EN validation

Workshop validation ledger

Overall verdict

All labs | Français | Evidence gallery

Lab 6 acceptance reconciled, 2026-10-26

The T017 documentary crosswalk is delivered with the updated guide, not whole-lab runtime acceptance. F6-01/02/03 are historical offline PASS, not runtime passes. F6-04 is PASS; F6-05 is PASS at document level with the displayed-excerpt limit below. F6-07 is PASS and F6-08 is response-only PASS; F6-06/12 are PARTIAL. F6-09 is PASS under an explicit language override; F6-10/11 are NOT_RUN. All twelve rows retain source provenance, public evidence pointers and limits. At the initial offline checkpoint on 2026-10-26, F6-04/05 were NOT_RUN, F6-08 was PARTIAL and the crosswalk reconciliation submitted zero questions/live attempts.

The exact multi-source check passes F6-07: one question, one completed native MCP invocation, two pinned-corpus records and two applicable groups, investigation procedures and regulatory compliance. The attributed comparison invents no rule and has a noncancelled completed plan. This prospective MCP proof does not join the historical original specialist's tool output.

The routing-repair receipt preserves the original F6-06 FAIL: the unanchored intake matched hi inside Which. Only that condition changed, adding ^...$ and reviewing a case; the saved native topic passed its checker. A separate four-question, four-conversation cohort had no in-cohort retries. F6-06 now passes supported MCP retrieval with two verified records and a completed plan, but remains PARTIAL because the original connected-Foundry-child invocation is unproved. F6-04/05 regressions pass within their existing limits. F6-09 FAIL is preserved: grounding and core lineage passed, but final prose was English. Topic component-version proof remains NOT_ESTABLISHED; no agent publish occurred.

The language-configured attempt added secondary fr-FR, leaving primary en-US unchanged. Its new English Send displayed Cannot send empty message: one intent/click, zero observed accepted questions within 240 seconds, server acceptance NOT_ESTABLISHED. French was not selected or sent; this attempt is NOT_RUN, not a replacement for the completed FAIL. Test-session language and English restoration remain unverified. Automatic detection and complete static-topic localization are not proven. Publication submits no questions.

The later single French-text check proved delivery, not French acceptance: one accepted question after verifying the real textarea and 60/2000 counter, but activity and reply locales were en-US and final prose was English. One completed MCP call returned three verified records, including both core records from the earlier supported English base; its corpus, instructions and model were unchanged, but French secondary was added afterward. The noncancelled plan completed in 71.232 seconds from server receipt to reply (76.134 seconds from Send). No new English question or resubmission occurred. Both language radios still exposed aria-checked=false after a separate native keyboard control without a question. That historical F6-09 remains FAIL; its selection cause is unproved. The previous delivery failure and guarded NOT_RUN remain intact.

The supported localization repair used official PAC 2.14.1 through the existing enterprise NuGet proxy, without authentication or TLS changes. The native French upload succeeded; independent exports prove all 155 English values unchanged and 97 translated runtime values persisted after label/trigger normalization. Complete Windows extended-length enumeration verifies all 39 existing core draft files byte-identical and 22 French translation files added; the earlier 35-file short-path subset was incomplete. After a full reload, French selection still leaves both radios aria-checked=false, and a no-question fresh session greets in English. At that localization-only checkpoint the French question was NOT_RUN and the last completed F6-09 was FAIL. Full static localization and the selection cause remain unproved; see #5.

The later explicit language-switch cohort passes F6-09: executed System.User.Language reads back French, the unchanged question receives French prose and a completed grounded plan, while actual posted/reply Activity.Locale remains en-US. The English regression also passes; F6-06 overall remains PARTIAL because the original Foundry-child route is unproved. Two fresh conversations contain five accepted messages: two domain questions and three deterministic controls, no resubmissions. Each domain turn completes one MCP operation; English retrieves one record and French two, sharing one ATO core record. English restoration is proven in the same second conversation. All 1,288 historical files retain hashes/timestamps; a fresh complete clone verifies all 63 non-token files and 22 translations unchanged. Native picker repair, automatic detection and continuous execution-version proof remain unproved. Earlier FAIL/NOT_RUN results and the collector's failed echo wait remain preserved; no message was replayed.

The later two static French wrappers pass their scoped native checks: one fresh conversation, four accepted messages (two lookups and two language controls), four Send clicks and zero retries. CTR and SAR replies exactly match the corrected French resources, retaining all 12 literal quotations and individual references. They are static snapshots, not live retrieval; native retrieval/MCP/Foundry invocation counts are zero. French System.User.Language and English restoration are proven; actual posted/reply Activity.Locale remains en-US. The persisted override is restored, not cleared. One localization upload persisted both wrappers. Strict candidate-map and two-file-only guards failed and remain failures: a read-only audit explains 41 reordered positions with all 42 Escalate phrases preserved, one materialized French fallback component and component bookkeeping. Complete clone accounting is 64 to 65 files; the native postflight preserves all 64 non-token files, 23 French translations and 50 components. All 158 English/153 French resources and 1,814 historical files remain exact at that postflight. A strict UTF-8 BOM failure is retained; BOM-aware validation used unchanged files without another export or message. This separate cohort does not prove full static localization, picker repair, automatic detection or T017 completion.

