoft mirror but also retains inconsistent older skills/status fields. A search-discovered seven-domain PDF is legacy evidence, not a replacement for the current outline. Booking and credential status come from the current credential page, not an isolated old API flag.
Expanded teaching now provides IDE setup, Chat context/options and an ordered suggestion-troubleshooting process. New policy examples distinguish locked feature defaults, model defaults, delegated settings and overridable review effort. The dated October rollout's one-time assignment does not mean later global-default changes rewrite already-set feature policies.
The current exclusion how-to says CLI exclusions are unsupported, while the September GA announcement says Business/Enterprise CLI/app exclusions are supported. Both remain cited, and the source gate stays blocked. Harmless client/plan/policy fixtures are proposed; no CLI or IDE test is claimed. MCP teaching distinguishes applicable-client managed settings, server allowlisting, tool selection, credentials and downstream authorization. Cloud MCP calls and assisted permission decisions are not human approval. The eighth lab and three new answered scenarios reinforce those boundaries. Fusion is qualified related context, without vendor benchmark or universal-client claims.
Tim Warner's current companion now targets September's class and the August blueprint; June defects are explicitly historical. The complete MIT license was restored and independently fetched, and its exact notice is preserved in Third-party notices. Pluralsight now lists six courses/six hours. A specific GitHub-owned CLI chapter and separately timed Skills exercises replace vague combined timing. Paid lessons, question items and actual training workflows were not accessed or executed.
AI-103¶
The 64 detailed objectives remain unchanged. Four shallow teaching areas now have concrete decision workflows: visual watermark/symbol/brand/content rules; custom speech training and held-out evaluation; audio-native versus transcript-only reasoning; and structured extraction with independent application validation. Supported direct-inference structured output is kept distinct from unsupported agent/audio combinations.
An original Python draft validator passed 13 local fixtures. Three accepted shapes include a deliberately wrong-source asset: that acceptance demonstrates why schema/type validation cannot establish the correct asset, permission or approval. Ten invalid fixtures were rejected. These results cover only local deterministic code, not a model, SDK, speech, image, agent, authorization service or cloud deployment.
Current notes distinguish task-specific SDK requirements, vector filter placement, detection from safety enforcement, staged regional safety updates, evaluator input trajectories and shared-tool version drift. Minimum package/Python versions are source observations, not an installed/verified environment. The new ninth worksheet and four more answered checks bring the total to 16.
Two exact Microsoft Learning function/MCP labs have publisher times and access/cost caveats. The MCP sample automatically approves matching requests, so it does not demonstrate human approval enforcement. A LevelUp landing shell does not verify its current session details or prove retirement; no exact successor URL was invented.
Evidence preservation and review mechanics¶
All prior source-review, deep-review and AI-audit results are retained. AB-100 already
had an earlier partial source review on the same calendar date. The new record uses
supersedes_review_id to preserve that earlier record's exact counts and checks. The
schema/validator require same-exam, same-day source-validation references without forks
or cycles. A superseded pass cannot satisfy a current source gate. Six focused regression
tests cover history, blocked successors, invalid references, forks/cycles, audit order
and the legacy fallback.
Intermediate scratch mapping captions/bindings were corrected before publication. The authoritative final matrix uses immutable pre-group guide copies plus complete final diffs, accepted/future objective hashes and exact final span hashes. Earlier scratch bindings are explicitly superseded; no historical repository receipt is rewritten.
The sequential collector checked 199 initially cited sources: 193 reachable, six automation-blocked and none missing/error. Seventeen new sources were registered; the 20 new/reused guide associations checked separately all returned HTTP 200. Final per-guide counts are in the ledger. Body reading, metadata access, HTTP health, local fixtures and live product execution are recorded separately.
Validation and delivery¶
A separate fresh-context reviewer read all three complete final guides, checked the 179 current and 74 future mappings and their exact bindings, and independently inspected selected primary sources. The bounded review passed with no open findings after two low-severity repairs: narrower GH-300 citation wording and removal of inapplicable Learn sign-in boilerplate from non-Learn source metadata. The complete review and repair verification are embedded in the operations ledger. This is not a full independent semantic audit, human review or live product test.
All 199 unit tests, repository validation, reviewed-resource synchronization, the strict MkDocs build, website validation and whitespace checks passed. The reviewer also independently reran the six same-day lineage regressions and 13 local draft-validator fixtures. All 286 earlier source-review rows, 163 deep-review receipts, 42 AI-audit batches and six accepted objective/status files are preserved. The final report/ledger assembly is followed by another repository and website gate before staging and delivery.
The base is commit 86d2b8a, which published the earlier validated Bob intake and first
Microsoft batches. GitHub's validation and Pages-build workflows passed for that base.
This grouping is staged, committed and pushed only after its final local gate; remote
workflow results are checked after push. A successful Pages build on the maintenance PR
does not establish a production deployment.
Continue the Microsoft review queue after this priority group, preserving the user's per-group stage/commit/push workflow and the existing maintenance branch. This report does not claim that the entire vendor refresh is complete.