The exact private-history check is F6-08 response-only PASS: one question, no replay or extra Hello, on the same frozen prospective configuration as F6-04/05. The complete final reply acknowledges that configured evidence cannot establish private customer history and refuses inference, without invented events or fictional-corpus substitution. A native final message and matching terminal fallback EndDialog/completed no-answer state are proven, but plan closure remains unproven: no DynamicPlanFinished within 240 seconds. One completed local search returned three records. An MCP label lacks correlated invocation/output; actual MCP and underlying Foundry counts are NOT_ESTABLISHED, not zero. This is distinct from original matrix case 9, not architecture equivalence or T017/workshop completion. Publication submits no questions.

The two exact native checks record the earlier cohort: two questions, one per fresh conversation, no preflight questions or replay. F6-04 completed clarification with activityProcessed=true, no retrieval or Foundry/MCP evidence calls (0.118 seconds from server receipt to final reply). F6-05 completed one local SearchAndSummarizeContent, a noncancelled DynamicPlanFinished and no Foundry/MCP evidence calls (70.618 seconds). Document-level support for Account Takeover is proven in another actually retrieved chunk of the same source; the displayed citation excerpt lacks the ATO row. Exact citation-chunk support, excerpt fidelity and global citation quality are not proven. This frozen prospective MCP-adapted existing original specialist configuration uses actual Ready MIT publisher Markdown, not historical baseline equivalence; original DOCX ingestion remains Error. Postflight root/settings, instruction digest, source versions, publisher hashes and archives were unchanged, without a second full component export. The live checks edited/published nothing and created no resources. Publication submits no questions. No further export or secret is needed for these two completed checks.

The original specialist native 4/4, hosted diagnostic extension and data-only MCP candidate are separate routes. Original corpus matches do not independently join the specialist's underlying native retrieval-tool output. The previous candidate 11/11 includes one explicit retry. The current post-control cohort is 10/11; case 9 failed twice, root cause unproved, no third replay. Preserve those failures and historical quotation passes without cross-cohort substitution.

Intake, local knowledge, grounded expertise, multi-source work, three datasets and cleanup guidance are source-backed. Syntax/layout checks, focused prompt variants and French/private-customer/denied-identity/missing-Activity controls are guide validations, not newly universal original requirements. F6-05 wording appears in setup instructions; execution was unproved at the initial offline checkpoint, before the new receipt. Remaining exact-variant gaps are reproducibility limits, not a reason to repeat verified source behavior. Additional protected-main requirements remain UNKNOWN, without bypass or reconstruction. Native annotations, publisher URLs and verbatim quotations are not universal original gates; historical tests retain their actual outcomes.

Lab 5 Search denial does not prove F6 Foundry denial; blueprint HTTP 400 does not prove the missing-Activity negative plus repaired retest. T003/T004's hosted v3 repair passed 3/3; hosted v5 remains retained and native delivery is not certified. Earlier cleanup receipts prove only their closed inventories, not current zero resources or final billing. T017 remains unchecked, workshop incomplete. Resume one unresolved criterion only with approved scope and a supported prerequisite, not whole-lab restarts. The chronology and original review baseline below are historical, not a replacement for this current reconciliation.

Earlier execution chronology

The integrated Search probe uses the native Microsoft Entra ID Integrated choice and documented ai:// namespace. One Create encounters an explicit consent-popup-blocked error; OAuth never completes, Next stays disabled and the picker never runs. No picker HTTP status/body or effective native principal is established. VS Code documents popup blocking; normal OAuth in a supported external browser is the next prerequisite, not another question or permission expansion. Cleanup verifies seven original connections and no persisted probe connection, without guessed deletion. Long-path-safe 64-file snapshots preserve 63 non-token files, 22 French translations and 49 components. Zero source additions, questions, agent saves, publications or Azure writes; no continuous external-state audit. This is not a repair of the earlier service-principal failure. S5-10/T014/T016 and native retrieval/paired-denial acceptance remain open.

The two new native Search OAuth service-principal cohorts each save Connected after one Create, then fail index selection with Something went wrong., console HTTP 400 and Add to agent disabled. Index-only Data Reader lacks metadata listing; adding a separate service-scoped Microsoft.Search/searchServices/indexes/read Action with no DataActions, writes or operations/read passes direct listing/own synthetic query, not the native picker. Its exactly two assignments are verified. Response body, cause and effective native principal remain unavailable/unproved; no fix is inferred. Zero sources and zero native positive/negative/model questions; no binding, mapping, readiness, post-revocation direct 403 or no-leak acceptance. Connected/direct controls are not native or partial paired acceptance. Both independent cleanups leave zero owned resources and remove credentials and protected secret files, preserving all original controls and 181 files. The first PAC filter and DNS verification errors remain recorded with separate recovery; the second cleanup/verification has no errors. Diagnostic issue #4 tracks the supported contract question. S5-10/T014/T016 remain open.

The current endpoint-model check returns two identical successful agent GETs. The original hosted v5 already exposes the new endpoint model with unchanged Activity/Invocations and Entra/BotServiceRbac; publication approval is not_published. No write or question occurs. New documentation does not prove a native fix or erase six legacy application failures; T005/T006 remain open.

The offline caller-ID correlation locates the historical mismatch: activity blueprint ID equals the approved user's directory object ID, while the summary ID matches the unchanged target blueprint. No new native test, identity override or supported fix is claimed.

The interrupted paired-SP cohort has zero authorized/denied native question intents. Owned credential, identity, draft, four roles and Azure fixture are independently absent after cleanup. Original inventories and 181 historical files are preserved. Native portal inventory redirects to sign-in; official PAC/Dataverse evidence is not portal absence verification. No paired native acceptance or finalized spending is claimed; S5-10, T014 and T016 stay open.

The native Entra Search follow-up passes S5-04/S5-07/S5-08/S5-09 for the isolated fictional fixture. The original 3/4 cohort retains its French-to-English failure. A separate French-enabled 4/4 cohort verifies actual native retrieval, exact facts, price abstention and individual citations. French is explicitly selected in the test pane, not inferred from the recorded en-US client locale. One genuine native click opens the immutable source file, without its section fragment. One diagnostic configuration payload is empty; independent before/after official clones verify all 17 authored files unchanged. The owned draft, connection and index are removed and original objects preserved. This is not an original hotel PDF, denied-identity or whole-lab pass; T014 stays open.

The October 26 guarded native recovery submits one first-use closing-time question after actual fresh authentication, the 600-second wait and a passing evidence guard. Direct Search is 403; native Search reports ConnectorRequestFailure / AzureAISearchClientError without an exposed HTTP status, with zero records/citations. The parent gives a generic rephrasing response, not a verified permission-specific error. No source fact is returned in that single response. This is not paired authorized-native/removal or independent-identity acceptance and does not close S5-10/T014. Cleanup and original inventories are independently verified; earlier outcomes are unchanged.

The separate isolated permission-removal run fails S5-10. Only a new test Reader grant is removed on a separate keyless Search service. Direct queries return 403, but one fresh native conversation returns 07:00 with the actual Harbor citation and two fixture records in its Search trace. This uses the same approved administrator, not an independently unprivileged identity; native effective principal and cause remain unproved. The official snapshot is recovered after a read timeout without question replay. Both configuration payloads match and all 17 authored files stay unchanged. All owned test objects are deleted, originals preserved and browser closed. Actual spending is not finalized. The earlier 4/4 is not promoted to an isolation pass; T014 and the full roadmap remain open.

(../assets/runtime-2026-10-25/lab1-history-citation-followup.json) The October 26 Dataverse/history check verifies recovered approved-admin Graph access, not Dataverse transcript rights. The original-agent metadata query returns 403; the fresh capability audit still lacks transcript-read and role-assignment privileges. Roles and policies remain unchanged. The exact original History run still exposes only its greeting and question, In progress. No replay, answer reconstruction or broad self-elevation occurs; T009 stays open.

The later original licensing History recovery finds the exact original question but no completed historical answer, even after one Refresh. One citation click in a separate completed history exposes no destination or PDF; it does not substitute for the original audit. No new question or configuration write occurs. T009 remains open. The verified wiki snapshot passes human-edit baseline review, normal publication and both rendered-language checks. The destination is public; visibility is unchanged. Publication progress does not close T016 or the remaining acceptance and retention gates.

The latest authenticated candidate continuation resolves administrator PAC sign-in and official candidate cloning. Corrected literal CTR/SAR lookups and source-free intake pass native 3/3, then 3/3 again on the final draft after a scoped broad-SAR instruction repair. Exact table values and individual attributions are verified; static snapshots do not certify native citations. The broad SAR route and coverage pass, but whole-output fidelity still fails: 15/16 quotations exact across two messages. The installed external Activity action schema supplies no supported message-suppression/capture switch. No unsupported configuration or protected-retriever change is applied. Independent preservation confirms the original's complete 26 rows unchanged, the candidate's intended 28 components, diagnostic Off/two specialists On, files, bindings, published timestamp and authentication preserved. No publication occurs. T007/T008 and the full roadmap remain open. Earlier unavailable-browser, wrong-PAC-tenant and clone-blocked statements below are chronological evidence, not current blockers; the plugin dependency-policy failure remains unresolved.

An isolated local MCP tool now passes real protocol discovery and one unchanged SAR request with two exact source records and correlated provenance. The actual tenant wizard exposes manual OAuth, but was cancelled without a connection. Local signed-token tests do not establish real Entra MCP OAuth.

The subsequent hosted preflight now passes separate App Service startup, HTTPS/TLS and disabled basic publishing, missing/invalid bearer 401, and correct protected-resource discovery 200. Two owned single-tenant apps, the scoped outbound managed-identity configuration and an official MCP connector with independently registered callback are persisted. The native consent popup is blocked; normal new-tab authorization reaches the administrator password prompt. No human response, successful delegated token, managed-identity retrieval, candidate binding or new native question is claimed. The temporary helper/redirect are removed. The B1 host remains billable at a published compute estimate of USD 0.018/hour, not an actual bill or approved budget. Fresh preservation confirms unchanged original/candidate content, files and bindings; the existing version-5 definition and principal/client IDs also match. Whole-output fidelity and the complete roadmap remain open.

The actual managed-identity runtime proof supersedes the outbound-retrieval boundary above. The unchanged SAR question returns rc-002 and ip-006 inside the actual App Service app container; both full records and rendered text match the pinned corpus independently. Microsoft's signature, intended tenant/identity, managed-identity-only credential chain, deployed source and protected retriever preservation are verified. Two earlier HTTP 403 attempts remain failures. Only one project-scoped custom role remains after removing three owned trial assignments. Its exact agent read/write actions also permit authoring; no agent-authoring request is made. The fresh full native readback passes again. Inbound delegated MCP OAuth, native binding, whole-output fidelity and the full eleven-case matrix remain unverified. No new native question, MCP tool invocation or publication is counted.

The later native MCP SAR recovery supersedes those OAuth/binding and bounded SAR-fidelity limits. Human-explicit tenant-wide consent is independently read back as AllPrincipals for Evidence.Read only. One saved candidate MCP binding and exact Author-validated instructions produce one complete native case-6 answer with two exact full-record quotations and individual canonical markers, plus actual local search/reference. The old emitting retrieval tool is saved Off; validation and MCP remain On. Original 26 rows, all other candidate definitions and missing/malformed-token 401 controls are preserved. No publication occurs. Raw DynamicPlanFinished, the fresh eleven-case matrix and broader roadmap gates remain open.

The initial full-matrix attempt passes cases 1-4 and stops at case 5's seven unattributed bullets. A saved repair passes fresh case 5, then the next SAR fails a misattributed renderer notice. The next saved repair removes that error but fails added punctuation in a local quotation. Actual generative completion events are now retained. Two actual evidence steps in one turn are counted separately from its single user submission. Strict quotation checks remain unchanged; no complete post-repair matrix passes yet.

The later complete eleven-case run now passes all original cases on one immutable saved configuration. Twelve submissions include one preserved case-7 stall and one separately guarded retry after exact readback and normal reconnection. Eight generative completion events, three deterministic processed-topic completions, seven genuine MCP invocations, actual intended local-source output and full reply/source/semantic audits are retained. Twelve local quotations remain exact; final MCP domain answers use individually attributed unquoted paraphrases. T008 closes for this run, not the roadmap. A separate unchanged validation control produces two identical replies through topic-then-direct-action invocation, leaving T007's control repair open.

The subsequent control repair and regression proves one real validation invocation/binding/reply after one counted recovery, without a hardcoded reply. Its fresh unchanged-question regression accepts 10/11 cases. Case 9 completes both searches and its plan without a parent answer; its counted retry has no plan or answer. Both failures remain. Fifteen total submissions include thirteen original questions and two controls. Fourteen available agent-content exports match; seven genuine MCP calls include the failed answer's successful retrieval. Case 2's lost export is not accepted until its separate recovery. Current case 7 has two exact attributed MCP quotations and allowed verified local context. Final full preservation passes and the owned browser closes. T007/current-configuration acceptance remain open; the preceding T008 run is not reused to certify this configuration.

The October 25 scoped inventory reads six owned ARM resources, including the isolated MCP site/B1 plan. A retained HTTP 429 billing failure precedes one counted successful recovery. Reported pretax cost is USD 11.26126391; only two resource IDs have posted rows, so this is not complete current spending or a final invoice, and excludes Copilot credits. No configuration writes, inference or native questions occur. User-unavailable form placeholders do not supply retention ownership, a review date, numerical budget or role-attributed sign-offs. T001/T002 remain open.

The latest verified Invocations delivery is active hosted version 5 with Activity retained, 54 service tests and seven passing hosted controls. Real local/hosted SAR output matches the exact two-record renderer. No-token denial, parser/body limits, abstention and corrected Activity delivery are verified; invalid inputs have no model calls in correlated logs. The deployment CLI timeout and first Activity missing-version HTTP 400 remain historical failures, not hidden successes. Independent readback and doctor (11 passed, zero failed, two skipped) establish active service without redeploying. The native preservation checks pass, but native acceptance and published SDK authorization remain unresolved. The shared browser is hidden and normal UI recovery fails; no new native question or configuration is submitted. No generated suite, retention decision, numerical budget or organizational signoff is invented from user unavailability.

The post-publication protocol follow-up recovers the verified administrator's browser but not tenant hosted-operation availability. Supported candidate cloning is blocked by an unavailable management skill; no YAML or remote configuration is invented. One fraud-candidate CTR citation check opens the expected uploaded source dialog, but its displayed passage does not support the threshold. This is not the original Lab 1 licensing-PDF navigation test, nor a native/full-matrix pass. The new failure remains retained.

The earlier authorized release deploys active version 3 and closes T003/T004 within hosted scope: 3/3 real positive, abstention and input-error Activity checks pass; 38 extension tests pass locally. The earlier version-2 blank-output failure is preserved. The reported AADSTS65002 login mismatch was resolved through one owned secretless PKCE client with current-user-only permissions. The later authenticated native canary uses the saved two-case test set and explicit connection binding: HTTP 202, Completed, one General Quality Pass and one Fail. Answer text and question-to-result mapping remain unavailable, so these grades do not establish whole-parent source acceptance or resolve the earlier native 403. The diagnostic was restored to Inactive with independent readback; its exact definition, original attachment and parent were preserved. The scoped transcript read was denied for missing prvReadconversationtranscript; no role was broadened. See authorized-release.json. Native/full-matrix, remaining labs, retention and final publication remain open.

The independent infrastructure reconciliation also completes T013's comparison work without reapplying resources. Official PAC 2.14.1 and checksum-verified Terraform 1.16.6 are restored. A byte-exact private copy of the original 23 Terraform files passes init/validate and a saved 15-instance plan; the plan-only group is absent. D5-02/D5-03 now pass. Historical corrected apply and cleanup remain separate from the unchanged plan's rejected model, public/authentication defaults and missing Search identity. No new deployment, full private-network or connected lab pass is inferred.

The earlier October 23 checkpoint adds fresh resource/cost inventory and authenticated saved-state reads. At that earlier checkpoint, the local SDK repair passed 33 tests but remained undeployed. T003/T004 later deployed the hosted repair and passed 3/3 v3 checks; the latest v5 remains retained without native certification. The diagnostic is Inactive; current Foundry metadata selects the unchanged active version-4 control. Neither proves a native conversation. The criterion-level lab review closes no outstanding acceptance criterion. Release approvals, retention ownership, browser access, policy and original live acceptance gates remain unresolved. The chronological evidence below is retained, not promoted to current completion.

Important

Latest continuation: the manual external Foundry connection is Connected, its native attachment and explicit topic persist, and direct Responses inference passed. The first corrective Activity test returned no response; a later instrumented native test received the genuine reply in 6.864 seconds, passing synthetic connected transport, not grounded scenario acceptance. The temporary diagnostic observer was deleted with independent 404s. Original-source review found native annotation/publisher-URL gates were added by the assistant. Canonical source metadata now preserves all 24 originals; 4/4 unchanged notebook questions pass extractive backend content and attribution. A later independent native run gives 4/4 original questions passing displayed content and individual canonical attribution after one targeted version-4 instruction fix. Five submissions include one unacknowledged combined attempt and one bounded transport reattempt. The earlier lost batch remains historical; its missing results were not reconstructed. Native MCP IDs remain unexposed. The original DOCX ingestion error remains, but an MIT-licensed official publisher guide now reaches Ready, with exact 24-record/four-query notebook equivalence. No protection was removed or DOCX equivalence claimed. Generative parent routing is enabled; a real CTR lookup cites the uploaded source without Foundry. An expanded QA attempt then failed intake, found incomplete CTR coverage and stopped on unacknowledged SAR delivery. After targeted rework, the exact first three scenarios receive real final replies: source-free intake, complete CTR and conditional SAR lookups with native local citations. The earlier failures remain recorded; these focused checks do not certify the full eleven-scenario natural/local/hybrid matrix or prove which intake topic executed. The independent fresh-session R2 retest stops on one unacknowledged Hello preflight before any scored target: all eleven cases remain unassessed in R2. Later native readback exposed an empty intake definition. Normal keyboard editing repaired it, and a saved/reopened readback plus actual topic/action traces now prove its execution. A startup-readiness barrier and normal Send button recovered native delivery without an external-browser result. All eleven R3 questions received responses, but six child-stage captures are not whole-parent acceptance evidence. Later captures require DynamicPlanFinished; repeated validation targets are counted separately. See the persistence recovery. The latest complete-turn ledger now contains all eleven original questions with the expected handler routes. This is merged evidence, not one fresh eleven-case run. Complete-turn auditing finds nonliteral quotations and grouped markers; detailed example coverage remains incomplete. An incident-only variant failed and was rolled back, with a genuine post-rollback canary. Full matrix/workshop acceptance remains failed. Standard capacity was corrected to 26,000 TPM / 26 RPM after a proven direct rate-limit failure. A separate GPT-4.1 comparison completed with 6,357 measured tokens but still failed conditional-duty grounding. Its temporary agent/deployment were deleted with independent 404 checks; it was not promoted to the native agent. Search Basic, embeddings, account/project and both agents are live and retained, unlike the earlier deleted runs. Its expiry tag does not delete resources. See the current original-requirement validation. The model-only comparison did not resolve narrative synthesis. See the later native and licensed-reference recovery for the genuine native and local passes, not full-workshop certification.

Earlier execution evidence: actual rerun record. The owner approved paid testing. A corrected 15-instance Terraform deployment, 14 keyless model/retrieval checks, and one actual synthetic Application Insights event with exact KQL correlation passed; the owned resources were deleted and absence verified. A further private PDF pipeline passed 25 checks, including an actual Blob indexer, nine child projections and integrated 1,536-dimensional embeddings. The latest record separates its cleanup from the earlier run. Lab 6's SDK failure and isolated restoration were also tested. Its later paid Foundry run generated 24 real embeddings and verified 22 retrieved passages. Direct authenticated MCP and single-source Foundry-agent grounding passed. Cross-source final exact quotations failed; Activity configuration passed without an Activity/Copilot conversation. That isolated environment was also deleted and its absence independently verified. The later approved quotation recovery passed all three exact quotations twice, but native citation fields were blank. The unsupported model abstained despite seven off-topic retrieved documents. Four new sequential test environments were deleted; independent checks confirmed four absent groups and eight account/Search ARM 404 responses. The backend continuation authenticated to Dataverse and cloned the actual saved agent through official PAC. Direct Line discovery and PAC evaluation authorization remain blocked; a clone is not a conversation. The later browser recovery did execute three fresh Preview checks: uploaded-PDF retrieval/answer capture, unsupported-price refusal and visible invalid-input validation 400. Evaluate UI access recovered, but no evaluation was created or run. An independent retained-PDF audit then passed all five checks for the new answer's source identity, edition, complete quotation and page 3 section. Historical full-answer auditing, citation navigation and connected-lab acceptance remain open. The October 25 historical full-answer audit attempt reverified the retained October 2026 PDF hash, but the complete October 19 answer could not be located and provenance-matched. The accuracy criterion is open/unavailable, not passed or contradicted; claim totals are unknown. No newer answer or clipped transcript was substituted. The separate licensing-PDF navigation failure and downstream connector error-body fidelity remain unchanged; Lab 1 is not certified. With explicit credit approval, one three-case General quality evaluation was then saved and submitted. It was observed queued, not graded; the 11:36 UTC browser recovery failed, so its current state is unknown. The approved-run evidence does not claim a score, credit total or repaired PAC permission. The next shared tab displayed a Dataverse setup error, but fresh authenticated reads returned HTTP 200 for the same provisioned environment, identity and saved bot. That portal message does not prove deletion or justify a replacement database. An offline Foundry citation-contract diagnosis also passed 45 checks, identifying unsupported JSON/title-binding assumptions without inventing missing native annotations. One supported-format live retest then preserved source sentences but invented all three document URLs and still produced zero native annotations. It stopped without repair prompting. The fifth owned environment was deleted; all five groups and ten account/Search ARM 404s were independently verified. The tables below retain the original review baseline unless superseded by those records.

The working session later recovered the original completed evaluation: 1 pass, 2 failures, 7 Copilot Credits, 00:02:44. No duplicate was submitted. The completed-run report preserves the refusal/relevance mismatch and PDF completeness/citation failure. A Standard authoring/test destination is now reachable, but merely reaching its integration controls does not pass the connected-lab gates. The later Lab 3 connected test did pass the remaining agent-originated telemetry boundary: one real native topic execution, privacy-settings readback and one independently matched Application Insights event with five synthetic properties. No synthetic emitter or published channel was used. Platform conversation/channel identifier keys still exist; their values are excluded from published evidence.

After explicit user approval, the trial extension removed the expired-trial dialog and restored generative settings. No recurring purchase, new environment, unrelated consent or web-search activation occurred. The expiry date was not displayed or verified. The advanced authentication contracts were then inspected: Search offers Entra Integrated, whereas prompt BYOM requires an API key incompatible with the enforced keyless policy. The actual Search connection attempt failed because the hosting browser blocked its sign-in pop-up. No consent screen, connection or connected question occurred, despite successful Azure vectorizer queries. Cleanup subsequently verified HTTP 404 for the owned group, Search, AI account, embedding deployment and four scoped role assignments. No active resources from this test remain; historical charges may still settle.

A newer formal AccessKey connection subsequently succeeded without an OAuth pop-up, on a different disposable Search service with Cognitive keyless security preserved. Across nine actual connected Search questions, the original licensing speculation failed, then native grounding instructions and a source-aware empty-answer branch corrected it. Final negative and positive regressions passed; the hotel answer matched its real location field and its negative abstained. The complete chronology retains every failure. A later native viewer check verified the authentic PDF at page 3, but not the original chat citation's click-through. Hotel document URLs remain absent. One real formal Foundry Entra connection attempt also hit the pop-up restriction; no Activity conversation ran. This newer environment was subsequently disconnected and independently deleted: remote removal and current-run cleanup. Project, scoped role, group, Search, AI account and embedding deployment returned HTTP 404 at 23:20 UTC; the external governance destination remained untouched. The revised unpublished topic persists without a live Search binding.

A later external Edge recovery reached the real Foundry Entra form, but native endpoint input did not persist and Create remained disabled. Zero authentication submissions occurred; neither an authentication rejection nor external pop-up behavior was tested. The requested manual handoff was unavailable. No model or inference was added. Fresh native list readback then verified no owned connection. The auth-only resources were deleted at 23:58:49 UTC, with project/account/group 404 and governance preserved. A soft-deleted account recovery record remains; no purge was attempted. An additional no-resource input prerequisite check stopped before the Legacy probe because the browser could not be foregrounded and the form did not open. A later shared integrated-browser input check opened the form with normal keyboard activation and verified a harmless endpoint probe, then canceled without submission. External Edge recovery and removal of the integrated authentication pop-up restriction remain unproven. The published Inference connector contract also requires an API key; it documents no Entra alternative for that connector.

Documentation review and publication are separate from live lab validation. The workshop is not certified as fully rerun. Lab 1 has user-provided evidence of its email outcome plus a fresh core-path rerun. Other labs now have real cloud subsystem evidence, but their complete connected-agent integrations remain unverified.

The browser initially reached an authentication barrier. The user then signed in and initialized the GitHub wiki. That removed the publication barrier; it did not automatically create Search, telemetry, models, indexes, knowledge bases, or connectors.

The synthetic Search run also passes S5-03/S5-05/S5-06 after explicit readiness correction. This proves retrieval and filtering on the exact fictional two-document fixture, not Blob ingestion, connected answers, navigable citations or denied-identity behavior. Both disposable indexes are deleted; original index schemas remain unchanged. The SDK portability preflight proves the live file API and existing binding, but the exported template loses instructions and five components. It is not a faithful published derivative. The new staging candidate has verified file hashes, binding and integrated authentication, but remains incomplete and unpublished. No SDK answer acceptance is inferred. After verified atomic reconstruction, the candidate is API-confirmed published. The owned-client SDK attempt still submits zero questions because silent Invoke authentication yields no token. Exact permissions/grant rollback passes; publication does not close native answer acceptance. The later bounded consent recovery does acquire a verified Invoke token, but SDK start yields only one typing activity and no question. Final independent restoration passes after transport errors; conversation-start and whole-answer evidence remain incomplete.

The subsequent header-aware capture does recover a server conversation ID and captures Hello plus unchanged SAR case 6, once each, with HTTP 200 and normally closed streams. Diagnostic start suppresses its event, so greeting acceptance is not established. Hello returns an investigation clarification; SAR returns a runtime connector Connect/Continue card, not a domain answer. Only DynamicPlanReceived, DynamicPlanStepTriggered and DynamicPlanStepBindUpdate are observed; no DynamicPlanFinished or source comparison. Connect/Continue remains uninvoked. Exact rollback and original-agent preservation pass. This proves parent response capture, not full native acceptance.

The connection follow-up stops before Connect because the website admin session cannot be verified through UIAutomation, despite the identified browser profile and environment. It makes no cloud writes, permission changes or new question submissions; independent preservation checks pass. Human account verification and connection completion remain required. This is not evidence of an MFA/OAuth failure or a completed SAR.

The later supported browser-page attempt does verify the exact website admin and target environment. One first-party connector Create submission returns Unable to sign in. Please try again. No success, Continue, fresh question or runtime event follows. Preservation checks pass. At that phase the blocker is actual connector sign-in, not unreadable website identity; the generic error does not prove its underlying cause.

The later native Studio association recovers per-user authorization by selecting the existing approved connection. Actual native SAR delivery then fails source fidelity: seven of nine and five of six quoted passages match their own labeled source groups; the final unquoted paraphrase is not a pass. The existing hosted diagnostic returns the caller-tenant HTTP 403; one bounded Entra-only experiment instead returns a blueprint-ID HTTP 400. Both diagnostics are restored Off. Exact endpoint/card, Entra plus BotServiceRbac, version-3 definition/identity, source manifest and independent preservation checks pass. No anonymous access, role change, deployment, publication or source-fidelity acceptance is claimed. The later published SDK retest then sends one unchanged SAR case and still receives the per-user Foundry connection card, not an answer or plan-finished event. HTTP 200 and normal stream closure are not authorization acceptance. No Connect/Continue follows; temporary owned manifest/current-user grants are restored exactly and the new Invoke token alone is removed. Maker access/valid refresh entries and all source/candidate/files/bindings/authentication/fixture/PAC checks are preserved.

The policy-compatible private fixture run adds genuine S5-01/S5-02 passes: two exact privately uploaded files, successful private Blob indexing with zero failed items/warnings, and two readable exact passages with real 1,536-dimensional vectors. All disposable resources and roles are removed and original Search/ARM state verified preserved. Earlier network, readback and expired-token/cleanup failures remain retained; no policy is weakened. These two new ingestion passes do not establish connected answers or navigation.

Status definitions

Status	Meaning
Observed user run	The provided evidence shows a specific outcome; not a new assistant rerun
Read-only check passed	A tool directly observed a narrow prerequisite or repository condition
Procedure reviewed	The source workflow was compared with official guidance; execution is still needed
Static check	A local file or code property was inspected without executing its cloud operations
Blocked	A prerequisite, permission, environment, or supported integration prevents completion
Not run	No execution result is available

An unavailable or untested step remains blocked/not run. A screenshot from a Word document, a generated diagram, or a plausible agent answer cannot upgrade that status.

Original review baseline

The following table preserves the initial review state. Use the linked current per-lab execution table for later real results.

Lab	Current result	Required evidence before a full pass
0	Partial read-only preflight; complete environment readiness not established	Environment, entitlement/capacity, policy, data permissions, quota, Terraform, and budget verified
1	Observed user run: send tool and received message; not every negative test executed	New controlled run with grounded answer, recipient resolution, send result, receipt, and negative tests
2	Procedure reviewed; no demonstrated complete retrieval run	Successful indexer, nonempty relevant Search results, connected Copilot answer with expected sources
3	Procedure reviewed; no demonstrated telemetry ingestion	Deliberate event correlated to a recent lab interaction and visible in the intended log query
4	Procedure reviewed; no demonstrated connected-model run	Direct model test, parsed retrieval, prompt invocation, grounded answer, explicit empty/error behavior
5.1	Static deployment review; no apply performed	Compatible providers, reviewed plan, approved apply, expected resources and private configuration
5	Source/static review; advanced query results not observed	Supplied corpus indexed, required queries executed, expected fields/ranking/filter behavior recorded
6	Source/static review; end-to-end connected-agent run not observed	Complete data/configuration, supported SDK/API, direct retrieval, Foundry agent and Copilot connection verified

Checks performed in this session

Check	Observed result	Scope limitation
Source inventory	Labs 0–6 and supplement 5.1 found, plus supporting files	Does not prove the runtime package is complete
Word source access	Readable ZIP-based documents extracted locally; compound-format files read through installed Word	Protected source documents are not republished
Lab numbering	Monitoring/BYOM internal titles contradict their filenames	Guides consistently follow filenames
Azure CLI	Installed version 2.91.0; selected account query succeeded	No claim about every subscription or tenant
Provider registration	Search, Cognitive Services, Insights observed Registered	Registration does not prove role rights, capacity, network access, or deployment success
Selected-subscription resource inventory	Original absence claim corrected by explicit recheck: 9 Cognitive accounts, 4 projects, 7 unrelated Insights components, 0 Search before disposable tests	Free Search tests and verified cleanup are in the latest runtime record
Terraform executable	Not found on PATH	No terraform init, validate, plan, or apply claimed
Copilot browser	Sign-in succeeded and Home was visible; agent navigation then returned a persistent blank page	Licensing advisor could not be opened; no new knowledge question was submitted
GitHub publication access	Private repository created; wiki enabled and initialized by the owner	Repository and wiki require separate commits and remote verification
Lab 1 attachments	Send-tool/result crop and matching received-email crop preserved	No independent transport-header or licensing-content audit
Current product guidance	Official Microsoft Learn references checked for the documented substitutions	Documentation checks are not runtime tests
Lab 6 notebook	11 of 11 code cells syntax-valid; 24 inline records identified	Syntax and data presence do not prove API, inference, retrieval, or integration compatibility
Advanced-lab examples	Four JSON examples checked; SVG diagrams parsed; acceptance IDs aligned between languages	Examples and diagrams do not prove cloud behavior

The advanced-lab review also identified Terraform identity/network gaps, disabled Foundry provisioning, embedding-dimension consistency requirements, API/SDK drift, and destructive reset behavior that must be resolved or consciously gated before a live run. Consult the detailed findings in Labs 5.1, 5, and 6 rather than applying the original package unchanged.

Advanced-package findings that need action

1. The supplied Terraform configuration disables the Foundry module, so the advertised Foundry project is not provisioned by that path.
2. The Search module does not apply the supplied allowed-IP/Entra options as expected, and the required service identity is not attached. Review effective resources and role assignments before treating the configuration as an identity/network baseline.
3. The deployment supplement uses 1,536-dimensional embeddings while Lab 6 expects 3,072. These are different configuration expectations, not interchangeable indexes. Align the model output, index field, and query embedding for the selected lab.
4. Preview SDK, ARM, and MCP contracts require compatibility reconciliation. The connected-agent route requires the supported Activity/new-portal agent contract.
5. Notebook reset code deletes fixed-name objects; the final cleanup is commented out. Do not run reset cells against a shared project.

The notebook also references a quick-reference Markdown file that is absent. The package supplies data\quick_reference_guide.docx instead. Renaming an extension does not convert a document; obtain an authorized source in the expected format.

Documentation and tooling gates

The offline gates are reproducible using:

python -m py_compile scripts\build_wiki.py scripts\validate_docs.py scripts\test_docs.py
python -m unittest discover -s scripts -p "test_*.py" -v
python scripts\build_wiki.py --output .wiki-build
python scripts\validate_docs.py --wiki .wiki-build

The offline checks passed for the publication snapshot: Python compilation, seven unit tests, 28 canonical Markdown files, 13 English/French page pairs, 29 manifested PNG figures, and 30 exported wiki Markdown files. There were no broken local links or missing image/provenance entries. No pre-existing linter or package build was configured in the initially empty repository.

The final publication is accepted only after these checks succeed and the remote repository/wiki commits match the local published commits. These checks validate the documentation deliverable, not cloud execution. See publication procedure for export details and maintenance.

Remaining live validation work

1. Confirm the selected workshop environment, harness, credits, permissions, quota, data policies, approved budget, and cleanup owner.
2. Preserve the working Lab 1 agent before changing its knowledge or topics.
3. Provision only the dedicated, approved Search and telemetry resources required by the selected labs. Do not reuse unrelated production resources.
4. Execute the test matrices in each guide in dependency order.
5. Capture new, redacted evidence with input, result, timestamp, source edition, model/API version, and downstream outcome where applicable.
6. Fix the smallest failing boundary, retest it, then retest the integrated path.
7. Update both language ledgers and distinguish partial tests from a complete lab pass.

The initial documentation-only phase created no billable resources. Later owner-approved paid tests and their cleanup are explicitly recorded in the execution ledger. No billing-policy change or modification of unrelated applications was made. Stop at blocked rather than inventing an integrated pass from subsystem results.